Paper deep dive

OSVE: One Step Video Editing with One Step Diffusion Models

Habin Lim, Gyeong-Moon Park

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 103

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%

Last extracted: 7/23/2026, 2:48:45 AM

Summary

The paper introduces OSVE, a framework for one-step video editing that adapts one-step Text-to-Image diffusion models to achieve high-quality, temporally consistent video editing with significantly reduced latency. Key innovations include a learnable encoder for single-pass inversion, a Structure-Aware Editing (SAE) loss to preserve geometry, and Unified-Frame Editing (UFE) to ensure temporal coherence via cross-frame attention.

Entities (8)

OSVE · framework · 98%Unified-Frame Editing · method · 94%Structure-Aware Editing Loss · methodloss-function · 92%Korea University · organization · 90%DMD2 · model · 85%ControlNet · method · 82%Prompt-to-Prompt · method · 82%RAVE · modelmethod · 80%

Relation Signals (8)

Gyeong-Moon Park → affiliatedwith → Korea University

confidence 95% · Habin Lim and Gyeong-Moon Park ⋆ Korea University

Habin Lim → affiliatedwith → Korea University

confidence 95% · Habin Lim and Gyeong-Moon Park ⋆ Korea University

OSVE → uses → Structure-Aware Editing Loss

confidence 95% · This encoder is trained with a novel Structure-Aware Editing (SAE) loss

OSVE → uses → Unified-Frame Editing

confidence 95% · For temporal coherence, we introduce Unified-Frame Editing (UFE)

OSVE → outperforms → RAVE

confidence 90% · OSVE achieves editing quality comparable or superior to state-of-the-art multi-step methods, while operating approximately 155--171 times faster.

Prompt-to-Prompt → failsin → One-Step Regime

confidence 88% · We empirically find that existing structure-preserving methods such as Prompt-to-Prompt [12] ... fail when constrained to a single step.

ControlNet → failsin → One-Step Regime

confidence 88% · We empirically find that existing structure-preserving methods such as ... ControlNet [73] fail when constrained to a single step.

OSVE → isbasedon → DMD2

confidence 85% · We instantiate and validate our framework on a publicly released one-step T2I backbone (DMD2 [68])

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Text-guided video editing with diffusion models is impractically slow, hindered by costly multi-step sampling and inversion. We present OSVE, the first framework to successfully adapt one-step Text-to-Image (T2I) models for high-quality video editing, addressing the core challenges of inversion, editability, and temporal consistency. To bypass slow iterative inversion, we train a learnable encoder that predicts the initial noise for each frame in a single forward pass. This encoder is trained with a novel Structure-Aware Editing (SAE) loss on a curated dataset of structurally-aligned image pairs, teaching it to preserve the source video's geometry during edits. For temporal coherence, we introduce Unified-Frame Editing (UFE), a technique that concatenates frame latents to facilitate cross-frame attention in a single generation step. Furthermore, for long videos, a sliding-window strategy with an anchor frame maintains global consistency. Our extensive experiments demonstrate that OSVE achieves editing quality comparable or superior to state-of-the-art multi-step methods, while operating approximately 155--171 times faster. This breakthrough paves the way for practical, real-time video editing applications. Code is available at this https URL.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

101,309 characters extracted from source content.

Expand or collapse full text

OSVE: One Step Video Editing with One Step Diffusion Models Habin Lim and Gyeong-Moon Park ⋆ Korea University, Seoul, Republic of Korea ha001211,gm-park@korea.ac.kr Abstract. Text-guided video editing with diffusion models is imprac- tically slow, hindered by costly multi-step sampling and inversion. We present OSVE, the first framework to successfully adapt one-step Text- to-Image (T2I) models for high-quality video editing, addressing the core challenges of inversion, editability, and temporal consistency. To bypass slow iterative inversion, we train a learnable encoder that predicts the ini- tial noise for each frame in a single forward pass. This encoder is trained with a novel Structure-Aware Editing (SAE) loss on a curated dataset of structurally-aligned image pairs, teaching it to preserve the source video’s geometry during edits. For temporal coherence, we introduce Unified- Frame Editing (UFE), a technique that concatenates frame latents to facilitate cross-frame attention in a single generation step. Furthermore, for long videos, a sliding-window strategy with an anchor frame main- tains global consistency. Our extensive experiments demonstrate that OSVE achieves editing quality comparable or superior to state-of-the-art multi-step methods, while operating approximately 155–171 times faster. This breakthrough paves the way for practical, real-time video editing applications. Code is available at https://github.com/KU-VGI/OSVE. 1 Introduction Text-guided video editing, i.e., modifying a source video’s visual content accord- ing to a natural-language instruction, is increasingly vital for content creation, filmmaking, and interactive media. Recent generative diffusion models, spanning Text-to-Image (T2I) [13,45,52,55] and Text-to-Video (T2V) [1,11,20,40,60] ar- chitectures, have enabled the generation of photorealistic, temporally coherent visual content from text alone. Building upon these powerful backbones, a grow- ing body of work performs training-free, zero-shot video editing by intervening in the diffusion process [6,9,15,17,29,61–63]. These methods typically follow a two- stage pipeline: (i) inverting the source video via multi-step inversion [36,53] to recover its latent trajectory, and (i) generating the edited frames by re-running the diffusion process under a modified text prompt, often with cross-frame at- tention control [12,59] to preserve spatial structure and temporal layout. Despite the editing quality these methods achieve, they inherit a critical bot- tleneck: extreme computational cost. Because every frame must be processed ⋆ Corresponding author. arXiv:2607.19895v1 [cs.CV] 22 Jul 2026 2H. Lim and G.-M. Park Fig. 1: Three frameworks for text-guided video editing. (a) Multi-to-Multi: The standard approach yields high-quality edits but is impractically slow. (b) Multi- to-One: Naïvely adopting a one-step diffusion model produces degraded, temporally inconsistent outputs. (c) One-to-One (Ours): Our proposed framework uses a learnable one-step encoder to enable fast, temporally consistent, and high-quality video editing, resolving the speed–quality trade-off. over tens to hundreds of denoising steps, wall-clock time scales with both the number of frames and the number of steps. For instance, RAVE [17], reported as one of the fastest T2I-based editing methods, still requires approximately 24 hours to edit a 5-minute, 30 FPS video 1 . T2V backbones, which incorpo- rate heavyweight temporal attention modules, demand even greater resources. This extreme latency renders current approaches impractical for real-time or streaming-style applications [64]. To overcome this bottleneck, we shift toward a one-step generation paradigm. One-step diffusion models [7,31,39,68,69], distilled from their multi-step teach- ers, have achieved competitive or even superior image quality, yet their potential for video editing remains unexplored. By adapting this ultra-fast regime to video, we build a One-to-One editing framework (one-step inversion followed by one- step generation) that replaces the costly iterative denoising loop with a single forward pass. We instantiate and validate our framework on a publicly released one-step T2I backbone (DMD2 [68]) to demonstrate its effectiveness with today’s most accessible one-step generators. However, directly applying one-step diffusion models to video editing presents three critical challenges (see Figure 1(b)): (i) Inversion–generation mismatch. Conventional multi-step inversion [53] is both time-consuming and fundamentally ill-suited for one-step generators. Since the inversion follows a multi-step trajectory while the generator performs a direct mapping, the mismatch between forward and reverse paths causes severe information loss, producing blurred and degraded results. (i) Structural collapse under single-step control. Preserving the source video’s spatial structure (e.g., a wolf’s rotating head) while altering its appear- ance (e.g., to a bear) is critical for faithful editing. We empirically find that existing structure-preserving methods such as Prompt-to-Prompt [12] and Con- 1 Measured on a single NVIDIA RTX 6000 Ada GPU at 512× 512 resolution. OSVE: One Step Video Editing with One Step Diffusion Models3 trolNet [73] fail when constrained to a single step. Unlike multi-step processes where control signals are distributed across steps and gradually absorbed, one- step models concentrate the same signal into a single pass with no room for correction, causing it to dominate rather than guide the output, a phenomenon we term “over-steering”. (i) Temporal inconsistency. One-step T2I diffusion models do not inher- ently encode temporal information, resulting in flickering or discontinuous frames when applied to video. Existing temporal consistency techniques, including at- tention fine-tuning [63], cross-frame attention control [6, 29, 42, 61], and noise shuffling across sampling steps [17], all rely on the multi-step structure and are therefore incompatible with the one-step regime. To address these challenges, we propose OSVE (One-Step Video Editing), a One-to-One video editing framework explicitly designed for one-step genera- tion. For inversion–generation mismatch, the core of OSVE is a learnable video encoder, initialized from a pre-trained U-Net to inherit a strong prior over the generator’s latent space, that replaces slow multi-step inversion with a sin- gle forward pass. For structural collapse, we introduce a Structure-Aware Editing (SAE) loss that trains the encoder to produce latents inherently pre- serving structural integrity under modified text prompts, eliminating the need for external guidance that fails in one step. By training on diverse editing pairs, the encoder learns structure-aware latent representations that maintain the source’s layout while permitting appearance changes (Figure 1(c)). For temporal in- consistency, we propose Unified-Frame Editing (UFE), which concatenates the inverted latents of consecutive frames into a single unified tensor, enabling the model’s self-attention to align features across frames in one pass. Combined with a sliding-window strategy for local coherence and anchor-frame selection for long-range stability, UFE ensures both short-term and long-term temporal consistency. In summary, our main contributions are: – We present OSVE, the first One-to-One video editing framework for the one-step diffusion regime, replacing multi-step inversion with a learnable single-pass encoder. – We propose a Structure-Aware Editing (SAE) loss that directly injects structural preservation into the inversion stage, mitigating the over-steering problem inherent in single-step control. – We introduce Unified-Frame Editing (UFE), which enforces temporal consistency via unified latent processing with a sliding window and anchor frame, ensuring both short- and long-range coherence. – Extensive experiments show that OSVE achieves sub-second latency, ∼ 155–171× faster than the previous fastest method, while delivering compa- rable or superior visual quality, enabling practical real-time video editing. 4H. Lim and G.-M. Park 2 Related Work Text-Guided Video Editing with Diffusion Models. Building on advances in image-level generative editing, including GAN inversion [37, 38, 44, 58] and diffusion-based manipulation [3,12,24,36], numerous zero-shot techniques have been proposed to extend such capabilities to video while preserving temporal consistency. These include fusing attention maps [41], propagating token fea- tures [9,61], using optical flow guidance [6], and shuffling noise [17]. While these approaches have significantly improved video coherence, they all operate within a multi-step sampling framework. This iterative nature, where operations are repeated for each frame across numerous steps, results in prohibitively long run- times. This computational cost remains a critical barrier, motivating our work on high-quality, single-step video editing. One-Step Diffusion Models. To overcome the significant latency of iterative diffusion sampling, a prominent line of research focuses on distilling multi-step teacher models into highly efficient one- or few-step generators. This paradigm was advanced by methods like Progressive Distillation and Consistency Models, with latent variants (LCM/LCM-LoRA) enabling 2–8 step generation for Stable Diffusion [34,47,54]. Subsequent works have further refined this approach. Rec- tified Flow and its successor, InstaFlow, straighten probability-flow trajectories to achieve SD-level quality in a single step via reflow-based distillation [30,31]. Other strategies, such as Adversarial Diffusion Distillation (e.g., SDXL-Turbo, SDXL-Lightning) and various distribution matching techniques (e.g., DMD, SwiftBrush), couple teacher supervision with GAN losses or novel distillation objectives to achieve high-quality synthesis in just 1–4 steps [7,25,39,48,68,69]. These one-step models have revolutionized image generation speed, but this progress has been exclusively image-centric. This creates a critical gap and a major opportunity, as the domain of video editing desperately needs such ac- celeration. Our work is the first to bridge this gap, introducing the necessary mechanisms to successfully adapt these fast generators for video, thereby un- locking their potential for practical, high-speed applications. 3 OSVE: One-Step Video Editing We present OSVE, a framework for one-step video editing built on two core ideas: (i) encoding source structure into the initial noise latent so that a single-pass generator can preserve it without iterative control, and (i) enforcing temporal consistency by letting all frames share context through a unified attention mech- anism. Section 3.1 develops the first idea through a structure-aware inversion encoder trained with a dedicated editing loss. Section 3.2 introduces Unified- Frame Editing (UFE) for the second, along with sliding-window and anchor- based schemes that scale it to long videos. Figure 2 illustrates an overview of the framework. OSVE: One Step Video Editing with One Step Diffusion Models5 132 2 padded 1 ⃜ Trimming K 2 1 ⃜ 12...K Structure-Aware Inversion <latexit sha1_base64="desf1fd+RoBa09yQ1RVhF83cDtk=">AAAB+XicbVBNS8NAFHypX7V+RT16WSyCp5KIVI9FL3qrYG2hCWGz3bRLN5uwuymU0H/ixYMiXv0n3vw3btoctHVgYZh5jzc7YcqZ0o7zbVXW1jc2t6rbtZ3dvf0D+/DoSSWZJLRDEp7IXogV5UzQjmaa014qKY5DTrvh+LbwuxMqFUvEo56m1I/xULCIEayNFNi2F2M9Ipjn97PAS0cssOtOw5kDrRK3JHUo0Q7sL2+QkCymQhOOleq7Tqr9HEvNCKezmpcpmmIyxkPaN1TgmCo/nyefoTOjDFCUSPOERnP190aOY6WmcWgmi5xq2SvE/7x+pqNrP2cizTQVZHEoyjjSCSpqQAMmKdF8aggmkpmsiIywxESbsmqmBHf5y6vk6aLhNhvNh8t666asowoncArn4MIVtOAO2tABAhN4hld4s3LrxXq3PhajFavcOYY/sD5/AM0fk8k=</latexit> I ω ❄ Input Video K 2 1 Output Video <latexit sha1_base64="TLZ4qdsJKRCIU0YIIrKsnHokoKc=">AAAB+HicbVDLSsNAFL3xWeujUZduBovgqiQi1WXRjeCmgn1AG8NkOmmHTiZhZiLU0C9x40IRt36KO//GSZuFth4YOJxzL/fMCRLOlHacb2tldW19Y7O0Vd7e2d2r2PsHbRWnktAWiXksuwFWlDNBW5ppTruJpDgKOO0E4+vc7zxSqVgs7vUkoV6Eh4KFjGBtJN+u9COsRwTzrD31bx8c3646NWcGtEzcglShQNO3v/qDmKQRFZpwrFTPdRLtZVhqRjidlvupogkmYzykPUMFjqjyslnwKToxygCFsTRPaDRTf29kOFJqEgVmMo+pFr1c/M/rpTq89DImklRTQeaHwpQjHaO8BTRgkhLNJ4ZgIpnJisgIS0y06apsSnAXv7xM2mc1t16r351XG1dFHSU4gmM4BRcuoAE30IQWEEjhGV7hzXqyXqx362M+umIVO4fwB9bnD4yYkwg=</latexit> V 0 K ⃜ K 2 1 ⃜ ... Inference with Sliding Window padded K-1 padded 2 1 Tiger padded K padded 2...K-11 K-1K padded 2...K-11 padded ⃜ 1...2 Tiger Tiger K Frames padded K-1 (b) Unified-Frame Editing (§3.2) w = 3 Inverse Attention Sharing (a) Structure-Aware One-Step Inversion (§3.1) <latexit sha1_base64="AnY0zmxQPt06dT+spaHEk74o1+Q=">AAACF3icbVBNS8MwGE7n15xfVY9egkPwIKMVmR6HXjxOcB/QlZKm6RaWpiVJhVH6L7z4V7x4UMSr3vw3pltF3Xwh5OF53ufNm8dPGJXKsj6NytLyyupadb22sbm1vWPu7nVlnApMOjhmsej7SBJGOekoqhjpJ4KgyGek54+vCr13R4SkMb9Vk4S4ERpyGlKMlKY8szGYznDE0Hczu2EVdWLN7h+QYy8jAVW5Z9a/ObgISr9VB2W1PfNjEMQ4jQhXmCEpHdtKlJshoShmJK8NUkkShMdoSBwNOYqIdLPpTjk80kwAw1jowxWcsr8dGYqknES+7oyQGsl5rSD/05xUhRduRnmSKsLx7KEwZVDFsAgJBlQQrNhEA4QF1btCPEICYaWjrOkQ7PkvL4LuacNuNpo3Z/XWZRlHFRyAQ3AMbHAOWuAatEEHYHAPHsEzeDEejCfj1XibtVaM0rMP/pTx/gW6OZvr</latexit> c edit <latexit sha1_base64="EFFT+q5eofYICKO5h7TqBztUMT0=">AAAB6HicbZC7SgNBFIbPxltcb1FLm8UgWIVdi2gjBm0sEzAXSJYwOzmbjJmdXWZmhbDkCWwsFLHVh7G3Ed/GyaXQ6A8DH/9/DnPOCRLOlHbdLyu3tLyyupZftzc2t7Z3Crt7DRWnkmKdxjyWrYAo5ExgXTPNsZVIJFHAsRkMryZ58w6lYrG40aME/Yj0BQsZJdpYNdotFN2SO5XzF7w5FC/e7fPk7dOudgsfnV5M0wiFppwo1fbcRPsZkZpRjmO7kypMCB2SPrYNChKh8rPpoGPnyDg9J4yleUI7U/dnR0YipUZRYCojogdqMZuY/2XtVIdnfsZEkmoUdPZRmHJHx85ka6fHJFLNRwYIlczM6tABkYRqcxvbHMFbXPkvNE5KXrlUrrnFyiXMlIcDOIRj8OAUKnANVagDBYR7eIQn69Z6sJ6tl1lpzpr37MMvWa/fKY2QLw==</latexit> c Edit Target Edit Target Source Recon. Target (a.2) Training Inversion Encoder Reconstruction " Editing Reconstruction Loss <latexit sha1_base64="EFFT+q5eofYICKO5h7TqBztUMT0=">AAAB6HicbZC7SgNBFIbPxltcb1FLm8UgWIVdi2gjBm0sEzAXSJYwOzmbjJmdXWZmhbDkCWwsFLHVh7G3Ed/GyaXQ6A8DH/9/DnPOCRLOlHbdLyu3tLyyupZftzc2t7Z3Crt7DRWnkmKdxjyWrYAo5ExgXTPNsZVIJFHAsRkMryZ58w6lYrG40aME/Yj0BQsZJdpYNdotFN2SO5XzF7w5FC/e7fPk7dOudgsfnV5M0wiFppwo1fbcRPsZkZpRjmO7kypMCB2SPrYNChKh8rPpoGPnyDg9J4yleUI7U/dnR0YipUZRYCojogdqMZuY/2XtVIdnfsZEkmoUdPZRmHJHx85ka6fHJFLNRwYIlczM6tABkYRqcxvbHMFbXPkvNE5KXrlUrrnFyiXMlIcDOIRj8OAUKnANVagDBYR7eIQn69Z6sJ6tl1lpzpr37MMvWa/fKY2QLw==</latexit> c Input <latexit sha1_base64="AnY0zmxQPt06dT+spaHEk74o1+Q=">AAACF3icbVBNS8MwGE7n15xfVY9egkPwIKMVmR6HXjxOcB/QlZKm6RaWpiVJhVH6L7z4V7x4UMSr3vw3pltF3Xwh5OF53ufNm8dPGJXKsj6NytLyyupadb22sbm1vWPu7nVlnApMOjhmsej7SBJGOekoqhjpJ4KgyGek54+vCr13R4SkMb9Vk4S4ERpyGlKMlKY8szGYznDE0Hczu2EVdWLN7h+QYy8jAVW5Z9a/ObgISr9VB2W1PfNjEMQ4jQhXmCEpHdtKlJshoShmJK8NUkkShMdoSBwNOYqIdLPpTjk80kwAw1jowxWcsr8dGYqknES+7oyQGsl5rSD/05xUhRduRnmSKsLx7KEwZVDFsAgJBlQQrNhEA4QF1btCPEICYaWjrOkQ7PkvL4LuacNuNpo3Z/XWZRlHFRyAQ3AMbHAOWuAatEEHYHAPHsEzeDEejCfj1XibtVaM0rMP/pTx/gW6OZvr</latexit> c edit <latexit sha1_base64="EFFT+q5eofYICKO5h7TqBztUMT0=">AAAB6HicbZC7SgNBFIbPxltcb1FLm8UgWIVdi2gjBm0sEzAXSJYwOzmbjJmdXWZmhbDkCWwsFLHVh7G3Ed/GyaXQ6A8DH/9/DnPOCRLOlHbdLyu3tLyyupZftzc2t7Z3Crt7DRWnkmKdxjyWrYAo5ExgXTPNsZVIJFHAsRkMryZ58w6lYrG40aME/Yj0BQsZJdpYNdotFN2SO5XzF7w5FC/e7fPk7dOudgsfnV5M0wiFppwo1fbcRPsZkZpRjmO7kypMCB2SPrYNChKh8rPpoGPnyDg9J4yleUI7U/dnR0YipUZRYCojogdqMZuY/2XtVIdnfsZEkmoUdPZRmHJHx85ka6fHJFLNRwYIlczM6tABkYRqcxvbHMFbXPkvNE5KXrlUrrnFyiXMlIcDOIRj8OAUKnANVagDBYR7eIQn69Z6sJ6tl1lpzpr37MMvWa/fKY2QLw==</latexit> c <latexit sha1_base64="mvNZ51H7HCspI67Oj6IpresiLvU=">AAAB/3icbVDLSgMxFM3UV62vUcGNm2ARKkiZEakui25cSQX7gM5QMmmmDU0yQ5IRytiFv+LGhSJu/Q13/o2ZdhbaeiBwOOde7skJYkaVdpxvq7C0vLK6VlwvbWxube/Yu3stFSUSkyaOWCQ7AVKEUUGammpGOrEkiAeMtIPRdea3H4hUNBL3ehwTn6OBoCHFSBupZx94inLocaSHGLH0dlJxTqF70rPLTtWZAi4SNydlkKPRs7+8foQTToTGDCnVdZ1Y+ymSmmJGJiUvUSRGeIQGpGuoQJwoP53mn8Bjo/RhGEnzhIZT9fdGirhSYx6YySyomvcy8T+vm+jw0k+piBNNBJ4dChMGdQSzMmCfSoI1GxuCsKQmK8RDJBHWprKSKcGd//IiaZ1V3Vq1dnderl/ldRTBITgCFeCCC1AHN6ABmgCDR/AMXsGb9WS9WO/Wx2y0YOU7++APrM8fBryU0Q==</latexit> →N(0,1) “A photo of a dog” Source Edited (a.1) Dataset Generation Structurally Simillar Structure-Aware Editing Loss <latexit sha1_base64="EFFT+q5eofYICKO5h7TqBztUMT0=">AAAB6HicbZC7SgNBFIbPxltcb1FLm8UgWIVdi2gjBm0sEzAXSJYwOzmbjJmdXWZmhbDkCWwsFLHVh7G3Ed/GyaXQ6A8DH/9/DnPOCRLOlHbdLyu3tLyyupZftzc2t7Z3Crt7DRWnkmKdxjyWrYAo5ExgXTPNsZVIJFHAsRkMryZ58w6lYrG40aME/Yj0BQsZJdpYNdotFN2SO5XzF7w5FC/e7fPk7dOudgsfnV5M0wiFppwo1fbcRPsZkZpRjmO7kypMCB2SPrYNChKh8rPpoGPnyDg9J4yleUI7U/dnR0YipUZRYCojogdqMZuY/2XtVIdnfsZEkmoUdPZRmHJHx85ka6fHJFLNRwYIlczM6tABkYRqcxvbHMFbXPkvNE5KXrlUrrnFyiXMlIcDOIRj8OAUKnANVagDBYR7eIQn69Z6sJ6tl1lpzpr37MMvWa/fKY2QLw==</latexit> c Generate Source Image Generate Edited Image DB Sampling <latexit sha1_base64="vVsLWiDRI1BsHgucHGTU9Za9GmE=">AAACM3icbVBJSwMxGM3UrdZt1KOXYBEqlDIjUj0WvYgHqWAX6AzlmzRtQzOLSUYow/wnL/4RD4J4UMSr/8F0A219EHi8b3t5XsSZVJb1amSWlldW17LruY3Nre0dc3evLsNYEFojIQ9F0wNJOQtoTTHFaTMSFHyP04Y3uBzVGw9USBYGd2oYUdeHXsC6jIDSUtu8TpzxkpboeW5il6wxitY8SZ37GDrYoQqwI5mPHR9UnwBPbtKCVcT2cdo287NuvEhmq/NoimrbfHY6IYl9GijCQcqWbUXKTUAoRjhNc04saQRkAD3a0jQAn0o3GdtN8ZFWOrgbCv0Chcfq74kEfCmHvqc7R07lfG0k/ldrxap77iYsiGJFAzI51I05ViEeBYg7TFCi+FATIIJpr5j0QQBROuacDsGe//IiqZ+U7HKpfHuar1xM48iiA3SICshGZ6iCrlAV1RBBj+gFvaMP48l4Mz6Nr0lrxpjO7KM/ML5/AEVGpec=</latexit> ω→N(0,1) <latexit sha1_base64="oU5pfFhT1G6cArt6Q4e0ANfG30M=">AAACP3icbVC7SgNBFJ31bXxFLW0Gg2AhYVdELUUbSwUThewSZmdv4uDszDJzV4zL/pmNv2Bna2OhiK2dkxhBowcGDufc15w4k8Ki7z96Y+MTk1PTM7OVufmFxaXq8krT6txwaHAttbmImQUpFDRQoISLzABLYwnn8dVR3z+/BmOFVmfYyyBKWVeJjuAMndSuNotwMKRlunFUBHV/gC1/lJShdEMT1i5ChBssQsuNyNCKW6BKCwtlSUOeaKQhICvb1dp3J/1LvtfUyBAn7epDmGiep6CQS2ZtK/AzjApmUHAJZSXMLWSMX7EutBxVLAUbFYPTS7rhlIR2tHFPIR2oPzsKllrbS2NXmTK8tKNeX/zPa+XY2Y8KobIcQfGvRZ1cUtS0HyZNhAGOsucIc3m4Wym/ZIZxdJFXXAjB6Jf/kuZ2Pdit757u1A4Oh3HMkDWyTjZJQPbIATkmJ6RBOLkjT+SFvHr33rP35r1/lY55w55V8gvexyfCJq02</latexit> ω noise ·ε <latexit sha1_base64="zCoKBXQAqf/8zwY1vL+3DQpS0ME=">AAACAHicbVBNS8NAEJ3Ur1q/oh48eAkWwVNJRKrHohfBS4V+QRPDZrtpl242YXcjlJCLf8WLB0W8+jO8+W/ctD1o64OBx3szzMwLEkalsu1vo7Syura+Ud6sbG3v7O6Z+wcdGacCkzaOWSx6AZKEUU7aiipGeokgKAoY6Qbjm8LvPhIhacxbapIQL0JDTkOKkdKSbx65I6QyN0JqhBHLOnnuZ3f5Q8s3q3bNnsJaJs6cVGGOpm9+uYMYpxHhCjMkZd+xE+VlSCiKGckrbipJgvAYDUlfU44iIr1s+kBunWplYIWx0MWVNVV/T2QoknISBbqzuFQueoX4n9dPVXjlZZQnqSIczxaFKbNUbBVpWAMqCFZsognCgupbLTxCAmGlM6voEJzFl5dJ57zm1Gv1+4tq43oeRxmO4QTOwIFLaMAtNKENGHJ4hld4M56MF+Pd+Ji1loz5zCH8gfH5A4YUlwU=</latexit> ˆ V T K <latexit sha1_base64="idCZ1WSfLlrEzQHkefIou0JqCLY=">AAACB3icbVDLSsNAFJ3UV62vqEtBgkVwVRKR6rLoRnBTwT6giWEymbRDJw9mboQSsnPjr7hxoYhbf8Gdf+OkzUJbD1w4nHMv997jJZxJMM1vrbK0vLK6Vl2vbWxube/ou3tdGaeC0A6JeSz6HpaUs4h2gAGn/URQHHqc9rzxVeH3HqiQLI7uYJJQJ8TDiAWMYFCSqx/aIwyZHWIYEcyzbp672Y2bUZ9Bnt+brl43G+YUxiKxSlJHJdqu/mX7MUlDGgHhWMqBZSbgZFgAI5zmNTuVNMFkjId0oGiEQyqdbPpHbhwrxTeCWKiKwJiqvycyHEo5CT3VWRws571C/M8bpBBcOBmLkhRoRGaLgpQbEBtFKIbPBCXAJ4pgIpi61SAjLDABFV1NhWDNv7xIuqcNq9lo3p7VW5dlHFV0gI7QCbLQOWqha9RGHUTQI3pGr+hNe9JetHftY9Za0cqZffQH2ucPJvmaJA==</latexit> ˆ V 0 K edit Inverted ... stride = 1 ... padded K <latexit sha1_base64="desf1fd+RoBa09yQ1RVhF83cDtk=">AAAB+XicbVBNS8NAFHypX7V+RT16WSyCp5KIVI9FL3qrYG2hCWGz3bRLN5uwuymU0H/ixYMiXv0n3vw3btoctHVgYZh5jzc7YcqZ0o7zbVXW1jc2t6rbtZ3dvf0D+/DoSSWZJLRDEp7IXogV5UzQjmaa014qKY5DTrvh+LbwuxMqFUvEo56m1I/xULCIEayNFNi2F2M9Ipjn97PAS0cssOtOw5kDrRK3JHUo0Q7sL2+QkCymQhOOleq7Tqr9HEvNCKezmpcpmmIyxkPaN1TgmCo/nyefoTOjDFCUSPOERnP190aOY6WmcWgmi5xq2SvE/7x+pqNrP2cizTQVZHEoyjjSCSpqQAMmKdF8aggmkpmsiIywxESbsmqmBHf5y6vk6aLhNhvNh8t666asowoncArn4MIVtOAO2tABAhN4hld4s3LrxXq3PhajFavcOYY/sD5/AM0fk8k=</latexit> I ω Reconstructed Edited DINO Anchor ⚓ Select Anchor <latexit sha1_base64="HZkHcstl+EoKiimdkgCoFKLvUDA=">AAAB73icbVBNS8NAEJ34WetX1aOXxSJ4KolI9Vj0oMcK9gPaUDbbTbt0s4m7E6GE/gkvHhTx6t/x5r9x2+agrQ8GHu/NMDMvSKQw6Lrfzsrq2vrGZmGruL2zu7dfOjhsmjjVjDdYLGPdDqjhUijeQIGStxPNaRRI3gpGN1O/9cS1EbF6wHHC/YgOlAgFo2il9m2vi0OOtFcquxV3BrJMvJyUIUe9V/rq9mOWRlwhk9SYjucm6GdUo2CST4rd1PCEshEd8I6likbc+Nns3gk5tUqfhLG2pZDM1N8TGY2MGUeB7YwoDs2iNxX/8zophld+JlSSIldsvihMJcGYTJ8nfaE5Qzm2hDIt7K2EDammDG1ERRuCt/jyMmmeV7xqpXp/Ua5d53EU4BhO4Aw8uIQa3EEdGsBAwjO8wpvz6Lw4787HvHXFyWeO4A+czx/xOY/t</latexit> G ω <latexit sha1_base64="HZkHcstl+EoKiimdkgCoFKLvUDA=">AAAB73icbVBNS8NAEJ34WetX1aOXxSJ4KolI9Vj0oMcK9gPaUDbbTbt0s4m7E6GE/gkvHhTx6t/x5r9x2+agrQ8GHu/NMDMvSKQw6Lrfzsrq2vrGZmGruL2zu7dfOjhsmjjVjDdYLGPdDqjhUijeQIGStxPNaRRI3gpGN1O/9cS1EbF6wHHC/YgOlAgFo2il9m2vi0OOtFcquxV3BrJMvJyUIUe9V/rq9mOWRlwhk9SYjucm6GdUo2CST4rd1PCEshEd8I6likbc+Nns3gk5tUqfhLG2pZDM1N8TGY2MGUeB7YwoDs2iNxX/8zophld+JlSSIldsvihMJcGYTJ8nfaE5Qzm2hDIt7K2EDammDG1ERRuCt/jyMmmeV7xqpXp/Ua5d53EU4BhO4Aw8uIQa3EEdGsBAwjO8wpvz6Lw4787HvHXFyWeO4A+czx/xOY/t</latexit> G ω <latexit sha1_base64="HZkHcstl+EoKiimdkgCoFKLvUDA=">AAAB73icbVBNS8NAEJ34WetX1aOXxSJ4KolI9Vj0oMcK9gPaUDbbTbt0s4m7E6GE/gkvHhTx6t/x5r9x2+agrQ8GHu/NMDMvSKQw6Lrfzsrq2vrGZmGruL2zu7dfOjhsmjjVjDdYLGPdDqjhUijeQIGStxPNaRRI3gpGN1O/9cS1EbF6wHHC/YgOlAgFo2il9m2vi0OOtFcquxV3BrJMvJyUIUe9V/rq9mOWRlwhk9SYjucm6GdUo2CST4rd1PCEshEd8I6likbc+Nns3gk5tUqfhLG2pZDM1N8TGY2MGUeB7YwoDs2iNxX/8zophld+JlSSIldsvihMJcGYTJ8nfaE5Qzm2hDIt7K2EDammDG1ERRuCt/jyMmmeV7xqpXp/Ua5d53EU4BhO4Aw8uIQa3EEdGsBAwjO8wpvz6Lw4787HvHXFyWeO4A+czx/xOY/t</latexit> G ω <latexit sha1_base64="HZkHcstl+EoKiimdkgCoFKLvUDA=">AAAB73icbVBNS8NAEJ34WetX1aOXxSJ4KolI9Vj0oMcK9gPaUDbbTbt0s4m7E6GE/gkvHhTx6t/x5r9x2+agrQ8GHu/NMDMvSKQw6Lrfzsrq2vrGZmGruL2zu7dfOjhsmjjVjDdYLGPdDqjhUijeQIGStxPNaRRI3gpGN1O/9cS1EbF6wHHC/YgOlAgFo2il9m2vi0OOtFcquxV3BrJMvJyUIUe9V/rq9mOWRlwhk9SYjucm6GdUo2CST4rd1PCEshEd8I6likbc+Nns3gk5tUqfhLG2pZDM1N8TGY2MGUeB7YwoDs2iNxX/8zophld+JlSSIldsvihMJcGYTJ8nfaE5Qzm2hDIt7K2EDammDG1ERRuCt/jyMmmeV7xqpXp/Ua5d53EU4BhO4Aw8uIQa3EEdGsBAwjO8wpvz6Lw4787HvHXFyWeO4A+czx/xOY/t</latexit> G ω <latexit sha1_base64="HZkHcstl+EoKiimdkgCoFKLvUDA=">AAAB73icbVBNS8NAEJ34WetX1aOXxSJ4KolI9Vj0oMcK9gPaUDbbTbt0s4m7E6GE/gkvHhTx6t/x5r9x2+agrQ8GHu/NMDMvSKQw6Lrfzsrq2vrGZmGruL2zu7dfOjhsmjjVjDdYLGPdDqjhUijeQIGStxPNaRRI3gpGN1O/9cS1EbF6wHHC/YgOlAgFo2il9m2vi0OOtFcquxV3BrJMvJyUIUe9V/rq9mOWRlwhk9SYjucm6GdUo2CST4rd1PCEshEd8I6likbc+Nns3gk5tUqfhLG2pZDM1N8TGY2MGUeB7YwoDs2iNxX/8zophld+JlSSIldsvihMJcGYTJ8nfaE5Qzm2hDIt7K2EDammDG1ERRuCt/jyMmmeV7xqpXp/Ua5d53EU4BhO4Aw8uIQa3EEdGsBAwjO8wpvz6Lw4787HvHXFyWeO4A+czx/xOY/t</latexit> G ω Concat. ... Padding Recon. Target <latexit sha1_base64="HZkHcstl+EoKiimdkgCoFKLvUDA=">AAAB73icbVBNS8NAEJ34WetX1aOXxSJ4KolI9Vj0oMcK9gPaUDbbTbt0s4m7E6GE/gkvHhTx6t/x5r9x2+agrQ8GHu/NMDMvSKQw6Lrfzsrq2vrGZmGruL2zu7dfOjhsmjjVjDdYLGPdDqjhUijeQIGStxPNaRRI3gpGN1O/9cS1EbF6wHHC/YgOlAgFo2il9m2vi0OOtFcquxV3BrJMvJyUIUe9V/rq9mOWRlwhk9SYjucm6GdUo2CST4rd1PCEshEd8I6likbc+Nns3gk5tUqfhLG2pZDM1N8TGY2MGUeB7YwoDs2iNxX/8zophld+JlSSIldsvihMJcGYTJ8nfaE5Qzm2hDIt7K2EDammDG1ERRuCt/jyMmmeV7xqpXp/Ua5d53EU4BhO4Aw8uIQa3EEdGsBAwjO8wpvz6Lw4787HvHXFyWeO4A+czx/xOY/t</latexit> G ω <latexit sha1_base64="HZkHcstl+EoKiimdkgCoFKLvUDA=">AAAB73icbVBNS8NAEJ34WetX1aOXxSJ4KolI9Vj0oMcK9gPaUDbbTbt0s4m7E6GE/gkvHhTx6t/x5r9x2+agrQ8GHu/NMDMvSKQw6Lrfzsrq2vrGZmGruL2zu7dfOjhsmjjVjDdYLGPdDqjhUijeQIGStxPNaRRI3gpGN1O/9cS1EbF6wHHC/YgOlAgFo2il9m2vi0OOtFcquxV3BrJMvJyUIUe9V/rq9mOWRlwhk9SYjucm6GdUo2CST4rd1PCEshEd8I6likbc+Nns3gk5tUqfhLG2pZDM1N8TGY2MGUeB7YwoDs2iNxX/8zophld+JlSSIldsvihMJcGYTJ8nfaE5Qzm2hDIt7K2EDammDG1ERRuCt/jyMmmeV7xqpXp/Ua5d53EU4BhO4Aw8uIQa3EEdGsBAwjO8wpvz6Lw4787HvHXFyWeO4A+czx/xOY/t</latexit> G ω ⃜ ⚓ ⚓ ⚓ ⚓ ⚓ ⚓ Fig. 2: Overview of our OSVE framework for one-step video editing. (a) Training: We train an inversion encoder using a novel dataset of structurally aligned image pairs generated via Prompt Perturbation. The encoder learns to predict an ini- tial noise that supports both faithful reconstruction and structure-aware editing. (b) Inference: Our Unified-Frame Editing (UFE) method inverts all video frames, concate- nates their latents into a single map, and processes them in one pass. This enables cross-frame attention, ensuring temporal consistency. An anchor-based sliding window scales the approach to long videos while maintaining global coherence. 3.1 Structure-Aware One-Step Inversion Recent one-step diffusion models have dramatically accelerated image and video generation, yet their application to editing remains severely limited. The core difficulty is structure preservation: an edit should maintain the spatial lay- out, object pose, and geometric relationships of the source while modifying only the intended semantics. Structure preservation has been the central concern in diffusion-based editing from the outset, as methods such as Prompt-to-Prompt (P2P) [12], ControlNet [73], and PnP [59] were all designed primarily to keep the source structure intact during semantic modification. Without it, edits suffer from spatial misalignment and temporal inconsistency (Figure 1(b)), rendering them practically unusable regardless of semantic quality. While multi-step diffu- sion models have largely addressed this through iterative inversion and control, one-step models lack such mechanisms, making structure preservation the central bottleneck that must be solved before one-step editing becomes practical. To understand why these existing solutions do not transfer, consider how they work in the multi-step setting. DDIM inversion [53] traces the denoising trajec- tory backward to recover a latent that approximately reconstructs the source, and the editing techniques listed above then intervene at selected steps along this trajectory. Crucially, these methods rely on the iterative nature of the pro- cess: each denoising step provides a corrective feedback loop that absorbs small perturbations introduced by the control signal, gradually steering the generation while keeping the structure intact. One-step models, however, collapse the entire 6H. Lim and G.-M. Park trajectory into a single forward pass, which creates two coupled difficulties. First, there is no multi-step trajectory to reverse, so standard DDIM-style inversion is inapplicable. Second, and more fundamentally, iterative control mechanisms lose the corrective loop they depend on. In iterative generation, the control signal is distributed across many denoising steps, and each step applies only a small perturbation that subsequent steps can absorb and correct. When the same con- trol is concentrated into a single pass, its full influence is applied at once with no opportunity for subsequent correction, causing the control signal to dominate the generation rather than gently guide it. The resulting “over-steering” artifacts are shown in Figure 4(b, c, e, f). This analysis points to a clear design principle: in a one-step setting, struc- ture preservation cannot be imposed during generation but should be encoded before generation begins. If the initial noise latent itself already carries the struc- tural information of the source frame, the one-step generator can preserve that structure in its single forward pass without any auxiliary control mechanism. We realize this principle by training an inversion encoder I φ (·) that directly maps a source frame to an initial noise latent encoding the spatial structure needed for both faithful reconstruction and meaningful editing. Concretely, given a VAE-encoded latent z 0 from a pre-trained autoencoder [19, 45] and a prompt embedding c = E(y) for a given text prompt y (where E is the frozen text encoder, e.g., CLIP [43]), our encoder predicts an initial noise latent ˆz T =I φ (z 0 ,c), where T denotes the terminal diffusion timestep (i.e., 999). Fol- lowing previous inversion methods [12,36,44,53,58,59], our objective is to predict a noise latent ˆz T that simultaneously satisfies two criteria: reconstruction and editability. To endow the encoder with structure-aware editability, we proceed in two stages: 1) Dataset generation with prompt perturbation: We cu- rate a dataset of structurally similar image pairs that share pose and layout but differ visually (Figure 2(a.1)). 2) Training with Structure-Aware Editing loss: We train the encoder using a newly proposed structure-aware editing loss that simulates editing and constrains the output to align with the paired target (Figure 2(a.2)). Dataset Generation with Prompt Perturbation. Training the inversion encoder to encode structure before generation requires data that teaches what “structure preservation under semantic change” looks like. Specifically, we need pairs of images that share spatial layout but differ in appearance, so that the encoder learns to produce latents capturing the shared structure while remaining agnostic to the varying semantics. To generate such a dataset DB = (z 0 ,z 0 edit ,c,c edit ), a natural approach would be to apply an off-the-shelf image editor to each source image. However, such editors typically rely on different generative backbones, and their outputs may fall outside the reachable manifold of the target one-step generator G θ . Training I φ on these distributionally mismatched targets leads to inconsistent gradients, as the encoder is asked to predict latents for images that G θ could never have produced in the first place. Table 1 confirms this: an encoder trained OSVE: One Step Video Editing with One Step Diffusion Models7 Table 1: Quantitative comparison of reconstruction quality for different in- version strategies. PSNR↑ LPIPS↓ MSE↓ SSIM↑ DDIM [53] 13.87 0.525 0.045 0.446 Off-the-shelf 2 17.54 0.400 0.032 0.543 Ours19.160.2610.0140.624 Fig. 3: Structural (DINO) and semantic (CLIP) distances as λ noise varies. on off-the-shelf 2 editing pairs yields notably lower reconstruction quality than ours across all metrics, indicating that the mismatched supervision corrupts the encoder’s overall latent representation. We also report DDIM 50-step inver- sion [53] as a baseline, which performs worst due to the fundamental inversion– generation mismatch discussed in Section 1. To avoid this mismatch, we propose a simple yet effective alternative termed prompt perturbation, which ensures that every target image is generator-aligned, i.e., lies within G θ ’s output distribution by construction. Concretely, we first sample a text prompt from a corpus (LAION [50], JourneyDB [56]) and a random noise vector z T ∼ N(0,I). Let c = E(y) be the text embedding of the sampled prompt y. We then construct a perturbed embedding as: c edit = c + λ noise · η, η ∼N(0,I),(1) where λ noise controls the perturbation intensity. Because both the source image G θ (z T ,c) and the edited image G θ (z T ,c edit ) are generated by the same frozen generator from the same initial noise, the resulting pairs are sampled from G θ ’s reachable output distribution and tend to share a similar structural layout while primarily differing in appearance. This generator-aligned supervision provides a well-posed training signal for I φ , which must predict latents that G θ can faithfully decode. We empirically study the effect of λ noise on both structure and semantics (Figure 3). Starting from the base prompt “a photo of a dog”, we vary λ noise ∈ 0.0, 0.1,..., 0.5 and, for each setting, generate 100 image pairs using the same initial noise z T . For each pair, we measure (i) a structural distance using DINO [4], and (i) a semantic distance using CLIP [43]. As λ noise increases, the semantic dis- tance grows, while for appropriately small noise (e.g., λ noise = 0.1) the structural distance remains low and stable, confirming that small perturbations preserve layout while inducing meaningful appearance changes. Guided by this analysis, we set λ noise = 0.1 in all experiments (see Supplementary B and G for training results of other parameters). 2 We use Google Gemini’s Nano Banana as a representative off-the-shelf image editor with default recommended settings. 8H. Lim and G.-M. Park Fig. 4: Comparison of structure-preserving methods for one-step editing. (a) Our method, trained with the proposed L sae , maintains structural integrity. (d) With- out L sae , our model fails, leading to structural collapse. (b, c, e, f) Similarly, existing control methods often produce degraded results, demonstrating their unsuitability for one-step generation. SA and CA denote the replacement of self-attention and cross- attention, respectively. Given (c,c edit ) and a shared z T , we obtain the reconstruction and editing targets as z 0 = G θ (z T ,c) |z Reconstruction Target , z 0 edit = G θ (z T ,c edit ) | z Editing Target . (2) This process yields a large dataset in which most pairs exhibit strong structural similarity but differ in appearance, effectively simulating the desired editing behavior. Training with Structure-Aware Editing Loss. The dataset provides pairs that share structure but differ in semantics, each pair generated from the same noise z T under the original and perturbed prompts. We translate this into a training objective that forces the inversion encoder to produce a latent ˆz T = I φ (z 0 ,c) serving both reconstruction and structure-preserving editing. Our loss is: L total = z 0 − G θ (I φ (z 0 ,c), c) 2 2 | z L mse + λ sae z 0 edit − G θ (I φ (z 0 ,c), c edit ) 2 2 | z L sae ,(3) where λ sae is the weight parameter (set to 1.0), z 0 = G θ (z T ,c), and z 0 edit = G θ (z T ,c edit ). The first term,L mse , ensures that the inverted latent allows G θ to accurately reconstruct the original image z 0 . Reconstruction alone, however, does not guar- antee that the same latent will preserve structure when the prompt changes. The second term, the Structure-Aware Editing (SAE) loss L sae , directly ad- dresses this: it requires that the same inverted latent ˆz T , when decoded with the perturbed prompt c edit , produces an output matching the pre-generated tar- get z 0 edit . Because z 0 edit was generated from the same noise z T as z 0 , it shares the source structure by construction, so matching it teaches the encoder to pro- duce latents whose structural content is invariant to prompt changes. Figure 4 OSVE: One Step Video Editing with One Step Diffusion Models9 Fig. 5: Mechanism of Unified-Frame Editing for Temporal Consistency. (a) Frame-Wise Editing: Processing frames individually restricts attention to within each frame, preventing information sharing and causing temporal incoherence. (b) Unified- Frame Editing (UFE): Our method concatenates all frame latents, allowing the atten- tion mechanism to operate globally. This enables features to be matched and aligned across all frames, enforcing temporal consistency throughout the edit. validates this design: without L sae , the model suffers from structural collapse (d), while existing control methods such as attention injection (b, c, e, f) simi- larly fail, confirming that structure preservation must come from the latent itself rather than from external control during generation. For implementation, we initialize the inversion encoder I φ with the pre- trained one-step model G θ , leveraging components from its U-Net. Detailed training settings and visual and statistical analyses of the latents ˆz T produced by I φ are provided in Supplementary C. 3.2 Unified-Frame Editing While the inversion encoder from Section 3.1 is effective for single images, ap- plying it frame-wise to video does not preserve temporal consistency. Most prior video editing methods enforce consistency by injecting guidance signals into in- termediate features or attention maps during multi-step denoising [6,9,17,29,61, 63,65]. However, as discussed in Section 3.1, such iterative guidance mechanisms are inapplicable in the one-step regime (Figure 4). This motivates a fundamentally different strategy: rather than manipulating internal attention, we encourage cross-frame communication by modifying only the input latent representation. We propose Unified-Frame Editing (UFE), an input-level latent unification scheme in which inverted latents from multi- ple frames are spatially concatenated and processed simultaneously in a single forward pass (Figure 2(b)). Concretely, given a video sequence of K frames V 0 K = (z 0 0 ,...,z 0 K−1 ), each frame is independently inverted via our encoder: ˆz T k =I φ (z 0 k , c), k = 0,...,K− 1.(4) We denote the set of inverted latents as ˆ V T K = (ˆz T 0 ,..., ˆz T K−1 ). Let ˆz T k ∈R C×H×W denote the inverted latent for the k-th frame. The unified latent map Z T UFE is 10H. Lim and G.-M. Park formed by concatenating all K latents along the width dimension: Z T UFE = Concat(ˆz T 0 , ˆz T 1 ,..., ˆz T K−1 , axis = W).(5) The resulting map Z T UFE has dimensionsR C×H×(W·K) and is passed to the one- step generator G θ (·) with the editing prompt c edit . This is feasible because the U-Net operates on spatial features and naturally accepts inputs of arbitrary size [32, 45, 46]. Crucially, this design allows the generator’s self-attention to function globally across all frames, leveraging its ability to align corresponding features [57] and thereby enforce temporal consistency without any external guidance module. Figure 5 visualizes this effect: unlike frame-wise editing where information remains isolated, a token representing a specific feature (e.g., the tiger’s nose) in one frame can directly attend to its counterpart in other frames, ensuring that edits are applied consistently across the entire sequence. Sliding Window with Anchor Frame. While UFE is effective, its VRAM requirement and inference time scale with the number of frames (Supplementary H), making it computationally challenging for long videos. To address this, we extend UFE with a sliding window mechanism guided by a global anchor frame. Figure 2(b) illustrates the pipeline of our method. The core idea is to process the video in manageable chunks while ensuring both local and global temporal consistency. The sliding window provides local coherence, while the anchor frame enforces a consistent global style and content reference across all windows. First, we identify a representative anchor frame by selecting the medoid in the DINO [4] feature space. Its inverted latent, ˆz T A , serves as a global anchor. We validate this choice against alternative strategies, including the use of multiple anchors for videos with scene transitions or dynamic content changes, in Supplementary D. We then process the video using an overlapping sliding window approach. For each window of w inverted latents, we prepend the anchor latent ˆz T A before feeding the concatenated map to the generator G θ . After generation, the out- put corresponding to the anchor is discarded, and we extract only the central s frames from the window’s output. These segments are then seamlessly stitched together to form the final edited video. This anchor-based sliding window strat- egy effectively prevents temporal drift by ensuring that each segment is coherent both locally and globally. We set the window size as w = 7 and a stride of s = 5 through ablation (Supplementary E). A detailed, step-by-step formulation of this process is provided in Supplementary F. 4 Experiments 4.1 Experimental Setup We evaluate our framework across three video editing settings. The first, Multi- to-Multi, uses multi-step diffusion models for both inversion and editing. We compare against five baselines: FLATTEN [6], TokenFlow [42], FRESCO [65], RAVE [17], and COVE [61], all built on Stable Diffusion 1.5 [45]. The second, Multi-to-One, performs video inversion via multi-step DDIM while editing OSVE: One Step Video Editing with One Step Diffusion Models11 Table 2: Performance comparison on short (20 frames) videos. Framework MethodSC BC TF MS AQ IQBQSFPS FLATTEN ‡ 0.965 0.970 0.964 0.972 0.625 0.6390.6110.072 TokenFlow ‡ 0.983 0.976 0.985 0.991 0.668 0.680 0.6630.075 Multi-to-Multi FRESCO ‡ 0.978 0.974 0.973 0.991 0.649 0.7290.6740.078 RAVE ‡ 0.982 0.976 0.975 0.986 0.637 0.6950.6530.091 COVE ‡ 0.983 0.976 0.984 0.989 0.645 0.6550.6390.061 Prompt Replacement † 0.921 0.946 0.948 0.979 0.593 0.5620.5484.732 Prompt-to-Prompt † 0.915 0.945 0.965 0.978 0.587 0.5660.5480.898 Multi-to-One ControlNet (Depth) † 0.968 0.961 0.970 0.984 0.658 0.6730.6460.578 ControlNet (Canny) † 0.960 0.969 0.966 0.983 0.582 0.6620.6030.578 Plug-and-Play † 0.953 0.971 0.995 0.995 0.486 0.2200.3450.398 One-to-One OSVE (Ours) † 0.983 0.977 0.978 0.991 0.678 0.7030.67915.625 ‡ Multi-Step Diffusion Model (SD1.5). † One-Step Diffusion Model (DMD2). Table 3: Performance comparison on long (90 frames) videos. Framework MethodSC BC TF MS AQ IQBQSFPS FLATTEN ‡ 0.933 0.960 0.965 0.975 0.648 0.6360.6150.052 TokenFlow ‡ 0.974 0.975 0.986 0.990 0.622 0.668 0.6330.084 Multi-to-Multi FRESCO ‡ 0.956 0.968 0.987 0.992 0.640 0.7020.6550.087 RAVE ‡ 0.962 0.963 0.978 0.983 0.665 0.6920.6590.102 COVE ‡ 0.955 0.964 0.985 0.988 0.651 0.650 0.6330.041 Prompt Replacement † 0.841 0.912 0.970 0.975 0.561 0.5280.5034.732 Prompt-to-Prompt † 0.900 0.956 0.969 0.974 0.423 0.267 0.3280.898 Multi-to-One ControlNet (Depth) † 0.945 0.944 0.971 0.978 0.626 0.6520.6130.578 ControlNet (Canny) † 0.934 0.952 0.967 0.976 0.590 0.6600.5980.578 Plug-and-Play † 0.841 0.912 0.994 0.995 0.561 0.528 0.5090.398 One-to-One OSVE (Ours) † 0.958 0.965 0.978 0.989 0.670 0.7230.67715.793 ‡ Multi-Step Diffusion Model (SD1.5). † One-Step Diffusion Model (DMD2). in a single step. As no prior work exists for this configuration, we construct baselines by adapting image editing techniques: Prompt Replacement, Prompt- to-Prompt [12], ControlNet [73] with depth and Canny guidance, and Plug-and- Play [59]. The third is our proposed One-to-One framework, which employs the pre-trained encoder for single-step inversion and edits via Prompt Replacement. For the Multi-to-One and One-to-One settings, we use DMD2 [68] as the one-step diffusion backbone. Further details are provided in Supplementary A. For evaluation, we curated 60 videos from open platforms such as Pixabay and prior works, comprising 51 short videos (20 frames) and 9 long videos (90 frames). We designed five editing prompts per video—three for local edits (e.g., object change, deletion) and two for global edits (e.g., style transfer, background modification)—resulting in 255 short-video and 45 long-video evaluation pairs. Short videos are processed as a single window, while long videos use a sliding window of 7 frames with a stride of 5. All experiments were conducted on a single NVIDIA RTX 6000 Ada GPU. 12H. Lim and G.-M. Park Fig. 6: Qualitative comparison with baseline methods. 4.2 Quantitative Comparison We adopt VBench [14] metrics for quantitative evaluation, organized into two categories. Temporal consistency metrics include: Subject Consistency (SC), measuring subject identity preservation via DINO [4] features; Background Con- sistency (BC), assessing background stability via CLIP [43] features; Tempo- ral Flickering (TF), where higher indicates less flickers; and Motion Smooth- ness (MS), evaluating motion fluidity using a pre-trained video interpolation model [23]. Per-frame quality metrics include: Aesthetic Quality (AQ), scored by the LAION aesthetic predictor [21], and Imaging Quality (IQ), quantifying perceptual artifacts using MUSIQ [18]. We also introduce the Balanced Qual- OSVE: One Step Video Editing with One Step Diffusion Models13 Fig. 7: Qualitative results with OSVE across various editing tasks. ity Score (BQS): BQS = SC+BC+TF+MS 4 × AQ+IQ 2 , which balances both dimensions multiplicatively, penalizing methods that excel in only one. Tables 2 and 3 present the results for short and long videos, respectively. Our method, OSVE, achieves the highest BQS in both settings, demonstrating strong temporal consistency and per-frame quality simultaneously. In terms of efficiency, OSVE is approximately 171× faster for short videos and 155× faster for long videos compared to RAVE, the fastest Multi-to-Multi baseline. See Sup- plementary A for the speed comparison setup. 4.3 Qualitative Comparison Figure 6 provides a qualitative comparison. OSVE successfully modifies objects while preserving their underlying structure, whereas the Multi-to-One baselines (rows 7–11) frequently exhibit structural collapse or visual degradation. Figure 7 illustrates the versatility of OSVE across diverse editing tasks, including object replacement (columns 1, 3, and 5), human identity modification (column 2), artistic style transfer (columns 3 and 4), and prompt-guided object removal (column 4). 4.4 User Study We conducted a user study with 26 participants. Each participant viewed six editing methods applied to three randomly sampled videos from our datasets and rated each result on Temporal Consistency (TC) and Visual Quality (VQ) using a 5-point Likert scale. As shown in Table 4, OSVE achieves the highest scores on both dimensions, outperforming all baselines. 4.5 Ablation Study Inversion Encoder. Table 5 validates the Structure-Aware Editing (SAE) Loss on the PIE-Bench [16], using structure distance [59] and CLIP scores [43]. While SAE Loss alone improves structure preservation, combining it with the recon- struction loss (L mse ) yields the best performance in both structural integrity and editability, indicating a synergistic effect between the two objectives. 14H. Lim and G.-M. Park Table 4: User study results. TC: Temporal Consistency, VQ: Visual Quality (higher is better). MethodTC VQ FLATTEN2.59 2.56 TokenFlow3.683.08 FRESCO3.02 2.76 RAVE2.74 2.70 COVE3.59 3.41 OSVE (Ours)3.853.65 Table 5: Ablation study on encoder training. We evaluate the effect of each loss component on PIE-Bench. L mse L sae Struct. Dist. (↓) CLIP Whole (↑) CLIP Edit (↑) ✓ –0.08721.79719.884 –✓0.07422.34919.863 ✓0.06422.32920.416 Table 6: Ablation study of UFE on 90-frame videos. Sliding window AnchorSCBCTFMSAQIQBQS –0.9310.9480.9720.980 0.6800.7220.671 ✓–0.9540.9630.9770.9880.6710.7230.676 –✓0.9430.9500.9760.9850.6720.7220.669 ✓0.9580.9650.9780.9890.6700.7230.677 Unified Frame Editing. Table 6 presents the ablation study of Unified-Frame Editing (UFE). Both the sliding window and anchor-frame strategies individually improve temporal consistency (SC, BC, TF, MS), and their combination yields the strongest results. We observe a marginal decrease in AQ when applying UFE, likely due to processing the expanded latent representation. However, this minor trade-off is perceptually negligible compared to the substantial reduction in flickering and inconsistencies. 5 Discussion We build OSVE on a one-step T2I backbone because current one-step T2V gen- erators, despite recent progress in T2V distillation [26, 66, 75], still suffer from severe temporal artifacts and blurriness that make them unsuitable as a reli- able editing backbone (see Supplementary G). In contrast, mature one-step T2I backbones have been reported to even surpass their multi-step teachers in gen- eration quality [68], providing a clean testbed to evaluate our core contributions independently of backbone limitations. The generative landscape is rapidly expanding toward T2V, T2AV, and other multimodal architectures, and one-step distillation of these models is an active area of research [8,22,26,28,74]. As their fidelity improves, extending one-step editing to these backbones will become both feasible and important. We believe the core principles behind OSVE, namely encoding structure into the initial latent and enforcing cross-frame consistency at the input level, are not specific to T2I and can generalize across modalities. OSVE: One Step Video Editing with One Step Diffusion Models15 6 Conclusion We presented OSVE, a framework that advances text-guided video editing by introducing a novel One-to-One pipeline. OSVE unlocks the potential of one-step diffusion models for this task, overcoming the speed limitations of conventional multi-step methods. Our two core contributions, a fast learnable inversion en- coder trained with SAE loss to preserve structure and the UFE mechanism to enforce temporal coherence, resolve the key bottlenecks that previously hindered this approach. Our experiments show that OSVE achieves competitive quality and consistency at a fraction of the computational cost, making real-time video editing practical and accessible. Acknowledgements This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2026-25480253), and in part by Korea Planning & Evaluation Institute of Industrial Technology (KEIT) grant funded by the Korea government (MOTIE) (RS-2024-00444344), and by the “Advanced GPU Utilization Support Program” funded by the Government of the Republic of Korea (Ministry of Science and ICT). Table of Contents A. Experimental Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .16 A.1 Multi-to-Multi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .16 A.2 Multi-to-One . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 A.3 One-to-One . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .19 A.4 Speed Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 B. Ablation on Prompt Perturbation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 C. Encoder Training and Analysis of Inverted Latents . . . . . . . . . . . . . .20 C.1 Training Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 C.2 Visualization and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 D. Anchor Ablations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .21 D.1 Anchor selection strategies ablation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 D.2 Multi-Anchor Ablation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 E. Ablation on Sliding Window and Anchor . . . . . . . . . . . . . . . . . . . . . . . . . 23 F. Unified-Frame Editing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .24 G. Additional Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 G.1 Detailed Ablation on λ noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 G.2 Discussion on One-Step T2V Generator Quality . . . . . . . . . . . . . . . . .26 H. GPU Memory Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 I. User Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 16H. Lim and G.-M. Park A Experimental Details In this section, we provide comprehensive implementation details for all methods evaluated in our paper. All training and inference experiments are conducted on a single NVIDIA RTX 6000 Ada graphics card (48 GB VRAM). Unless stated otherwise, we follow the hyperparameter settings reported in each respective paper. For cases where specific hyperparameters are not explicitly declared, we adopt the default values provided in the official open-source implementations. To ensure fair comparison across all methods, we keep the input video reso- lution and frame count consistent throughout our experiments. All input videos are resized to 512× 512 resolution. We use the same set of source videos and editing prompts for every method under the same experimental setting. A.1 Multi-to-Multi The Multi-to-Multi setting refers to the conventional video editing pipeline where both the inversion and sampling processes require multiple diffusion steps. In this setting, we adopt Stable Diffusion v1-5 3 as the pre-trained multi-step diffusion model for all methods. For the inversion process, we uniformly apply the DDIM Inversion scheduler [53] with 50 inversion steps across all methods, which reconstructs the latent noise representation from each source video frame. We describe the specific configurations for each baseline method below. FLATTEN [6]. We use the official PyTorch implementation of FLATTEN 4 , which leverages optical flow-guided attention to improve temporal consistency in video editing. Following the settings from the original paper, we set the classifier- free guidance scale to 20 and use 50 DDIM sampling steps for the denoising pro- cess. FLATTEN modifies the attention mechanism in the U-Net by incorporating optical flow information, which guides the attention maps to be temporally co- herent across frames. We use the default optical flow estimation model provided in the official repository. TokenFlow [9]. We use the official PyTorch implementation of TokenFlow 5 . To- kenFlow enforces temporal consistency by propagating diffusion features across frames based on inter-frame correspondences. Following the paper’s methodol- ogy, we utilize the Plug-and-Play (PnP) editing framework [59] as the underlying editing mechanism. The classifier-free guidance scale is set to 7.5, and we use 50 DDIM sampling steps. TokenFlow first identifies keyframes, computes nearest- neighbor correspondences between them, and then propagates the edited tokens to all other frames to maintain visual coherence. FRESCO [65]. We use the official PyTorch implementation of FRESCO 6 , which introduces spatial-temporal correspondences for zero-shot video editing. As rec- 3 https://huggingface.co/runwayml/stable-diffusion-v1-5 4 https://github.com/yrcong/flatten 5 https://github.com/omerbt/TokenFlow 6 https://github.com/williamyang1991/FRESCO OSVE: One Step Video Editing with One Step Diffusion Models17 ommended in the paper, we employ ControlNet [73] with HED (Holistically- Nested Edge Detection) conditioning 7 to provide structural guidance during the editing process. We use 20 DDIM sampling steps and set the classifier-free guid- ance scale to 7.5. We select one keyframe every 8 frames for the editing process. After editing the keyframes, we use EbSynth 8 , a patch-based video stylization tool, to propagate the edits from keyframes to all remaining frames and produce the final edited video. All other parameters, including the spatial-temporal at- tention fusion weights, follow the default settings in the official implementation. RAVE [17]. We use the official PyTorch implementation of RAVE 9 , which achieves temporally consistent editing by rearranging video frames into a grid layout and processing them simultaneously. Following the settings described in the paper, we employ a ControlNet model conditioned on depth maps 10 to pre- serve the geometric structure of the source video. The grid size is set to 3×3 (i.e., 9 frames are arranged in a single grid image), the classifier-free guidance scale is set to 7.5, and we perform 50 DDIM sampling steps. Depth maps are estimated from the source video frames using the default depth estimation model provided in the repository. COVE [61]. We use the official PyTorch implementation of COVE 11 , which ex- ploits video-level correspondences to guide the diffusion process. Following the settings from the paper, we set the classifier-free guidance scale to 7.5 and the DIFT (Diffusion Features) up-sampling block index to 2, which controls the gran- ularity of the extracted diffusion features used for correspondence computation. The sliding window size for computing inter-frame correspondences is set to 7, the correspondence-guidance scale (which controls the strength of the correspon- dence constraint during sampling) is set to 3, and the token merging ratio is set to 50%, meaning half of the tokens are merged based on their correspondence similarities to reduce computational cost while maintaining temporal coherence. A.2 Multi-to-One The Multi-to-One setting represents a hybrid approach where the inversion process still requires multiple diffusion steps, but the sampling (generation) is performed in a single step using a distilled one-step diffusion model. In this setting, we use DMD2 [68] as our pre-trained one-step diffusion model, which is a distribution matching distillation model capable of generating high-quality images in a single forward pass. For the inversion process, we apply the DDIM Inversion scheduler [53] with 50 inversion steps to obtain the latent noise rep- resentation from each source frame, identical to the Multi-to-Multi setting. All inference in this setting is performed in a single denoising step using DMD2. 7 https://huggingface.co/lllyasviel/sd-controlnet-hed 8 https://github.com/jamriska/ebsynth 9 https://github.com/RehgLab/RAVE 10 https://huggingface.co/lllyasviel/sd-controlnet-depth 11 https://github.com/wangjiangshan0725/COVE 18H. Lim and G.-M. Park Below, we describe the specific configuration of each baseline method adapted to this setting. Prompt Replacement. This is the simplest baseline method, which performs editing by directly replacing the source text prompt with the target text prompt during the one-step sampling process. For example, the source prompt “A photo of a dog” is replaced with the target prompt “A photo of a cat,” while the inverted latent noise is kept unchanged. No additional structural guidance or attention manipulation is applied. This method serves as a lower bound to evaluate the effectiveness of more sophisticated editing techniques. Prompt-to-Prompt [12]. We use the official PyTorch implementation of Prompt- to-Prompt 12 , adapted for the one-step sampling setting. Prompt-to-Prompt achieves localized editing by manipulating the cross-attention maps within the U-Net dur- ing the denoising process. Specifically, the cross-attention maps corresponding to the source prompt tokens are injected into the denoising process with the target prompt, allowing the model to preserve the spatial layout of the source image while applying the desired edit. In our evaluation, we inject only the cross- attention maps and deliberately exclude the self-attention maps, as we empir- ically observed that injecting self-attention maps into the one-step generation pipeline leads to significant degradation in output image quality (see Figure 4 in the main paper). ControlNet (Depth) [73]. We employ a pre-trained ControlNet model con- ditioned on depth maps 13 to provide structural guidance during the one-step sampling process. Depth maps are extracted from each source video frame and fed into the ControlNet branch, which injects depth-aware features into the U- Net to preserve the spatial structure of the original scene. We set the ControlNet conditioning scale (which controls the strength of the structural guidance) to 0.8. This value was chosen to balance between preserving the source structure and allowing sufficient flexibility for the target prompt to take effect. ControlNet (Canny) [73]. Similarly, we use a pre-trained ControlNet model conditioned on Canny edge maps 14 as another structural guidance baseline. Canny edges are detected from each source video frame and used as condition- ing input. The ControlNet conditioning scale is set to 0.8, consistent with the depth-conditioned variant. Compared to depth conditioning, Canny edge condi- tioning provides finer-grained structural constraints based on edge boundaries, which can be advantageous for preserving sharp object boundaries but may also restrict the editing flexibility in some cases. Plug-and-Play [59]. We use the official PyTorch implementation of Plug-and- Play 15 , adapted for the one-step generation setting. Plug-and-Play achieves structure- preserving editing by injecting intermediate spatial features from the source im- 12 https://github.com/google/prompt-to-prompt 13 https://huggingface.co/lllyasviel/sd-controlnet-depth 14 https://huggingface.co/lllyasviel/sd-controlnet-canny 15 https://github.com/MichalGeyer/plug-and-play OSVE: One Step Video Editing with One Step Diffusion Models19 age’s denoising process into the target image’s generation process. Specifically, we extract the spatial features from the convolutional layers of the U-Net during the source image reconstruction and inject them at the corresponding layers dur- ing the target image generation, as described in the original paper. This feature injection constrains the spatial layout and structural details of the edited output to align with the source image while allowing semantic changes dictated by the target prompt. A.3 One-to-One The One-to-One setting represents the most efficient pipeline, where both the inversion and sampling processes are performed in a single step. This eliminates the computational overhead of iterative DDIM inversion entirely, enabling real- time or near-real-time video editing. OSVE (Ours). In this setting, we use our trained OSVE (One-Step Video Editing) encoder to perform single-step inversion. Given a source video frame, the OSVE encoder predicts the corresponding noise latent in a single forward pass, bypassing the need for iterative DDIM inversion. The predicted noise is then directly passed to the DMD2 [68] one-step diffusion model for sampling with the target text prompt. Since our inversion encoder is specifically designed and trained to preserve the structural and geometric information of the source image within the predicted noise representation, edits in this setting are performed using only the target text prompt without any additional structural guidance (e.g., ControlNet or attention injection). This design choice results in a highly streamlined editing pipeline that requires only two forward passes per frame: one through the OSVE encoder for inversion and one through the DMD2 model for generation. A.4 Speed Comparison For a fair and accurate speed comparison across all methods, we measure only the U-Net inference time, excluding the time required for VAE encoding and decoding operations. This is because the VAE computation cost is identical across all methods that share the same base diffusion model, and thus does not contribute to meaningful differences in computational efficiency. The reported time corresponds to the total inversion and inference time aggregated over all frames in a given video clip. To ensure consistency, all operations are executed sequentially rather than asynchronously or in parallel, so that the measured time accurately reflects the per-frame computational cost of each method. We repeat each timing measurement 10 times and report the average to mitigate variance from GPU scheduling and memory caching effects. B Ablation on Prompt Perturbation Table A presents an ablation study on the noise injection parameter for the SAE loss. The results indicate that a parameter of 0.1 achieves the lowest struc- 20H. Lim and G.-M. Park λ noise Structure Distance (↓) CLIP Whole (↑) CLIP Edit (↑) 0.00.08721.797 19.884 0.10.06422.32920.416 0.20.06822.314 20.406 0.30.07522.449 20.422 0.40.08522.531 20.402 0.50.10122.641 20.424 Table A: Ablation on noise injection parameter λ noise . ture distance, signifying the best preservation of the source content’s geometry. Conversely, as the injection parameter increases, the CLIP score also rises. This suggests a trade-off: a larger parameter pushes the inverted latent to be more editable and better aligned with the target prompt, but at the cost of structural fidelity. Given that our primary objective is to perform edits while faithfully preserving the original structure, we set this parameter to 0.1 for our main ex- periments. We note that the increase is primarily observed in CLIP Whole, while CLIP Edit remains largely stable across all values of λ noise (ranging from 20.40 to 20.42). This is because CLIP Edit measures similarity only within the edited region, which is already well-aligned with the target prompt at λ noise =0.1; fur- ther noise injection mainly affects the non-edited regions, which are captured by CLIP Whole. C Encoder Training and Analysis of Inverted Latents C.1 Training Details We trained the inversion encoder for 72 hours using a batch size of 6. The total number of trained samples is 426,000. The weight for the SAE loss, λ sae , was set to 1. For optimization, we used the AdamW [33] optimizer with a learning rate of 1× 10 −5 , betas of (0.9, 0.999), an epsilon of 1× 10 −8 , and a weight decay of 1× 10 −2 . The loss curves during training are visualized in Figures D and E. C.2 Visualization and Analysis of Inverted Latents We analyze the statistical and spectral properties of latents produced by our SAE-trained inversion encoder, comparing against pure Gaussian noise z ∼ N(0,I). Element-wise Gaussianity. Table B (top) reports normality diagnostics over 1000 inverted latents. The Q–Q R 2 of 0.999, KS statistic of 0.069, and near-zero skewness and kurtosis confirm that the marginal distribution is nearly indistin- guishable from N(0, 1), verifying compatibility with the diffusion prior. OSVE: One Step Video Editing with One Step Diffusion Models21 Table B: Analysis of inverted latents. Top: element-wise Gaussianity. Bottom: Low/High Ratio measuring spatial structure in the power spectrum (Section C.2). MetricGaussian Ours Q–Q R 2 1.000 0.999 KS Statistic0.000 0.069 Skewness0.000 −0.032 Kurtosis0.000 −0.083 Low/High Ratio–7.145 Fig. A: Visualization of inverted latents from our encoder. Despite being element-wise Gaussian (Table B), the latents exhibit visible low-frequency structure inherited from the source image. Spectral Analysis. Element-wise Gaussianity does not imply spatial indepen- dence. We compute the 2D power spectrum of each latent, average over channels, and extract radial profiles. We define the Low/High Ratio as the mean power for r < 0.15R divided by that for r > 0.50R. For spatially independent noise, this ratio is approximately 1. As shown in Table B, our latents exhibit a ratio of 7.145—low-frequency energy is approximately 7× larger than high-frequency energy, indicating that coarse spatial structure is implicitly encoded. This is also visually apparent in Figure A, where the inverted latent retains discernible low- frequency patterns from the source image despite appearing noise-like overall. Interpretation. Our inverted latents are element-wise Gaussian yet spatially structured: they appear as valid noise to the denoiser while encoding source- specific layout information in their spatial correlations. This dual property steers the denoising trajectory toward faithful reconstruction without violating the diffusion prior, explaining the superior inversion quality over DDIM reported in Table 1 of the main paper. D Ablation of Anchor D.1 Anchor selection strategies ablation To determine the optimal anchor selection strategy, we conducted an ablation study comparing our chosen method against six alternatives. The evaluated methods include: 1) Random, which selects a random frame; 2) First, which 22H. Lim and G.-M. Park Anchor method 90 Frames SCBCMSAQIQBQS Random0.9500.9630.9870.6700.7230.673 First0.9580.9620.9880.6700.7230.675 Pixel-Centroid0.9570.9620.9880.6700.7230.675 Pixel-Medoid0.9570.9640.9880.6700.7230.675 CLIP Top-10.9570.9630.9880.6700.7230.675 DINO-Medoid0.9580.9650.9890.6700.7230.676 Hybrid0.9560.9620.9880.6700.7230.675 Table C: Ablation of anchor Methods (90 Frames). always uses the initial frame; 3) Pixel-Centroid, which selects the frame closest to the pixel-wise mean of all frames; 4) Pixel-Medoid, which selects the frame minimizing the sum of pixel-wise distances to all other frames; 5) DINO-Medoid, which identifies the medoid in the DINO [4] feature space; 6) CLIP Top-1, which selects the frame with the highest CLIP [43] similarity to the source prompt; and 7) Hybrid, which considers both DINO and CLIP scores. The ablation was per- formed on 90-frame videos to capture the long-term effects of the anchor choice. As shown in Table C, the DINO-Medoid strategy consistently yielded the best performance, achieving the highest scores across our metrics. Notably, the Hy- brid strategy does not outperform DINO-Medoid despite combining more infor- mation. We conjecture that aggregating two heterogeneous scores introduces a weighting ambiguity: the selected frame becomes a compromise that does not maximize either criterion. Given that all non-random strategies already select reasonably representative frames and the absolute differences are small, a single well-chosen criterion (DINO feature similarity) proves sufficient and avoids the need for additional hyperparameter tuning to balance the two scores. D.2 Multi-Anchor Ablation We further demonstrate that our single-anchor framework can be naturally ex- tended to a multi-anchor strategy to handle more complex video scenarios. When a video contains only a single object or a single coherent scene, a single anchor is sufficient to maintain temporal consistency throughout the editing process. However, for videos composed of multiple distinct scenes—for example, a video that transitions from a cat walking to a dog walking (Figure B(a))—relying on a single anchor becomes problematic. In such cases, the anchor selected from one scene (e.g., the dog’s scene) fails to provide meaningful correspondences for frames belonging to a different scene (e.g., the cat’s scene), resulting in notice- able degradation in editing quality (Figure B(b)). To address this limitation, we introduce a simple yet effective multi-anchor strategy. We first apply an off- the-shelf scene change detector (e.g., PySceneDetect 16 ) to segment the input video into individual scenes. We then independently select an anchor frame for 16 https://github.com/Breakthrough/PySceneDetect OSVE: One Step Video Editing with One Step Diffusion Models23 Fig. B: Overview of the multi-anchor strategy. (a) A multi-scene video example (dog to cat). (b) When using only a single anchor derived from the dog’s scene, the editing quality of the cat’s region degrades noticeably (see red circles). (c) By employing mul- tiple anchors, our method preserves visual consistency for each object across different scenes. each detected scene. During inference, anchors are dynamically swapped at scene boundaries so that each segment is always paired with its corresponding anchor. As shown in Figure B(c), this strategy enables our method to preserve per-scene consistency without compromising editing quality across scene transitions. E Ablation on Sliding Window and Anchor In this section, we conduct an ablation study on the sliding window size and stride to find the optimal balance between editing quality and processing speed. As shown in Tables D and E, a window size of 7 and a stride of 5 (with an anchor) provides a strong trade-off, which we adopt for our main experiments. We note that SC, BC, and MS consistently improve as the window size in- creases, confirming that larger windows provide stronger temporal context. In contrast, AQ exhibits a slight decrease (e.g., from 0.680 at w=1 to 0.670 at w=7 in Table E). This is because tiling more frames into a single grid reduces the effective per-frame resolution seen by the UNet, which marginally affects per- frame aesthetic quality. However, the magnitude of this decrease is negligible 24H. Lim and G.-M. Park Window sizew Stride SCBC MS AQIQ BQS FPS 11 0.931 0.948 0.980 0.680 0.723 0.668 20.925 31 0.950 0.962 0.988 0.676 0.723 0.676 11.646 5 1 0.952 0.963 0.988 0.674 0.722 0.676 7.068 3 0.952 0.962 0.987 0.673 0.722 0.675 17.409 7 1 0.954 0.964 0.988 0.670 0.722 0.674 4.769 3 0.954 0.963 0.988 0.671 0.722 0.674 12.354 50.9540.9630.9880.6720.7230.67518.140 9 1 0.955 0.963 0.989 0.669 0.723 0.674 3.351 3 0.954 0.962 0.988 0.670 0.723 0.674 9.203 5 0.954 0.965 0.988 0.669 0.723 0.674 13.862 7 0.955 0.964 0.988 0.670 0.723 0.674 17.599 Table D: Ablation on Unified-Frame Editing without anchor on 90-frame videos. Window sizew Stride SCBC MS AQIQ BQS FPS 11 0.943 0.950 0.985 0.680 0.722 0.669 15.332 31 0.958 0.966 0.988 0.670 0.723 0.679 8.698 5 1 0.957 0.965 0.988 0.670 0.723 0.678 5.644 3 0.958 0.963 0.988 0.670 0.723 0.677 14.324 7 1 0.957 0.966 0.988 0.670 0.723 0.677 3.959 3 0.957 0.964 0.988 0.670 0.723 0.676 10.551 50.9580.9650.9890.6700.7230.67715.793 9 1 0.957 0.963 0.989 0.670 0.723 0.675 2.870 3 0.957 0.961 0.988 0.670 0.723 0.675 7.886 5 0.957 0.964 0.988 0.670 0.723 0.676 12.109 7 0.958 0.964 0.988 0.670 0.723 0.676 15.614 Table E: Ablation on Unified-Frame Editing with anchor on 90-frame videos. (∼0.01), and the substantial gains in temporal consistency metrics justify the trade-off. Comparing the two tables, the anchor consistently improves consistency met- rics (e.g., SC increases from 0.931 to 0.943 at w=1), but incurs an FPS reduction (e.g., 20.9 to 15.3 at w=1). This is expected: adding an anchor increases the num- ber of grid slots from w to w+r (cf. Section H), and the relative overhead is most pronounced at small window sizes where the anchor constitutes a larger fraction of the total grid. F Details of Unified-Frame Editing We prepare the sequence for sliding window processing. To ensure sufficient context for boundary frames, we apply reflect padding to the sequence of inverted OSVE: One Step Video Editing with One Step Diffusion Models25 latents ˆ V T K = (ˆz T 0 ,..., ˆz T K−1 ). Let p = (w− 1)/2 be the padding size (for an odd w). The padded sequence ˆ V T pad is constructed as: ˆ V T pad = (ˆz T p ,..., ˆz T 1 | z left reflection , ˆz T 0 ,..., ˆz T K−1 | z original , ˆz T K−2 ,..., ˆz T K−1−p |z right reflection ) The padded sequence is processed to generate a series of output segments. The total number of segments, N w , is given by N w =⌈K/s⌉. For each segment index i ∈ 0, 1,...,N w − 1, we first extract a window of w latents from the padded sequence. These are spatially concatenated to form the window map W T i : W T i = Concat(ˆz T i·s ,..., ˆz T i·s+w−1 , axis = 2) The final input for the generator, W ′T i , is then constructed by prepending the anchor latent ˆz T A to this window map: W ′T i = Concat(ˆz T A ,W T i , axis = 2) The generator processes this input, producing a unified output map W ′0 i = G θ (W ′T i ,c edit ). From this output, we first discard the initial portion correspond- ing to the anchor frame to get W 0 i : W 0 i = W ′0 i [:, :,W :] (Dimensions:R C×H×(W·w) ) Next, we extract the central s frames from W 0 i to form the final segment S i : S i = W 0 i [:, :,W · offset : W · (offset + s)] (where offset = (w− s)/2) After generating all N w segments, they are concatenated. The resulting map is truncated to the original length of K frames to form the final edited latent map, Z 0 edit : Z 0 edit = Truncate(Concat(S 0 ,S 1 ,...,S N w −1 , axis = 2),K). Finally, this unified map is partitioned back into a sequence of individual frame latents: ˆz 0 k edit = Z 0 edit [:, :,k· W : (k + 1)· W] for k = 0,...,K− 1. The final edited video, V 0 edit , is the sequence of these K latents: ˆ V 0 edit = (ˆz 0 0 edit , ˆz 0 1 edit ,..., ˆz 0 K−1 edit ). G Additional Analysis G.1 Detailed Ablation on Noise Parameterλ noise Detailed quantitative results are reported in Table F. As λ noise increases, both the CLIP [43] distance (CLIP-D) and the structure distance [4] with respect to the original image increase, and their standard deviations also grow. Neverthe- less, for moderate values of λ noise (e.g., λ noise = 0.1), we can generate images that are semantically slightly different while remaining structurally similar to the original, with only a small variability across samples. 26H. Lim and G.-M. Park G.2 Discussion on One-Step T2V Generator Quality As discussed in Section 5 (Discussion) of the main paper, we build OSVE on a one-step T2I backbone because one-step text-to-video (T2V) generators have not yet reached a level of quality sufficient for reliable video editing applications. In this section, we review the current state of accelerated T2V generation and discuss why one-step T2V remains an open and challenging problem. Current State of T2V Distillation. State-of-the-art T2V diffusion models such as CogVideoX [67], HunyuanVideo [20], and Wan2.1 [60] typically require 30–50 denoising steps, resulting in generation times on the order of minutes even on high-end GPUs. While the image domain has seen remarkable progress in distillation—with one-step T2I generators such as DMD2 [68] and SDXL- Turbo [49] achieving quality on par with their multi-step teachers—extending these techniques to video has proven substantially more difficult. Recent efforts have explored trajectory-based approaches [47,54], distribution matching meth- ods [68,70], and their combinations [2,71,72,76], but video distillation remains in a nascent stage where even few-step (2–8 step) methods exhibit measurable quality gaps relative to their full-step teachers in terms of temporal coherence, motion dynamics, text alignment, and generation diversity [10,51,76]. One-Step T2V: Fundamental Quality Limitations When pushed to the extreme of a single denoising step, the quality degradation becomes severe. The few works that have attempted one-step T2V generation consistently report fundamental limitations. APT [27] proposes adversarial post-training for one-step video generation, producing 2-second 720p videos in real time. However, the authors themselves acknowledge significant quality degradation, and the approach has been demon- strated only on short clips with limited resolution. POSE [5] introduces phased adversarial equilibrium for one-step video distillation and identifies three funda- mental limitations of existing approaches: (i) an efficiency bottleneck in scaling to large models (>10B parameters), (i) temporal consistency degradation in- cluding flickering artifacts and broken long-range dependencies, and (i) limited task generalization to conditional downstream tasks. OSV [35] combines con- sistency distillation with GAN training for one-step image-to-video generation, but requires multi-step refinement to match the teacher’s quality. Even rCM [76], one of the strongest distillation frameworks to date, reports a VBench score gap of ∼2 points between its 1-step and 4-step variants on the same Wan2.1 back- bone, and explicitly notes that the distilled model does not surpass the teacher in diversity and physical consistency. We further corroborate this observation by running Wan-rCM 1.3B in the one-step setting; as shown in Figure C, the generated frames suffer from severe blurriness and temporal discontinuities, con- firming that one-step T2V generation remains far from practical quality. These limitations stem from challenges that are unique to or amplified in the video domain: OSVE: One Step Video Editing with One Step Diffusion Models27 Fig. C: Qualitative results of one-step generation using Wan-rCM 1.3B. The outputs exhibit noticeable quality degradation, including spatial blurriness, temporal flickering, and discontinuous motion across frames, illustrating the fundamental limitations of current one-step T2V distillation discussed in this section. 28H. Lim and G.-M. Park – Temporal consistency collapse. One-step generators must produce tem- porally coherent motion in a single forward pass, which often leads to flick- ering, jittering, or static outputs [5,71]. – Spatiotemporal mode averaging. The mode averaging problem—where the model averages over conflicting denoising directions—is dramatically am- plified in the high-dimensional spatiotemporal setting of video, with no op- portunity for iterative correction in one-step generation. – Training instability. POSE [5] observes that adversarial distillation from Gaussian noise faces severe training instability in the video domain due to the high dimensionality of video latents, often resulting in mode collapse or training divergence. Implications for Our Approach. The contrast between T2I and T2V distil- lation is stark: one-step T2I generators reliably match their multi-step teachers, whereas T2V distillation—at any step count—still exhibits notable quality gaps, with one-step T2V suffering from particularly severe degradation. Our OSVE therefore adopts a one-step T2I backbone combined with explicit temporal con- sistency mechanisms, which provides precise per-frame structural control while maintaining temporal coherence—a combination that current T2V distillation methods cannot reliably offer regardless of step count. As T2V distillation tech- nology matures, future work may revisit this design choice. H GPU Memory Analysis A key advantage of the sliding window formulation above is that GPU memory consumption is determined solely by the window size w and is independent of the total sequence length K, since only a single window resides on the GPU at any time. Concretely, let r and c denote the number of rows and columns used to tile the w latent codes in a window onto a 2D spatial grid, so that w = r· c. For each window, our implementation constructs a single tiled latent tensor of shape C × (rH ℓ )× (c ′ W ℓ ), where H ℓ and W ℓ are the spatial dimensions of a single latent code, and feeds it to the denoising UNet and the VAE decoder. When no anchor is used, we have c ′ = c and the number of latent slots in the grid equals the window size: N eff (w) = r· c = w.(6) When an anchor latent ˆz T A is prepended (cf. the construction of W ′T i ), the grid must accommodate w + 1 latent codes in r rows, requiring c ′ = ⌈(w + 1)/r⌉ columns. The effective number of grid slots therefore becomes N eff (w) = r w + 1 r .(7) Since w = r· c by construction, this simplifies to c ′ = c + 1 and thus N eff (w) = r(c + 1) = w + r. OSVE: One Step Video Editing with One Step Diffusion Models29 The peak GPU memory can then be approximated as M VRAM (w) ≈ M 0 + αN eff (w),(8) where M 0 is a constant overhead (model weights, optimizer states, etc.) and α is a per-slot cost that depends on the latent resolution and the UNet architecture. Substituting the anchor case gives M VRAM (w) = M 0 +α (w+r), confirming that the memory requirement grows linearly in the window size, M VRAM = O(w). Crucially, the total number of frames K affects only the number of windows N w = ⌈K/s⌉ and hence the processing time, not the peak memory. This allows our method to scale to arbitrarily long videos without increasing GPU memory, as validated in the ablation study (Tables D and E). I Details of User Study We recruited 26 participants anonymously through social media platforms. No restrictions were placed on professional background, and participants received no information about which method corresponded to which result, ensuring an unbiased evaluation. For each trial, a participant was shown the source video and the editing results produced by six methods side by side. The presentation order of the six methods was randomized independently for every trial and every participant to eliminate positional bias. Each participant evaluated three randomly sampled videos from our datasets, yielding 26×3×6 = 468 individual ratings per criterion. Participants were asked to rate each result on two criteria using a 5-point Likert scale (1 = Severe issues, 5 = No issues): – Temporal Consistency (TC): whether the edited video maintains smooth and coherent motion without flickering or abrupt changes across frames. – Visual Quality (VQ): whether the edited frames are visually pleasing and free from noticeable artifacts. An example of the survey interface presented to participants is shown in Figure F. 30H. Lim and G.-M. Park Fig. D: Graph of MSE loss while training. Fig. E: Graph of SAE loss while training. OSVE: One Step Video Editing with One Step Diffusion Models31 Fig. F: Screenshot of the survey interface presented to participants. 32H. Lim and G.-M. Park λ noise CLIP Distance (CLIP-D) Structure Distance meanstdmean std 0.000 1.000.0000.000 0.000 0.0100 0.994 0.00778 0.00230 0.00324 0.0200 0.987 0.0139 0.00660 0.00779 0.0300 0.981 0.01970.0106 0.0138 0.0400 0.970 0.03210.0163 0.0206 0.0500 0.968 0.03370.0227 0.0337 0.0600 0.956 0.03920.0305 0.0393 0.0700 0.958 0.03530.0301 0.0384 0.0800 0.949 0.04390.0359 0.0358 0.0900 0.944 0.04130.0374 0.0332 0.100 0.944 0.04090.0415 0.0329 0.110 0.936 0.04500.0414 0.0312 0.120 0.931 0.04790.0501 0.0378 0.130 0.925 0.05130.0533 0.0406 0.140 0.921 0.05600.0554 0.0426 0.150 0.915 0.05330.0601 0.0496 0.160 0.911 0.05940.0711 0.0677 0.170 0.909 0.05740.0714 0.0566 0.180 0.903 0.05760.0713 0.0647 0.190 0.898 0.05950.0856 0.0775 0.200 0.897 0.06130.0803 0.0704 0.210 0.886 0.06840.0839 0.0664 0.220 0.891 0.06480.0885 0.0682 0.230 0.882 0.06510.0831 0.0692 0.240 0.866 0.06510.102 0.0762 0.250 0.870 0.07010.0924 0.0783 0.260 0.866 0.07040.102 0.0878 0.270 0.862 0.06070.101 0.0770 0.280 0.856 0.07210.107 0.0867 0.290 0.852 0.06950.106 0.0764 0.300 0.854 0.06900.111 0.0829 0.310 0.840 0.07380.115 0.0843 0.320 0.839 0.07100.112 0.0762 0.330 0.829 0.07070.114 0.0855 0.340 0.827 0.07290.117 0.0918 0.350 0.818 0.08380.124 0.0911 0.360 0.822 0.07860.111 0.0752 0.370 0.817 0.08170.125 0.102 0.380 0.809 0.07090.122 0.0868 0.390 0.808 0.08240.128 0.0979 0.400 0.778 0.09130.129 0.0839 0.410 0.784 0.08180.140 0.114 0.420 0.778 0.08890.141 0.109 0.430 0.772 0.08150.135 0.0894 0.440 0.772 0.07160.141 0.108 0.450 0.763 0.08180.145 0.103 0.460 0.752 0.09020.137 0.102 0.470 0.767 0.08210.145 0.123 0.480 0.744 0.08460.151 0.123 0.490 0.726 0.09210.151 0.117 0.500 0.734 0.07050.147 0.123 Table F: Ablation on the noise parameter λ noise . We report the mean and standard deviation of CLIP distance (CLIP-D, lower is better) and structure distance. OSVE: One Step Video Editing with One Step Diffusion Models33 References 1. Bar-Tal, O., Chefer, H., Tov, O., Herrmann, C., Paiss, R., Zada, S., Ephrat, A., Hur, J., Liu, G., Raj, A., Li, Y., Rubinstein, M., Michaeli, T., Wang, O., Sun, D., Dekel, T., Mosseri, I.: Lumiere: A space-time diffusion model for video genera- tion. arXiv preprint arXiv:2401.12945 (2024). https://doi.org/10.48550/arXiv. 2401.12945 2. Berner, J., et al.: Transition matching distillation for fast video generation. arXiv preprint arXiv:2601.09881 (2026) 3. Brooks, T., Holynski, A., Efros, A.A.: InstructPix2Pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800 (2022) 4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. p. 9650–9660 (2021) 5. Cheng, o.: POSE: Phased one-step adversarial equilibrium for video diffusion mod- els. arXiv preprint arXiv:2508.21019 (2025) 6. Cong, Y., Xu, M., Simon, C., Chen, S., Ren, J., Xie, Y., Perez-Rua, J.M., Rosen- hahn, B., Xiang, T., He, S.: Flatten: optical flow-guided attention for consistent text-to-video editing (2024), https://arxiv.org/abs/2310.05922 7. Dao, T., Nguyen, T.H., Le, T., Vu, D., Nguyen, K., Pham, C., Tran, A.: Swift- brush v2: Make your one-step diffusion model better than its teacher. In: European Conference on Computer Vision. p. 176–192. Springer (2024) 8. Ding, Z., Jin, C., Liu, D., Zheng, H., Singh, K.K., Zhang, Q., Kang, Y., Lin, Z., Liu, Y.: Dollar: Few-step video generation via distillation and latent reward optimiza- tion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2025) 9. Geyer, M., Bar-Tal, O., Bagon, S., Dekel, T.: Tokenflow: Consistent diffusion fea- tures for consistent video editing. arXiv preprint arXiv:2307.10373 (2023) 10. GPD Authors: GPD: Guided progressive distillation for fast and high-quality video generation. arXiv preprint arXiv:2602.01814 (2025) 11. HaCohen, Y., Chiprut, N., Brazowski, B., Shalem, D., Moshe, D., Richardson, E., Levin, E., Shiran, G., Zabari, N., Gordon, O., Panet, P., Weissbuch, S., Kulikov, V., Bitterman, Y., Melumian, Z., Bibi, O.: Ltx-video: Realtime video latent diffusion (2024), https://arxiv.org/abs/2501.00103 12. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-or, D.: Prompt-to-prompt image editing with cross-attention control. In: The Eleventh In- ternational Conference on Learning Representations (2023), https://openreview. net/forum?id=_CDixzkzeyb 13. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS p. 6840–6851 (2020) 14. Huang, Z., He, Y., Yu, J., Zhang, F., Si, C., Jiang, Y., Zhang, Y., Wu, T., Jin, Q., Chanpaisit, N., et al.: Vbench: Comprehensive benchmark suite for video gener- ative models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 21807–21818 (2024) 15. Jiang, Z., et al.: Vace: All-in-one video creation and editing. arXiv preprint arXiv:2503.07598 (2025). https://doi.org/10.48550/arXiv.2503.07598 16. Ju, X., Zeng, A., Bian, Y., Liu, S., Xu, Q.: Pnp inversion: Boosting diffusion-based editing with 3 lines of code. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=FoMZ4ljhVw 34H. Lim and G.-M. Park 17. Kara, O., Kurtkaya, B., Yesiltepe, H., Rehg, J.M., Yanardag, P.: Rave: Random- ized noise shuffling for fast and consistent video editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 6507–6516 (2024) 18. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. p. 5148–5157 (2021) 19. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: International Con- ference on Learning Representations (ICLR) (2014) 20. Kong, W., Tian, Q., Zhang, Z., Min, R., Dai, Z., Zhou, J., Xiong, J., Li, X., Wu, B., Zhang, J., Wu, K., Lin, Q., Yuan, J., Long, Y., Wang, A., Wang, A., Li, C., Huang, D., Yang, F., Tan, H., Wang, H., Song, J., Bai, J., Wu, J., Xue, J., Wang, J., Wang, K., Liu, M., Li, P., Li, S., Wang, W., Yu, W., Deng, X., Li, Y., Chen, Y., Cui, Y., Peng, Y., Yu, Z., He, Z., Xu, Z., Zhou, Z., Xu, Z., Tao, Y., Lu, Q., Liu, S., Zhou, D., Wang, H., Yang, Y., Wang, D., Liu, Y., Jiang, J., Zhong, C.: Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603 (2024). https://doi.org/10.48550/arXiv.2412.03603 21. LAION-AI: aesthetic-predictor. https://github.com/LAION- AI/aesthetic- predictor (2022), accessed: 2025-09-25 22. Li, J., Feng, W., Fu, T.J., Wang, X., Basu, S., Chen, W., Wang, W.Y.: T2v-turbo: Breaking the quality bottleneck of video consistency model with mixed reward feedback. In: Advances in Neural Information Processing Systems (2024) 23. Li, Z., Zhu, Z.L., Han, L.H., Hou, Q., Guo, C.L., Cheng, M.M.: Amt: All-pairs multi-field transforms for efficient frame interpolation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 9801– 9810 (2023) 24. Lim, H., Won, Y., Seo, J., Park, G.M.: Conceptsplit: Decoupled multi-concept per- sonalization of diffusion models via token-wise adaptation and attention disentan- glement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2025) 25. Lin, S., Wang, A., Yang, X.: Sdxl-lightning: Progressive adversarial diffusion dis- tillation. arXiv preprint arXiv:2402.13929 (2024) 26. Lin, S., Xia, X., Ren, Y., Yang, C., Xiao, X., Jiang, L.: Diffusion adversarial post- training for one-step video generation. arXiv preprint arXiv:2501.08316 (2025) 27. Lin, S., Xia, X., Ren, Y., Yang, C., Xiao, X., Jiang, L.: Diffusion adversarial post- training for one-step video generation. arXiv preprint arXiv:2501.08316 (2025) 28. Lin, S., Yang, X.: Animatediff-lightning: Cross-model diffusion distillation. arXiv preprint arXiv:2403.12706 (2024) 29. Liu, S., Zhang, Y., Li, W., Lin, Z., Jia, J.: Video-p2p: Video editing with cross- attention control. In: Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition. p. 8599–8608 (2024) 30. Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. In: The Eleventh International Conference on Learning Representations (ICLR) (2023) 31. Liu, X., Zhang, X., Ma, J., Peng, J., et al.: Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In: The Twelfth Interna- tional Conference on Learning Representations (2023) 32. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. p. 3431–3440 (2015) OSVE: One Step Video Editing with One Step Diffusion Models35 33. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations 34. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency mod- els: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378 (2023) 35. Mao, X., et al.: OSV: One step is enough for high-quality image to video generation. In: CVPR (2025) 36. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inver- sion for editing real images using guided diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. p. 6038–6047 (2023) 37. Moon, S.J., Kim, C., Park, G.M.: WINE: Wavelet-guided GAN inversion and edit- ing for high-fidelity refinement. In: Proceedings of the IEEE/CVF Winter Confer- ence on Applications of Computer Vision (WACV) (2025) 38. Moon, S.J., Park, G.M.: Interestyle: Encoding an interest region for robust stylegan inversion. In: European Conference on Computer Vision (ECCV). p. 460–476. Springer (2022) 39. Nguyen, T.H., Tran, A.: Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 7807–7816 (2024) 40. OpenAI: Video generation models as world simulators. Tech. rep., OpenAI (Feb 2024) 41. Qi, C., Cun, X., Zhang, Y., Lei, C., Wang, X., Shan, Y., Chen, Q.: Fatezero: Fusing attentions for zero-shot text-based video editing. ICCV (2023) 42. Qu, L., Zhang, H., Liu, Y., Wang, X., Jiang, Y., Gao, Y., Ye, H., Du, D.K., Yuan, Z., Wu, X.: Tokenflow: Unified image tokenizer for multimodal understanding and generation (2024), https://arxiv.org/abs/2412.03069 43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. p. 8748–8763. PMLR (2021) 44. Roich, D., Mokady, R., Bermano, A.H., Cohen-Or, D.: Pivotal tuning for latent- based editing of real images. ACM Transactions on graphics (TOG) 42(1), 1–13 (2022) 45. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. p. 10684–10695 (2022) 46. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi- cal image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. p. 234–241. Springer (2015) 47. Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512 (2022) 48. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distilla- tion. In: European Conference on Computer Vision. p. 87–103. Springer (2024) 49. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distilla- tion. In: ECCV (2025) 50. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) 51. Self-Forcing++ Authors: Self-forcing++: Towards minute-scale high-quality video generation. arXiv preprint arXiv:2510.02283 (2025) 36H. Lim and G.-M. Park 52. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsuper- vised learning using nonequilibrium thermodynamics. In: Bach, F., Blei, D. (eds.) Proceedings of the 32nd International Conference on Machine Learning. Proceed- ings of Machine Learning Research, vol. 37, p. 2256–2265. PMLR, Lille, France (07–09 Jul 2015), https://proceedings.mlr.press/v37/sohl-dickstein15. html 53. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) 54. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models (2023) 55. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020) 56. Sun, K., Pan, J., Ge, Y., Li, H., Duan, H., Wu, X., Zhang, R., Zhou, A., Qin, Z., Wang, Y., et al.: Journeydb: A benchmark for generative image understanding. Advances in neural information processing systems 36, 49659–49678 (2023) 57. Tang, L., Jia, M., Wang, Q., Phoo, C.P., Hariharan, B.: Emergent correspon- dence from image diffusion. Advances in Neural Information Processing Systems 36, 1363–1389 (2023) 58. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG) 40(4), 1–14 (2021) 59. Tumanyan, N., Geyer, M., Bagon, S., Dekel, T.: Plug-and-play diffusion features for text-driven image-to-image translation. In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition. p. 1921–1930 (2023) 60. Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.W., Chen, D., Yu, F., Zhao, H., Yang, J., Zeng, J., Wang, J., Zhang, J., Zhou, J., Wang, J., Chen, J., Zhu, K., Zhao, K., Yan, K., Huang, L., Feng, M., Zhang, N., Li, P., Wu, P., Chu, R., Feng, R., Zhang, S., Sun, S., Fang, T., Wang, T., Gui, T., Weng, T., Shen, T., Lin, W., Wang, W., Wang, W., Zhou, W., Wang, W., Shen, W., Yu, W., Shi, X., Huang, X., Xu, X., Kou, Y., Lv, Y., Li, Y., Liu, Y., Wang, Y., Zhang, Y., Huang, Y., Li, Y., Wu, Y., Liu, Y., Pan, Y., Zheng, Y., Hong, Y., Shi, Y., Feng, Y., Jiang, Z., Han, Z., Wu, Z.F., Liu, Z.: Wan: Open and advanced large-scale video generative models (2025), https://arxiv.org/abs/2503.20314 61. Wang, J., Ma, Y., Guo, J., Xiao, Y., Huang, G., Li, X.: Cove: Unleashing the diffusion feature correspondence for consistent video editing. Advances in Neural Information Processing Systems 37, 96541–96565 (2024) 62. Wang, Y., et al.: Videodirector: Precise video editing via text-to-video models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025) 63. Wu, J.Z., Ge, Y., Wang, X., Lei, W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to- video generation (2023), https://arxiv.org/abs/2212.11565 64. Xing, Z., Feng, Q., Chen, H., Dai, Q., Hu, H., Xu, H., Wu, Z., Jiang, Y.G.: A survey on video diffusion models. ACM Computing Surveys 57(2), 1–42 (2024) 65. Yang, S., Zhou, Y., Liu, Z., Loy, C.C.: Fresco: Spatial-temporal correspondence for zero-shot video translation (2024), https://arxiv.org/abs/2403.12962 66. Yang, Y., Li, X., Zhang, H., Li, Z., Shechtman, E., Li, Y.: Towards one-step causal video generation via adversarial self-distillation. arXiv preprint arXiv:2511.01419 (2025) OSVE: One Step Video Editing with One Step Diffusion Models37 67. Yang, Z., Teng, J., Zheng, W., Ding, M., Huang, S., Xu, J., Yang, Y., Hong, W., Zhang, X., Feng, G., Yin, D., Zhang, Y., Wang, W., Cheng, Y., Xu, B., Gu, X., Dong, Y., Tang, J.: Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072 (2024) 68. Yin, T., Gharbi, M., Park, T., Zhang, R., Shechtman, E., Durand, F., Freeman, B.: Improved distribution matching distillation for fast image synthesis. Advances in neural information processing systems 37, 47455–47487 (2024) 69. Yin, T., Gharbi, M., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T., Park, T.: One-step diffusion with distribution matching distillation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. p. 6613– 6623 (2024) 70. Yin, T., Gharbi, M., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T., Park, T.: One-step diffusion with distribution matching distillation. In: CVPR (2024) 71. Yin, T., Zhang, Q., Zhang, R., Freeman, W.T., Durand, F., Shechtman, E., Huang, X.: From slow bidirectional to fast autoregressive video diffusion models. In: CVPR (2025) 72. Zhang, J., Zheng, K., Jiang, K., Wang, H., Stoica, I., Gonzalez, J.E., Chen, J., Zhu, J.: Turbodiffusion: Accelerating video diffusion models by 100-200 times. arXiv preprint arXiv:2512.16093 (2025) 73. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543 (2023) 74. Zhang, Z., Li, Y., Wu, Y., Xu, Y., Kag, A., Skorokhodov, I., Menapace, W., Siaro- hin, A., Cao, J., Metaxas, D., Tulyakov, S., Ren, J.: Sf-v: Single forward video generation model. In: Advances in Neural Information Processing Systems (2024) 75. Zheng, K., Wang, Y., Ma, Q., Chen, H., Zhang, J., Balaji, Y., Chen, J., Liu, M.Y., Zhu, J., Zhang, Q.: Large scale diffusion distillation via score-regularized continuous-time consistency. arXiv preprint arXiv:2510.08431 (2025) 76. Zheng, K., Wang, Y., Ma, Q., Chen, H., Zhang, J., Balaji, Y., Chen, J., Liu, M.Y., Zhu, J., Zhang, Q.: Large scale diffusion distillation via score-regularized continuous-time consistency. arXiv preprint arXiv:2510.08431 (2025)

Instant research discovery