Paper deep dive

VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System

Haodong Li, Tianfei Ren, Xiaoxiao Ma, Chunmei Qing, Zhen Fang, Sipeng He, Ziyu Guo, Haoyu Wu, Juanxi Tian, Yihang Zou, Ruichuan An, Dongzhi Jiang, Boxue Yang, Ji Xie, Xu Huang, Wenhao Yan, Jialv Zou, Zhengrong Yue, Yaxin Luo, Xiaotong Li, Yuzhu Wang, Junyan Ye, Jinjing Zhao, Zehui Chen, Lin Chen, Renye Yan, Feng Zhao, Pheng-Ann Heng

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 47

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%

Last extracted: 8/1/2026, 1:27:33 AM

Summary

The paper introduces VideoCoCo, an agentic dual-engine framework for physically consistent video generation. It uses executable Blender code as a process-level chain of thought to generate deterministic spatiotemporal drafts, which are then refined into photorealistic videos by a generative video engine. The authors also construct VideoCoCo-3K, a dataset of draft-instruction-target triplets, to train the video editor. The method achieves state-of-the-art performance on PhyGenBench and VBench-2.0.

Entities (8)

VideoCoCo · framework · 98%VideoCoCo-3K · dataset · 97%PhyGenBench · benchmark · 96%VBench-2.0 · benchmark · 96%Blender · software · 95%OmniWeaving · baseline-model · 94%Code-as-CoT · concept · 93%Seedance 2.0 · model · 92%

Relation Signals (7)

VideoCoCo → utilizes → Code-as-CoT

confidence 96% · VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation

VideoCoCo → evaluatedon → VBench-2.0

confidence 95% · and from 52.18 to 77.88 on VBench-2.0

VideoCoCo → evaluatedon → PhyGenBench

confidence 95% · VideoCoCo improves the OmniWeaving baseline ... on PhyGenBench

VideoCoCo → uses → Blender

confidence 95% · executable Blender code serves as a process-level chain of thought

VideoCoCo → produces → VideoCoCo-3K

confidence 94% · we construct VideoCoCo-3K, a curated dataset of draft–instruction–target triplets

VideoCoCo → outperforms → OmniWeaving

confidence 92% · VideoCoCo improves the OmniWeaving baseline from 0.475 to 0.558 on PhyGenBench

Seedance 2.0 → usedasteacherfor → VideoCoCo-3K

confidence 90% · We instantiate G T as Seedance 2.0 ... The resulting dataset is D VideoCoCo-3K

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Text-to-video models have achieved remarkable visual quality, yet they still struggle to generate physically consistent dynamics because the temporal evolution of a scene must be inferred implicitly from a highly compressed text prompt. Existing chain-of-thought approaches introduce intermediate plans or visual states, but these representations are typically non-executable or temporally sparse, limiting their ability to instantiate and control the complete spatiotemporal process. To address this limitation, we introduce VideoCoCo, an agentic dual-engine framework in which executable Blender code serves as a process-level chain of thought. Given a text prompt, a coding agent synthesizes a Blender program that explicitly specifies the scene and its temporal evolution. The executable simulation engine runs the program to produce a deterministic spatiotemporal draft, which is subsequently transformed into a photorealistic video by a generative video engine through draft-conditioned editing. This decomposition separates process-level reasoning from high-fidelity visual realization. To adapt the video editor to simulated drafts, we construct VideoCoCo-3K, a curated dataset of draft-instruction-target triplets. VideoCoCo improves the OmniWeaving baseline from 0.475 to 0.558 on PhyGenBench and from 52.18 to 77.88 on VBench-2.0, achieving the best average score on both benchmarks. These results demonstrate that executable code provides an effective, controllable, and inspectable intermediate representation for physically consistent video generation.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

46,177 characters extracted from source content.

Expand or collapse full text

VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System Haodong Li 1,∗,‡ ,Tianfei Ren 2,∗ ,Xiaoxiao Ma 2,∗ ,Chunmei Qing 3,† ,Zhen Fang 2,‡ , Sipeng He 2 ,Ziyu Guo 1 ,Haoyu Wu 4 ,Juanxi Tian 5 ,Yihang Zou 3 , Ruichuan An 6 ,Dongzhi Jiang 1 ,Boxue Yang 7 ,Ji Xie 8 ,Xu Huang 6 , Wenhao Yan 9 ,Jialv Zou 10 ,Zhengrong Yue 7 ,Yaxin Luo 11 ,Xiaotong Li 6 , Yuzhu Wang 9 ,Junyan Ye 1 ,Jinjing Zhao 1 ,Zehui Chen 2 ,Lin Chen 2 , Renye Yan 6 ,Feng Zhao 2,† ,Pheng-Ann Heng 1,† 1 CUHK 2 USTC 3 SCUT 4 HKU 5 NTU 6 PKU 7 SJTU 8 CMU 9 THU 10 HUST 11 MBZUAI ∗ Equal contribution., † Corresponding authors., ‡ Project leaders. Abstract Text-to-video models have achieved remarkable visual quality, yet they still struggle to gener- ate physically consistent dynamics because the temporal evolution of a scene must be inferred implicitly from a highly compressed text prompt. Existing chain-of-thought approaches intro- duce intermediate plans or visual states, but these representations are typically non-executable or temporally sparse, limiting their ability to instantiate and control the complete spatiotemporal process. To address this limitation, we introduceVideoCoCo, an agentic dual-engine framework in which executable Blender code serves as a process-level chain of thought. Given a text prompt, a coding agent synthesizes a Blender program that explicitly specifies the scene and its tempo- ral evolution. The executable simulation engine runs the program to produce a deterministic spatiotemporal draft, which is subsequently transformed into a photorealistic video by a genera- tive video engine through draft-conditioned editing. This decomposition separates process-level reasoning from high-fidelity visual realization. To adapt the video editor to simulated drafts, we constructVideoCoCo-3K, a curated dataset ofdraft–instruction–targettriplets. VideoCoCo improves the OmniWeaving baseline from 0.475 to 0.558 on PhyGenBench and from 52.18 to 77.88 on VBench-2.0, achieving the best average score on both benchmarks. These results demonstrate that executable code provides an effective, controllable, and inspectable intermediate representa- tion for physically consistent video generation. Date:August 2026 Correspondence: mickyhimself4@gmail.com Project Page:https://github.com/micky-li-hd/VideoCoCo arXiv:2607.27380v1 [cs.CV] 29 Jul 2026 Planner Prompt Video Generator Video Keyframes Plan Layout ... ① Planning CoT ② Test-Time Search CoT Video Generator Candidate Videos Video Revision Prompt ③ Visual-State CoT Prompt Video Generator Video State 1 ... State N Prompt Code Agent Code Generation Sandbox Draft Video Generator Video VideoCoCo (Ours) Figure 1Comparison of chain-of-thought paradigms for video generation.(1) Planning CoTexternalizes reasoning as textual plans, keyframes, or layouts that condition a video generator.(2) Test-Time Search CoTsamples multiple candidate videos and revises or selects among them at inference time.(3) Visual- State CoTreasons through a sequence of intermediate visual states within the generated video. In contrast, VideoCoCo (Ours)uses a code agent to synthesize executable code, renders a deterministic sandbox draft, and conditions the video generator on this draft, yielding a complete and inspectable process-level chain of thought rather than sparse or non-executable intermediates. 1Introduction The pursuit of Artificial General Intelligence (AGI) requires machines to understand and model how the world evolves over time. Video generation provides a concrete means of representing such dynamics by transforming natural-language descriptions into temporally evolving visual sequences. Recent advances have enabled increasingly realistic and temporally coherent videos, positioning video generation as a promising pathway toward generative world modeling. Realizing this potential, however, requires more than producing visually plausible videos. In the prevailing text-to-video paradigm, a model maps a text prompt to a complete visual sequence. Such prompts specify an event only at a highly compressed semantic level, leaving the physical principles governing its evolution largely unstated. The model must therefore recover a complete spatiotemporal process while simultaneously synthesizing its visual appearance. We refer to this mismatch between compressed intent and fully realized evolution asCausal Opacity. This raises a central question:How can the process implicit in a prompt be externalized before final video generation? 2 A natural way to address this question is to reason through an intermediate representation rather than directly in pixels. Recent work has begun to extend chain-of-thought reasoning to video generation: external-planning methods use text, layouts, or keyframes to guide synthesis [1,2]; frame-chain methods reason through generated visual states [3,4]; and test-time methods search over or revise candidate trajectories [5,6]. Despite their different forms, these intermediates remain largely descriptive or selective, rather than instantiating a complete process that can be directly inspected and revised (Figure1). Recent advances in code agents point to a different possibility: executable code itself can serve as a process-level intermediate representation for video generation. Building on this possibility, we introduceVideoCoCo, an agentic dual-engine framework in which executable code serves as the chain of thought. Given a text prompt, a coding agent synthesizes a Blender program that explicitly specifies the scene and its temporal evolution. The executable simulation engine runs the program in a sandboxed Blender environment to produce a deterministic spatiotemporal draft. Conditioned on this draft, the generative video engine realizes the simulated process as a photorealistic video through draft-conditioned editing. This division separates process- level reasoning from visual realization: the coding agent determines how the scene evolves, while the video editor realizes the instantiated process with high visual fidelity. Realizing this dual-engine design requires supervision that connects simulated drafts with photoreal- istic targets while preserving their underlying spatiotemporal processes. To this end, we develop an agentic data-construction pipeline and constructVideoCoCo-3K, a dataset of draft–instruction– target triplets. These triplets adapt the generative video engine to realize simulated processes as photorealistic videos while retaining the spatiotemporal structure encoded in the drafts. Experi- ments on PhyGenBench and VBench-2.0 demonstrate that VideoCoCo substantially improves phys- ical consistency over strong video generation baselines, validating executable code as an effective process-level intermediate representation. In summary, our contributions are as follows: •We introduceVideoCoCo, an agentic dual-engine framework for physically consistent video generation. VideoCoCo uses executable Blender code as a process-level chain of thought, producing a spatiotemporal draft through simulation and realizing it as a photorealistic video through draft-conditioned editing. •We constructVideoCoCo-3K, a dataset ofdraft–instruction–targettriplets generated through an agentic data construction pipeline. The dataset provides aligned supervision for adapting the generative video engine to simulated drafts. •VideoCoCo achieves state-of-the-art average performance on both PhyGenBench and VBench- 2.0, substantially improving the OmniWeaving baseline, with particularly large gains in ther- mal and material dynamics. 2Related Work 2.1Text-to-Video Diffusion Models Diffusion models have become the dominant paradigm for text-to-video (T2V) generation due to their strong visual quality and favorable scaling behavior. Early methods established the founda- tions of temporally coherent synthesis [7–9]. Building on these foundations, latent-diffusion ap- proaches improved fidelity and eﬀiciency through lightweight adaptation and tuning [ 10–13]. More 3 recently, large-scale video diffusion backbones have further strengthened temporal modeling, pow- ering systems such as Open-Sora, HunyuanVideo, Wan, CogVideoX, and Step-Video-T2V [14–18]. Beyond video synthesis itself, these models have also been explored as general visual and world rep- resentations [19–21]. Complementary work strengthens generative world representations through reconstruction alignment, explicit coupling to 3D geometry, and scalable multi-agent, multi-view modeling [22–24]. Nevertheless, generating physically plausible and temporally coherent dynamics remains an open challenge [25–28]. 2.2Physics-Aware Video Generation Existing approaches to physically plausible video generation can be broadly divided into explicit and implicit paradigms. Explicit methods directly incorporate physical knowledge through simulators, trajectory guidance, or structured constraints [29–33]. Although effective for predefined phenomena, they often rely on fixed rules or templates that generalize poorly to open-world prompts. To reduce this dependence on hand-specified structure, implicit methods instead learn physical priors from data through preference optimization, reinforcement learning, or feature alignment [34–38]. Complementary work makes multimodal generative criteria explicit by deriving auto-rubrics as reward signals [39]. This shift improves applicability, but offers limited control over prompt-specific processes because physical knowledge remains encoded indirectly through rewards, preferences, or learned representations. Despite their complementary strengths, neither paradigm typically grounds generation in an explicit spatiotemporal realization of the target process. VideoCoCo addresses this limitation by having an agent write and execute simulation code to produce a deterministic draft video, which subsequently guides a video-editing model toward a realistic result. 2.3Chain-of-Thought for Visual Generation Chain-of-thought reasoning [40] has recently been extended from language to visual generation. Ex- isting video approaches generally follow three directions. One line of work externalizes reasoning through textual plans, layouts, or keyframes, as in VChain and VideoDirectorGPT [ 1,2]. A sec- ond line embeds reasoning within the generated sequence itself, using intermediate visual states as in VideoRLVR and ChEaP [3,4]. Rather than prescribing intermediate states, a third line allo- cates additional computation at inference time to search over or revise generated trajectories, as in Video-T1 and temporal backtracking [ 5,6]. Beyond these video-generation paradigms, DraCo uses a draft as chain of thought for text-to-image preview and rare-concept generation [41], while CoCo further represents the intermediate as executable code [42]. MetaPoint and VideoCoF introduce structured spatial control and temporal reasoning for agentic visual generation and unified video editing [ 43,44]. Although prior video CoT methods improve planning and structural consistency, their intermediate representations remain sparse, non-executable, or specialized to constrained envi- ronments. VideoCoCo instead extends executable Code-as-CoT from image preview to a complete temporal process: it renders a deterministic draft video that provides the downstream editor with dense spatiotemporal guidance over the target dynamics. 3Method We instantiate the Code-as-CoT idea asVideoCoCo, an agentic dual-engine framework that sepa- ratesprocess-levelreasoning fromvisualrealization. As illustrated in Figure2, the two engines play complementary roles. Anexecutablesimulationengine externalizes the process implicit in a text prompt as a runnable Blender program and renders it into a deterministic spatiotemporal draft, 4 Executable Simulation Engine Plan & Write Code A cinematic slow push-in watches as a knob of butter in a sizzling skillet foams, browns and pools, leaving melted butter. User Prompt Spatiotemporal Draft (Low-fidelity) Generative Video Engine 퐀 퐀 퐀 퐀 퐀 퐀 퐀 퐀 Draft video (from stage 1) A cinematic slow push-in watches as a knob of butter in a sizzling skillet foams, browns and pools, leaving melted butter. Prompt (from stage 1) Final Video (High-fidelity) A pale golden knob of butter at the center of a dark cast-iron skillet, product-style close-up. Under heat it softens, slumps downward, and spreads into a thin sizzling pool with foamy edges, ripples, and nutty amber browning. Warm directional light with a bright rim creates wet gleaming highlights. Fixed slightly elevated three-quarter view, subtle push-in, ~85m, shallow depth of field. Warm butter-gold and toasted amber tones, rich savory mood, photorealistic, cinematic, crisp detail. --resolution 720p --duration 5 --ratio 16:9 Compose Editing Instruction E d i t M o d e l Instruction Draft Invoke Editing Model Figure 2Overview of the VideoCoCo dual-engine framework.Stage 1 (Executable Simulation Engine): given a user prompt, a coding agent plans and writes a self-contained Blender program that specifies the scene and its temporal evolution; executing this program in a sandbox renders a low-fidelity spatiotemporal draft. Stage 2 (Generative Video Engine):conditioned on the stage-1 prompt and draft video, an instruction agent composes an editing instruction, which together with the draft is passed to a video editing model to produce the final high-fidelity, physically consistent video. The executable draft supplies the process-level dynamics, while the editing model supplies photorealistic appearance. thereby committing the system to a concrete realization of the target dynamics before any pixels are synthesized. Agenerativevideoengine then takes this draft as a structural condition and turns it into a photorealistic video, focusing its capacity on appearance rather than on reconstructing the process from scratch. This division is deliberate: it decouples the two aspects of physical- video generation that are hardest to satisfy simultaneously—faithful dynamics and photorealistic appearance—and lets each engine specialize in what it does best. The remainder of this section follows this division. We first describe how the process is externalized as executable code, then how the resulting draft is realized as a photorealistic video, how we equip the video engine with the ability to read simulated drafts through a dedicated dataset, and finally how the editor is adapted and how the whole pipeline runs end-to-end at inference time. 3.1Executable Simulation Engine A text prompt compresses a physical event into a few words, forcing a text-to-video model to reconstruct the entire spatiotemporal process from language alone. This burden is precisely where existing generators fail: appearance can be memorized from data, but a prompt-specific process rarely can. Instead of asking the video model to bear this burden implicitly, we let an agent write the process down explicitly. Our first engine therefore turns the prompt into an executable program and runs it in a sandbox, so that the target dynamics are instantiated before any pixels 5 Dry ice sublimating directly as temperature rises. A sealed plastic bottle collapsing as air is extracted. PromptDraft VideoBaseline (OmniWeaving)Ours (VideoCoCo) An egg is thrown forcefully at a rough rock surface and shatters on impact. A wooden toy is placed gently on the surface of a bowl of water. Figure 3 Qualitative comparison on representative physical processes.OmniWeaving often pro- duces visually plausible videos that violate the requested dynamics, whereas VideoCoCo follows the executable spatiotemporal draft and better preserves the intended processes, including sublimation, vacuum-induced col- lapse, impact shattering, and buoyancy. are synthesized. Code-as-CoT Program Synthesis.Given an input promptp, a coding agentA code synthesizes a self-contained Blender Python program c=A code (p),(1) that specifies the scene, the objects and their physical properties, and the temporal evolution of the target event. We deliberately choose Blender Python as the representation of reasoning, rather than natural-language plans or sparse keyframes, for three reasons. First, it is explicit: every object, motion, and interaction must be declared, leaving no room for the underspecification typical of textual plans. Second, it isexecutable: the program commits the agent to a concrete process that can actually be run, not merely described. Third, it is inspectable: the code can be read, edited, and re-executed, so the intermediate reasoning is transparent rather than hidden inside latent activations. Together, these properties mark the key departure from prior chain-of-thought representations for video, which remain either non-executable (text, layouts) or temporally sparse (isolated keyframes). Sandboxed Execution and Draft Rendering.Code alone does not yet realize the process. We therefore execute the program in an isolated Blender environmentBto obtain a rendered draft d=B(c).(2) Executing the program in a controlled runtime is essential: the sandbox provides a standardized set of primitives and enforces deterministic behavior, so that the same program always yields the 6 same draft, and rendering errors are returned as diagnostic signals for code revision rather than remaining silent failures. The resulting draft is intentionally low-fidelity—a white-clay simulation without photorealistic materials or complex lighting—but it istemporallydense: every frame corre- sponds to a physically instantiated state of the programmed process, from initial condition through intermediate stages to final outcome. In this sensedis not a candidate output but a structural scaffold: it fixes what happens and when, leaving the question of how it should look to the next engine. 3.2Generative Video Engine The simulation draft solves the process, but not the appearance. Bridging this gap is the job of the second engine, whose role is deliberately narrow: rather than re-reasoning about the dynamics, it only translates the already instantiated process into a photorealistic video. This narrow scope is what allows the framework to be both physically consistent and visually faithful—the editor never has to guess what should happen, only how it should look. Editing Instruction Construction.The original promptpand the rendered draftddescribe the target at very different levels of granularity. The prompt is compressed, abstract, and biased toward what the event is, while the draft is dense, stylized, and biased toward how the eventunfolds. Passing pdirectly to a video editor would therefore leave the appearance underspecified and risk generating motion that contradicts the draft. To reconcile the two, we introduce an instruction agentA edit that reads both signals and composes an appearance-focused editing instruction: e=A edit (p, d).(3) The instructioneis written to describe the target subjects, materials, lighting, and cinematic style, and is explicitly discouraged from redefining the motion already present ind. It therefore complements, rather than competes with, the draft:ddictateswhathappens, whileedictateswhat itlookslike. Draft-Conditioned Video Editing.The two conditions are then fed jointly into a draft-conditioned video editorG θ , which produces the final video ˆv=G θ (d, e).(4) The two conditions carry disjoint responsibilities:danchors the spatiotemporal structure of the target process, ensuring that the physical evolution is preserved, whileespecifies the photorealistic appearance the final video should exhibit. Under this decomposition,G θ is asked torestyle an already instantiated process rather than to imagine one from scratch—a substantially easier problem than direct text-to-video generation, and one that closely aligns with the strengths of modern video-editing models (see Figure 3for a qualitative illustration). However, the draft follows a distinctive white-clay simulation style that lies outside the training distribution of off-the-shelf editors, which either ignore its motion or hallucinate appearance that conflicts with it. TurningG θ into a competent draft-conditioned editor therefore requires targeted supervision, which motivates the dataset introduced next. 3.3VideoCoCo-3K: Data for Draft-Conditioned Editing The core diﬀiculty in adaptingG θ is that publicly available video-editing datasets contain only natural-video pairs. There is essentially no supervision that pairs a simulated white-clay draft 7 with a photorealistic video sharing thesame spatiotemporal process, which is precisely what our second engine needs to learn from. We therefore constructVideoCoCo-3K, a dataset ofdraft– instruction–targettriplets that provides exactly this alignment. Teacher-Based Triplet Construction.We build the dataset by running the pipeline of Sec. 3.1–3.2 at scale. For each promptp i collected from our source distribution, the executable simulation engine produces a rendered draftd i and the instruction agent composes an editing instructione i . To obtain the corresponding photorealistic target, we invoke a high-fidelity teacher editorG T once per triplet: y i =G T (d i , e i ).(5) We instantiateG T as Seedance 2.0, chosen because it is strong enough to produce photorealistic frames while faithfully preserving the motion carried byd i —the two properties that a training target for draft-conditioned editing must simultaneously satisfy. The resulting dataset is D VideoCoCo-3K =(d i , e i , y i ) 3000 i=1 ,(6) in which each triplet aligns a physically grounded process with its photorealistic realization. We exclude all prompts from the evaluation benchmarks, along with their near-duplicates. We addi- tionally retain each original prompt and Blender program as metadata, allowing the dataset to be inspected, extended, or regenerated as the pipeline evolves. 3.4Editor Adaptation and Inference With the aligned triplets in place, the remaining question is how to teachG θ to consume simulated drafts and how the whole pipeline runs at inference time from a text prompt alone. Training Objective.We adapt the draft-conditioned editorG θ on the VideoCoCo-3K triplets (d i , e i , y i ), where the draft and instruction serve as conditions and the photorealistic target pro- vides supervision. Letz 0 denote the latent of the targetyandz t its noised version at diffusion timestept. The editor is trained with a standard conditional denoising objective L(θ) =E (d,e,y), t, ε [ ∥ε−ε θ (z t , t, d, e)∥ 2 2 ] ,(7) withε∼N(0, I). Minimizing this objective teaches the editor to synthesize a photorealistic video whose noise structure is consistent with the target while remaining conditioned on both the simulated draft and the editing instruction. We initializeG θ from the OmniWeaving base generator and consider two adaptation strategies—full fine-tuning and parameter-eﬀicient LoRA—whose empirical comparison is deferred to Sec. 4.3. End-to-End Inference.At inference time, VideoCoCo operates end-to-end from a text prompt alone. Given a promptp, the coding agent synthesizes a Blender programc, the sandbox renders a deterministic draftd, the instruction agent composes an editing instructione, and the adapted editor produces the final videoˆv. No ground-truth draft, human annotation, or auxiliary input is required: the user supplies only the prompt, and the executable draft is generated automatically as the process-level chain of thought. Because every intermediate artifact—the program, the draft, and the instruction—is either code or a rendered video, the entire pipeline is fullyautomatic,inspectable, andreproducible, which we regard as a defining property of Code-as-CoT for video generation. 8 MethodMechanics (↑) Optics (↑) Thermal (↑) Material (↑) Average (↑) Closed Pika [45]0.350.560.430.390.44 Gen-3 [46]0.450.570.490.510.51 Kling [47]0.450.580.500.400.49 Open CogVideoX [17]0.390.550.400.420.45 Open-Sora V1.2 [14]0.430.500.440.370.44 LaVie [12]0.300.440.380.320.36 Vchitect-2.0 [48]0.410.560.440.370.45 HunyuanVideo [15]0.330.390.260.300.33 Wan2.2-TI2V-5B [49]0.550.580.530.500.54 Cosmos-Predict2.5 [50]0.300.330.410.390.35 LTX-Video-2B [51]0.510.580.480.450.51 Ours OmniWeaving [52]0.480.560.430.390.48 + VideoCoCo0.560.610.510.530.56 Table 1Physical-consistency comparison on PhyGenBench across closed- and open-source video generators. We report per-category consistency scores in[0,1]for mechanics, optics, thermal, and material dynamics, together with their average; higher is better. Adding VideoCoCo to the OmniWeaving base generator lifts every category and raises the average from 0.48 to 0.56, the best overall result, with the largest gains on material and thermal dynamics, where appearance-driven priors are weakest. OrangeandChampagne cells denote the best (top-1) and second-best (top-2) results per column. MethodMechanics (↑) Thermotics (↑) Material (↑) Average (↑) Closed Sora [25]62.22%43.36%64.94%56.84% Kling 1.6 [ 53]65.55%59.46%68.00%64.34% Open HunyuanVideo [15]76.09%56.52%64.37%65.66% CogVideoX-1.5 [17]80.80%67.13%83.19%77.04% Ours OmniWeaving [52]62.79%52.08%41.67%52.18% + VideoCoCo92.31%72.92%68.42%77.88% Table 2Physical-plausibility comparison on VBench-2.0, reported as per-dimension plausibility percentages over mechanics, thermotics, and material, with their average; higher is better. Coupling VideoCoCo with OmniWeaving improves the average from 52.18% to 77.88% (a gain of 25.70 points) and attains the best mechanics and thermotics scores among all systems, trailing only CogVideoX-1.5 on material.Orangeand Champagnecells denote the best (top-1) and second-best (top-2) results per column. 4Experiments This section shows that Code-as-CoT drafting improves the physical consistency of generated video, achieving the best average score on both PhyGenBench and VBench-2.0, and that both the exe- cutable draft and the editing stage contribute to the gain. 9 MethodMech.(↑) Opt.(↑) Therm.(↑) Mat.(↑) Avg.(↑) OmniWeaving [52]0.48 0.560.430.39 0.48 + VideoCoCo (Tune-Free) 0.50 0.530.480.510.51 + VideoCoCo (Full-Tune) 0.51 0.610.510.49 0.54 + VideoCoCo (LoRA-Tune)0.560.610.510.53 0.56 Table 3Ablation of video-editor adaptation strategies on PhyGenBench. All variants share the same ex- ecutable draft-generation pipeline and differ only in how the editor is adapted: Tuning-Free, Full-Tune, or LoRA-Tune. The Tuning-Free variant already surpasses OmniWeaving (0.48→0.51), isolating the contri- bution of executable drafting, while LoRA-Tune yields the best average (0.56), showing that drafting and editor adaptation are complementary.Boldand underlinedenote the best (top-1) and second-best (top-2) results per column. 4.1Experimental Setup Datasets.We evaluate on two physics-focused video benchmarks.PhyGenBench[54] evaluates physical commonsense across four categories: mechanics, optics, thermal, and material dynamics. We follow its oﬀicial evaluation protocol, which uses GPT-4o [55] as the MLLM judge to assess how faithfully a generated video adheres to the physical principles implied by its prompt, providing a comprehensive measure of the physical knowledge instantiated by a generative model.VBench- 2.0[ 56] targets intrinsic faithfulness beyond surface quality; we adopt its per-dimension protocol on a pre-registered subset that balances temporal and perceptual/semantic factors. We focus on its physical dimensions—Mechanics (↑; e.g., gravity, buoyancy, stress), Thermotics (↑; e.g., vaporiza- tion, freezing), and Material (↑; e.g., color mixing, solubility)—each probed by a dedicated prompt suite and scored to reflect adherence to real-world physical laws. Baselines.We group baselines by access. Closed-source, API-only systems include Pika [ 45], Gen-3 [46], Kling [47], and Sora [25]. Open-source systems with publicly released weights include CogVideoX [17], Open-Sora [14], LaVie [12], Vchitect-2.0 [48], HunyuanVideo [15], Wan2.2-TI2V- 5B [ 49], Cosmos-Predict2.5 [50], and LTX-Video [51]. VideoCoCo augments the OmniWeaving [52] base generator, which we also report on its own to isolate our contribution. Metrics.PhyGenBench reports a per-category consistency score in[0,1]and their average. VBench- 2.0 reports per-dimension plausibility as a percentage. Higher is better on every metric. Implementation.VideoCoCo writes Blender programs, renders deterministic draft videos in a sandbox, and refines them with a video editor trained on VideoCoCo-3K. We evaluate three editor regimes: tuning-free, LoRA, and full tuning. 4.2Code-as-CoT Improves Physical Consistency PhyGenBench.Table1compares VideoCoCo against a broad set of closed- and open-source gen- erators, including recent strong open models such as Wan2.2-TI2V-5B [49], Cosmos-Predict2.5 [50], and LTX-Video [51]. Adding VideoCoCo to OmniWeaving raises the average consistency score from 0.475 to 0.558, the best overall result in the table, ahead of the strongest open baseline Wan2.2-TI2V- 5B at 0.544. VideoCoCo attains the best per-category score on mechanics (0.558), optics (0.613), and material (0.525), and ranks a close second on thermal (0.511), where Wan2.2-TI2V-5B leads 10 narrowly at 0.533. The largest gains over the OmniWeaving base appear on material (+0.133) and thermal (+0.078), precisely the categories where appearance-driven generators, lacking an explicit dynamics reference, are weakest. Takeaway.Executable drafting helps most where visual priors are weakest, which confirms that code contributes physical dynamics rather than surface realism. VBench-2.0.Table2reports physical plausibility on VBench-2.0. OmniWeaving alone scores 52.18% on average, and adding VideoCoCo raises the average to 77.88%, a gain of 25.70 points over the base model and the best average among all systems. VideoCoCo attains the best mechanics (92.31%) and thermotics (72.92%) scores, improving the base generator by 29.52 and 20.84 points respectively, and ranks second on material (68.42%), trailing only CogVideoX-1.5 (83.19%). Takeaway.The VBench-2.0 pattern matches PhyGenBench: simulation-grounded drafts lift the dimensions that data-driven priors handle worst. 4.3Both Drafting and Tuning Contribute Table3ablates the editor adaptation regime on PhyGenBench, holding the executable draft- generation pipeline fixed so that only the way the video editor is adapted varies. Thetuning-free variant, which conditions the unmodified OmniWeaving editor directly on the simulation draft, al- ready improves the average from 0.475 to 0.506 and lifts the thermal and material categories most (0.433→0.478 and 0.392→0.508). Since no editor parameters are updated, this gain is attributable entirely to the executable draft, confirming that the draft—not additional training—supplies the physical dynamics. Adapting the editor on VideoCoCo-3K adds a further, complementary gain:full tuningraises the average to 0.535, andLoRA tuningattains the best result, 0.558, improving every category over the base generator. Notably, LoRA outperforms full fine-tuning (0.558 vs. 0.535) despite updating far fewer parameters. We attribute this to the narrow nature of the adaptation: the editor does not need to relearn general video priors, but only to acquire the transferable skill of restyling a white-clay simulation draft into a photorealistic video while preserving its motion. Restricting adaptation to a low-rank subspace retains the strong visual priors of the base generator and learns this draft-to-realistic mapping, whereas full fine-tuning is more prone to overfitting the limited triplet data and drifting away from those priors. Takeaway.Each stage contributes. Drafting supplies most of the physical gain, and editor tuning refines appearance while preserving the simulated dynamics; a lightweight LoRA adaptation is suﬀicient—and preferable—to full fine-tuning. Qualitative results.Figure 3illustrates these gains: on prompts such as dry-ice sublimation and vacuum-induced bottle collapse, the OmniWeaving baseline produces plausible frames that violate the intended process, whereas VideoCoCo follows the executable draft and realizes the correct dynamics with photorealistic appearance. 5Conclusion In this paper, we introduced VideoCoCo, an agentic dual-engine framework that uses executable code as a process-level chain of thought to enforce physical consistency in text-to-video genera- tion. By decoupling spatiotemporal simulation from photorealistic editing, VideoCoCo achieves 11 state-of-the-art physical plausibility on PhyGenBench and VBench-2.0. However, our approach in- troduces additional inference latency and is bounded by the expressiveness of the underlying Blender simulator, making highly complex phenomena like turbulent fluids challenging to synthesize zero- shot. Future work will explore integrating specialized physical engines (e.g., Taichi) and investigate knowledge distillation to internalize these executable priors directly into end-to-end video models, ultimately eliminating the need for inference-time simulation. References [1]Ziqi Huang, Ning Yu, Gordon Chen, Haonan Qiu, Paul Debevec, and Ziwei Liu. Vchain: Chain-of-visual- thought for reasoning in video generation. InFindingsoftheAssociationforComputationalLinguistics (ACLFindings), 2026. [2]Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. InConferenceonLanguageModeling(COLM), 2024. [3]Tinghui Zhu, Sheng Zhang, James Y. Huang, Selena Song, Xiaofei Wen, Yuankai Li, Hoifung Poon, and Muhao Chen. Video models can reason with verifiable rewards. arXivpreprintarXiv:2605.15458, 2026. [4]Kaleb Newman, Tyler Zhu, and Olga Russakovsky. Video models reason early: Exploiting plan commit- ment for maze solving.arXivpreprintarXiv:2603.30043, 2026. [5]Fangfu Liu, Hanyang Wang, Yimo Cai, Kaiyan Zhang, Xiaohang Zhan, and Yueqi Duan. Video-t1: Test-time scaling for video generation. In IEEE/CVFInternationalConferenceonComputerVision (ICCV), 2025. [6]Sejoon Jun, Zheng Ding, Huangyuan Su, Weirui Ye, and Yilun Du. Temporal backtracking search for test-time generative video reasoning.arXivpreprintarXiv:2606.13861, 2026. [7]Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models.Advancesinneuralinformationprocessingsystems, 35:8633–8646, 2022. [8]Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data.arXiv preprintarXiv:2209.14792, 2022. [9]Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models.arXivpreprintarXiv:2210.02303, 2022. [10]Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In ProceedingsoftheIEEE/CVFconferenceoncomputervisionandpatternrecognition, pages 22563– 22575, 2023. [11]Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. InProceedingsof theIEEE/CVFconferenceoncomputervisionandpatternrecognition, pages 7310–7320, 2024. [12]Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. InternationalJournalofComputerVision, 133(5):3059–3078, 2025. [13]Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ProceedingsoftheIEEE/CVFInternationalConferenceonComputer Vision(ICCV), pages 7623–7633, October 2023. 12 [14]Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing eﬀicient video production for all.arXivpreprint arXiv:2412.20404, 2024. [15]Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprintarXiv:2412.03603, 2024. [16]Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXivpreprint arXiv:2503.20314, 2025. [17]Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXivpreprintarXiv:2408.06072, 2024. [18]Guoqing Ma, Haoyang Huang, Kun Yan, Liangyu Chen, Nan Duan, Shengming Yin, Changyi Wan, Ranchen Ming, Xiaoniu Song, Xing Chen, Yu Zhou, Deshan Sun, Deyu Zhou, Jian Zhou, Kaijun Tan, Kang An, Mei Chen, Wei Ji, Qiling Wu, Wen Sun, Xin Han, Yanan Wei, Zheng Ge, Aojie Li, Bin Wang, Bizhu Huang, Bo Wang, Brian Li, Changxing Miao, Chen Xu, Chenfei Wu, Chenguang Yu, Dapeng Shi, Dingyuan Hu, Enle Liu, Gang Yu, Ge Yang, Guanzhe Huang, Gulin Yan, Haiyang Feng, Hao Nie, Haonan Jia, Hanpeng Hu, Hanqi Chen, Haolong Yan, Heng Wang, Hongcheng Guo, Huilin Xiong, Huixin Xiong, Jiahao Gong, Jianchang Wu, Jiaoren Wu, Jie Wu, Jie Yang, Jiashuai Liu, Jiashuo Li, Jingyang Zhang, Junjing Guo, Junzhe Lin, Kaixiang Li, Lei Liu, Lei Xia, Liang Zhao, Liguo Tan, Liwen Huang, Liying Shi, Ming Li, Mingliang Li, Muhua Cheng, Na Wang, Qiaohui Chen, Qinglin He, Qiuyan Liang, Quan Sun, Ran Sun, Rui Wang, Shaoliang Pang, Shiliang Yang, Sitong Liu, Siqi Liu, Shuli Gao, Tiancheng Cao, Tianyu Wang, Weipeng Ming, Wenqing He, Xu Zhao, Xuelin Zhang, Xianfang Zeng, Xiaojia Liu, Xuan Yang, Yaqi Dai, Yanbo Yu, Yang Li, Yineng Deng, Yingming Wang, Yilei Wang, Yuanwei Lu, Yu Chen, Yu Luo, Yuchu Luo, Yuhe Yin, Yuheng Feng, Yuxiang Yang, Zecheng Tang, Zekai Zhang, Zidong Yang, Binxing Jiao, Jiansheng Chen, Jing Li, Shuchang Zhou, Xiangyu Zhang, Xinhao Zhang, Yibo Zhu, Heung-Yeung Shum, and Daxin Jiang. Step-video-t2v technical report: The practice, challenges, and future of video foundation model, 2025. URL https://arxiv.org/abs/2502.10248. [19]Haoxiao Wang, Antao Xiang, Haiyang Sun, Peilin Sun, Changhao Pan, Yifu Chen, Minjie Hong, Weijie Wang, Shuang Chen, Yue Chen, et al. Diffusion model as a generalist segmentation learner.arXiv preprintarXiv:2604.24575, 2026. [20]Zizhao Tong, Hongfeng Lai, Zeqing Wang, Zhaohu Xing, Kexu Cheng, Haoran Xu, Zhao Pu, Shangwen Zhu, Ruili Feng, Jian Zhao, et al. Scope: Simulating cross-game operations in playable environments for fps world models.arXivpreprintarXiv:2605.23345, 2026. [21]Yixuan Ye, Xuanyu Lu, Yuxin Jiang, Yuchao Gu, Rui Zhao, Qiwei Liang, Jiachun Pan, Fengda Zhang, Weijia Wu, and Alex Jinpeng Wang. Mind: Benchmarking memory consistency and action control in world models. arXivpreprintarXiv:2602.08025, 2026. [22]Ji Xie, Trevor Darrell, Luke Zettlemoyer, and XuDong Wang. Reconstruction alignment improves unified multimodal models. arXivpreprintarXiv:2509.07295, 2025. [23]Haoyu Wu, Diankun Wu, Tianyu He, Junliang Guo, Yang Ye, Yueqi Duan, and Jiang Bian. Geometry forcing: Marrying video diffusion and 3D representation for consistent world modeling. InTheFourteenth InternationalConferenceonLearningRepresentations, 2026. URLhttps://openreview.net/forum? id=ULXYZCms41. [24]Haoyu Wu, Jiwen Yu, Yingtian Zou, and Xihui Liu. MultiWorld: Scalable multi-agent multi-view video world models. arXivpreprintarXiv:2604.18564, 2026. [25]Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Leo Jing, David Schnurr, Joe Taylor, 13 Troy Luhman, Eric Luhman, et al. Video generation models as world simulators.OpenAIBlog, 1(8):1, 2024. [26]Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense-based benchmark for video generation.arXivpreprintarXiv:2410.05363, 2024. [27]Hritik Bansal, Clark Peng, Yonatan Bitton, Roman Goldenberg, Aditya Grover, and Kai-Wei Chang. Videophy-2: A challenging action-centric physical commonsense evaluation in video generation.arXiv preprintarXiv:2503.06800, 2025. [28]Ke Zhang, Cihan Xiao, Jiacong Xu, Yiqun Mei, and Vishal M Patel. Think before you diffuse: Infusing physical rules into video diffusion.arXivpreprintarXiv:2505.21653, 2025. [29]Jing Wang, Ao Ma, Ke Cao, Jun Zheng, Zhanjie Zhang, Jiasong Feng, Shanyuan Liu, Yuhang Ma, Bo Cheng, Dawei Leng, et al. Wisa: World simulator assistant for physics-aware text-to-video generation. arXivpreprintarXiv:2503.08153, 2025. [30]Qiyao Xue, Xiangyu Yin, Boyuan Yang, and Wei Gao. Phyt2v: Llm-guided iterative self-refinement for physics-grounded text-to-video generation. InProceedingsoftheComputerVisionandPattern RecognitionConference, pages 18826–18836, 2025. [31]Xindi Yang, Baolu Li, Yiming Zhang, Zhenfei Yin, Lei Bai, Liqian Ma, Zhiyong Wang, Jianfei Cai, Tien-Tsin Wong, Huchuan Lu, et al. Vlipp: Towards physically plausible video generation with vision and language informed physical prior. InProceedingsoftheIEEE/CVFInternationalConferenceon ComputerVision, pages 12360–12370, 2025. [32]Chen Wang, Chuhao Chen, Yiming Huang, Zhiyang Dou, Yuan Liu, Jiatao Gu, and Lingjie Liu. Physctrl: Generative physics for controllable and physics-grounded video generation.arXivpreprint arXiv:2509.20358, 2025. [33]Haoze Zhang, Tianyu Huang, Zichen Wan, Xiaowei Jin, Hongzhi Zhang, Hui Li, and Wangmeng Zuo. Physchoreo: Physics-controllable video generation with part-aware semantic grounding.arXivpreprint arXiv:2511.20562, 2025. [34]Runtao Liu, Haoyu Wu, Ziqiang Zheng, Chen Wei, Yingqing He, Renjie Pi, and Qifeng Chen. VideoDPO: Omni-preference alignment for video diffusion generation.arXivpreprintarXiv:2412.14167, 2024. [35]Yuanhao Cai, Kunpeng Li, Menglin Jia, Jialiang Wang, Junzhe Sun, Feng Liang, Weifeng Chen, Felix Juefei-Xu, Chu Wang, Ali Thabet, et al. Phygdpo: Physics-aware groupwise direct preference optimiza- tion for physically consistent text-to-video generation.arXivpreprintarXiv:2512.24551, 2025. [36]Peiyao Wang, Weining Wang, and Qi Li. Physcorr: Dual-reward dpo for physics-constrained text-to- video generation with automated preference selection.arXivpreprintarXiv:2511.03997, 2025. [37]Chenyu Li, Oscar Michel, Xichen Pan, Sainan Liu, Mike Roberts, and Saining Xie. Pisa experiments: Exploring physics post-training for video diffusion models by watching stuff drop.arXivpreprint arXiv:2503.09595, 2025. [38]Qiyuan Zhang, Biao Gong, Shuai Tan, Zheng Zhang, Yujun Shen, Xing Zhu, Yuyuan Li, Kelu Yao, Chunhua Shen, and Changqing Zou. Physrvg: Physics-aware unified reinforcement learning for video generative models.arXivpreprintarXiv:2601.11087, 2026. [39]Juanxi Tian, Fengyuan Liu, Jiaming Han, Yilei Jiang, Yongliang Wu, Yesheng Liu, Haodong Li, Furong Xu, and Wanhua Li. Auto-rubric as reward: From implicit preferences to explicit multimodal generative criteria. arXivpreprintarXiv:2605.08354, 2026. [40]Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Advances inNeuralInformationProcessingSystems(NeurIPS), 2022. 14 [41]Dongzhi Jiang, Renrui Zhang, Haodong Li, Zhuofan Zong, Ziyu Guo, Jun He, Claire Guo, Junyan Ye, Rongyao Fang, Weijia Li, et al. DraCo: Draft as CoT for text-to-image preview and rare concept generation.arXivpreprintarXiv:2512.05112, 2025. [42]Haodong Li, Chunmei Qing, Huanyu Zhang, Dongzhi Jiang, Yihang Zou, Hongbo Peng, Dingming Li, Yuhong Dai, ZePeng Lin, Juanxi Tian, Yi Zhou, Siqi Dai, and Jingwei Wu. CoCo: Code as CoT for text-to-image preview and rare concept generation, 2026. URLhttps://arxiv.org/abs/2603.08652. [43]Dewei Zhou, Xinyu Huang, Xun Wang, Ji Xie, Yabo Zhang, Liang Li, Kunchang Li, Zongxin Yang, and Yi Yang. MetaPoint: Unlocking precise spatial control in agentic visual generation. arXivpreprint arXiv:2606.05031, 2026. [44]Xiangpeng Yang, Ji Xie, Yiyuan Yang, Yue Ma, Yan Huang, Min Xu, and Qiang Wu. VideoCoF: Unified video editing with temporal reasoner. InProceedingsoftheIEEE/CVFConferenceonComputerVision andPatternRecognition(CVPR), pages 37940–37949, June 2026. [45]Pika Labs. Pika.https://pika.art, 2023. Accessed: 2026-07-20. [46]Runway. Introducing gen-3 alpha: A new frontier for video generation.https://runwayml.com/ research/introducing-gen-3-alpha, 2024. Accessed: 2026-07-20. [47]Kuaishou Technology. Kling ai: Text-to-video generation service.https://klingai.com/, 2024. Ac- cessed: 2026-07-20. [48]Weichen Fan, Chenyang Si, Junhao Song, Zhenyu Yang, Yinan He, Long Zhuo, Ziqi Huang, Ziyue Dong, Jingwen He, Dongwei Pan, et al. Vchitect-2.0: Parallel transformer for scaling up video diffusion models. arXivpreprintarXiv:2501.08453, 2025. [49]Wan Team. Wan: Open and advanced large-scale video generative models.arXivpreprint arXiv:2503.20314, 2025. [50]NVIDIA. Cosmos world foundation model platform for physical ai.arXivpreprintarXiv:2501.03575, 2025. [51]Yoav HaCohen et al. LTX-Video: Realtime video latent diffusion.arXivpreprintarXiv:2501.00103, 2024. [52]Kaihang Pan et al. Omniweaving: Towards unified video generation with free-form composition and reasoning.arXivpreprintarXiv:2603.24458, 2026. [53]Kuaishou Technology. Kling ai 1.6: Text-to-video generation service.https://klingai.com/, 2024. Accessed: 2026-07-20. [54]Fanqing Meng et al. Towards world simulator: Crafting physical commonsense-based benchmark for video generation. arXivpreprintarXiv:2410.05363, 2024. [55]OpenAI. GPT-4o system card.arXivpreprintarXiv:2410.21276, 2024. [56]Dian Zheng et al. VBench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness. arXivpreprintarXiv:2503.21755, 2025. 15

Instant research discovery