Paper deep dive
FlashVSR+: Efficient Video Super-Resolution via Hybrid Attention and Reward-Yield Distribution Matching
Ruchang Yao, Shi Guo, Junhao Zhuang, Mingde Yao, Xin Cai, Jiahao Zhan, Bin Fu, Xinyuan Chen, Yu Qiao, Tianfan Xue
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Recent video super-resolution (VSR) models achieve efficient one-step generation through diffusion distillation. However, high-resolution inference remains costly, and distilled models are difficult to align with human preferences. These challenges call for improvements in both model architecture and training strategy. We present \textbf{FlashVSR+} to address both limitations. On the architecture side, we introduce hybrid attention based on the strong locality observed in pretrained VSR models, while retaining sparse global interaction for long-range consistency. It achieves up to $2\times$ DiT speedup and 45\% lower peak memory, enabling spatially untiled 4K VSR on a single 24GB RTX 4090. On the training side, we propose Reward-Yield Distribution Matching Distillation (Ry-DMD) for preference-aligned one-step generation. Ry-DMD uses reward-weighted teacher rollouts to incorporate black-box rewards into distribution matching, without backpropagating through the reward model. We further show that Ry-DMD generalizes beyond VSR to T2I generation. Together, FlashVSR+ delivers stronger restoration quality and substantially improves high-resolution efficiency, with consistent gains in user preference. We release our code, models, and training data at \href{this https URL}{\textcolor{FlashBlue}{\textbf{this http URL}}}.
Tags
Links
- Source: https://arxiv.org/abs/2610.17552v1
- Canonical: https://arxiv.org/abs/2610.17552v1
Trouble viewing inline? Open PDF directly →
Full Text
60,579 characters extracted from source content.
Expand or collapse full text
FlashVSR+: Efficient Video Super-Resolution via Hybrid Attention and Reward-Yield Distribution Matching Ruchang Yao Shi Guo Junhao Zhuang Mingde Yao Xin Cai Affiliation: Shanghai AI Lab Affiliation: MMLAB CUHK Affiliation: Tsinghua University Affiliation: CPII under InnoHK† denotes corresponding authors. Jiahao Zhan Bin Fu Xinyuan Chen Yu Qiao Tianfan Xue Affiliation: Shanghai AI Lab Affiliation: MMLAB CUHK Abstract Recent video super-resolution (VSR) models achieve efficient one-step generation through diffusion distillation. However, high-resolution inference remains costly, and distilled models are difficult to align with human preferences. These challenges call for improvements in both model architecture and training strategy. We present FlashVSR+ to address both limitations. On the architecture side, we introduce hybrid attention based on the strong locality observed in pretrained VSR models, while retaining sparse global interaction for long-range consistency. It achieves up to 2×2× DiT speedup and 45% lower peak memory, enabling spatially untiled 4K VSR on a single 24GB RTX 4090. On the training side, we propose Reward-Yield Distribution Matching Distillation (Ry-DMD) for preference-aligned one-step generation. Ry-DMD uses reward-weighted teacher rollouts to incorporate black-box rewards into distribution matching, without backpropagating through the reward model. We further show that Ry-DMD generalizes beyond VSR to T2I generation. Together, FlashVSR+ delivers stronger restoration quality and substantially improves high-resolution efficiency, with consistent gains in user preference. We release our code, models, and training data at w.flashvsr.tech. Figure 1: FlashVSR+ achieves high-quality and efficient video restoration. Left: We recover fine visual details while maintaining natural textures. Right: At 4K resolution, we suppress visible artifacts and enable substantially faster inference on both H200 equivalent GPU and a RTX 4090. 1 Introduction Video super-resolution (VSR) reconstructs high-resolution videos from low-resolution inputs. It supports applications such as smartphone photography, social media, live streaming, and cost-effective upscaling of AI-generated videos to 2K or 4K resolution. Reconstruction-based VSR aggregates complementary information across frames to improve distortion oriented fidelity (Chan et al., 2022a; Chan et al., 2022b; Shi et al., 2022). These methods benefit from stronger propagation, alignment, and spatiotemporal modeling, but the perception–distortion trade-off often leads to over-smoothed textures when high-frequency details are ambiguous (Blau and Michaeli, 2018). Generative VSR addresses this limitation by utilizing learned image or video priors to synthesize perceptually realistic details (Zhou et al., 2024; Chen et al., 2024; Xu et al., 2025c; Li et al., 2025), and has achieved substantial progress in visual quality. Despite these advances, efficient and flexible generative VSR remains hindered by two bottlenecks. (1) Computational efficiency: Diffusion-based approaches rely on costly multi-step denoising trajectories (Zhou et al., 2024; Chen et al., 2024). Although recent one-step VSR methods greatly reduce this latency by distilling the multi-step generation trajectory into a single step (Song et al., 2023; Yin et al., 2024a; Yin et al., 2024b; Sun et al., 2026; Zhuang et al., 2026; Lv et al., 2026), the computation and memory costs of spatiotemporal attention remain prohibitive for high-resolution videos (e.g., 2K and 4K resolution). As a result, inference on consumer-grade GPUs (e.g., RTX 4090) often requires spatial tiling, which limits global context and may introduce boundary artifacts, as illustrated in Fig. 1. (2) Preference-aligned distillation: Standard one-step distillation method DMD (Yin et al., 2024a; Yin et al., 2024b) matches the teacher distribution but cannot explicitly favor outputs that better align with human preferences, such as more accurate text rendering. Existing reward-guided methods typically rely on multi-step stochastic trajectories (Black et al., 2024; Fan et al., 2023) or differentiable reward signals (Clark et al., 2024; Jia et al., 2025), making them ill-suited to deterministic one-step generators with black-box rewards. To address these bottlenecks, we present FlashVSR+, a one-step VSR framework with hybrid attention for efficient high-resolution restoration and Ry-DMD for preference-aligned distillation with black-box rewards. Together, they make one-step generative VSR more efficient, enabling spatially untiled 4K VSR inference on a single 24GB RTX 4090, while better aligning outputs with human preferences. For high-resolution efficiency, we find that pretrained VSR models such as FlashVSR (Zhuang et al., 2026) exhibit strong spatiotemporal attention locality, but with substantial variation across layers. Attention is highly concentrated within compact neighborhoods in many blocks, suggesting that dense attention is unnecessary throughout the backbone, while fully local attention would be overly restrictive. We therefore introduce a hybrid attention architecture that replaces most original attention blocks with efficient local spatiotemporal attention while retaining a small number of sparse global blocks. This design achieves up to 2×2× DiT acceleration and nearly halves peak memory, enabling spatially untiled 4K VSR on a single 24GB RTX 4090 while maintaining comparable restoration quality. For preference-aligned distillation, one-step generators are particularly challenging because stochasticity is limited to the input seed, while black-box rewards provide no gradients for direct optimization. We propose Reward-Yield Distribution Matching Distillation (Ry-DMD), which uses multiple teacher rollouts to expose the stochasticity available to a one-step generator and reweights them with black-box rewards. The difference between the reward-weighted and original teacher predictions provides a preference direction, which is added to standard DMD to favor higher-reward teacher-supported outputs. We validate Ry-DMD on OCR-guided VSR and further demonstrate its applicability to text-to-image generation. To further scale up one-step VSR distillation, a natural strategy is to increase the training budget and data size. However, we find that longer DMD training does not yield consistent gains, revealing non-monotonic behavior. Instead, student initialization and teacher capacity play a more important role in final performance. 2 Related Work 2.1 Video Super-Resolution Conventional VSR methods recover missing details through alignment, propagation, and spatiotemporal aggregation (Wang et al., 2019; Chan et al., 2021; Chan et al., 2022a; Shi et al., 2022; Liang et al., 2022), but often produce over-smoothed textures in perceptually ambiguous regions (Blau and Michaeli, 2018). Generative VSR instead leverages learned priors to synthesize realistic details, including adversarial (Chu et al., 2020; Xu et al., 2025c) and diffusion-based approaches (Rota et al., 2024; Zhou et al., 2024; Chen et al., 2024; Li et al., 2025; Chen et al., 2026c). To reduce the cost of iterative sampling, recent methods increasingly target one-step restoration through direct fine-tuning, adversarial post-training, or distillation (Liu et al., 2025; Chen et al., 2026d; Wang et al., 2026; Sun et al., 2026; Zhuang et al., 2026; Lv et al., 2026). Recent image SR and restoration methods have also explored reward or preference guided post-training (Yi et al., 2026; Zhao et al., 2026; Xu et al., 2025a). Nevertheless, one-step VSR still faces challenges in both high-resolution inference efficiency and reward-guided post-training. FlashVSR+ targets these two aspects with hybrid attention and Ry-DMD, respectively. 2.2 Efficient Attention for Visual Generation The quadratic cost of dense attention has motivated sparse and sub-quadratic alternatives for visual generation. Video diffusion methods exploit structured spatiotemporal sparsity or locality-aware computation to reduce attention cost (Xi et al., 2025; Yang et al., 2026; Shen et al., 2025; Xu et al., 2025b; Li et al., 2026). Another line of work adopts efficient attention backbones, including linear attention and hybrid architectures that periodically retain full softmax attention for global information exchange (Xie et al., 2025a; Xie et al., 2025b; Chen et al., 2026b; Chen et al., 2026a). In contrast, FlashVSR+ exploits the locality observed in a pretrained VSR model, replacing most global-attention blocks with spatiotemporal local attention while retaining a small number of global blocks. 2.3 Reward Optimization for One/Few-Step Generative Models Reward-guided post-training has been extensively studied for multi-step diffusion models, where denoising trajectories enable policy optimization, differentiable reward backpropagation, and preference learning (Fan et al., 2023; Black et al., 2024; Clark et al., 2024; Wallace et al., 2024; Liu et al., 2026; Zheng et al., 2026). Extending these approaches to few-step and one-step generators remains challenging. While few-step models still preserve short denoising trajectories, one-step generators directly map noise to samples and lack such intermediate optimization signals. Existing few-step approaches explore reward-guided distribution matching and related optimization strategies (Jiang et al., 2026; Huang et al., 2026; Dong et al., 2026), while one-step methods typically rely on inference-time optimization or differentiable reward surrogates (Eyring et al., 2024; Jia et al., 2025; Zhu et al., 2026). Ry-DMD instead enables black-box reward optimization for deterministic one-step generators by introducing a reward-tilted teacher target within the DMD framework. 3 Method 3.1 Overview Figure 2: Overview of the FlashVSR+ training pipeline. Stage 1 trains a 14B full-attention multi-step teacher on image and video data. Stage 2 trains the hybrid-attention student to obtain a strong initialization. Stage 3 distills the student into a one-step generator using Ry-DMD, where reward feedback modifies the teacher term estimation without requiring reward gradients. Following FlashVSR (Zhuang et al., 2026), FlashVSR+ adopts the three-stage training framework shown in Fig. 2. In Stage 1, we train a 14B full-attention multi-step model with flow matching as the teacher. In Stage 2, we replace the 1.3B backbone with our hybrid-attention architecture and train it to provide a strong initialization for one-step generation. In Stage 3, we distill the initialized student into a one-step VSR model using distribution matching together with the pixel loss. Ry-DMD further injects scalar reward feedback into the teacher term without reward backpropagation. Training details are provided in Appendix A. 3.2 Efficient One-Step VSR with Hybrid Attention Figure 3: Attention locality and adaptation in causal VSR. (a) Attention concentrates in compact spatiotemporal neighborhoods. (b) Local attention mass varies across DiT blocks, and hybrid training reduces the locality of retained global-attention blocks. Spatiotemporal attention is a major scalability bottleneck for causal VSR at high resolution. To understand the potential redundancy of full attention, we examine the attention patterns of a pretrained one-step VSR model. As shown in Fig. 3(a), attention is highly concentrated within local neighborhoods, which cover only ∼1% 1\% of valid key blocks yet receive ∼20% 20\% of the total attention mass. We quantify this behavior using the local attention mass, defined as the fraction of attention assigned to valid local key blocks. Higher local attention mass indicates stronger concentration within local neighborhoods, suggesting that many non-local interactions are weakly weighted yet still require computation, leading to unnecessary overhead. As shown in Fig. 3(b), local attention is strongly enriched across DiT blocks but varies substantially across layers, motivating a hybrid rather than purely local design. Motivated by this observation, we introduce hybrid attention, which replaces most attention blocks with efficient spatiotemporal local attention while retaining sparse global-attention blocks. Specifically, we denote the local attention configuration as (s,t,b)(s,t,b), where s and t denote the numbers of spatial and preceding temporal key blocks used in local attention, and b denotes the number of DiT blocks replaced by local attention. We then analyze how attention adapts after hybrid training. For visualization, we use the setting (s,t,b)=(3,1,20)(s,t,b)=(3,1,20), with after-training measurements taken only on retained global-blocks. As shown in Fig. 3(b), the retained global-attention blocks exhibit reduced local attention mass after hybrid training, indicating a shift toward non-local interactions. This suggests that hybrid attention allows the retained global-attention blocks to better capture non-local interactions, leading to more efficient attention computation. In the final model, we adopt a regular interleaving of local and global blocks and use (s,t,b)=(3,1,25)(s,t,b)=(3,1,25), retaining one global-attention block every five DiT blocks (Fig. 2, Stages 2&3). This design achieves up to 2×2× DiT acceleration and nearly halves peak memory usage, with detailed evaluations in Section 4.2. 3.3 Reward-Yield Distribution Matching Distillation Figure 4: Conceptual illustration of Ry-DMD. Adapting a deterministic one-step generator to a downstream objective or preferences is not straightforward. Continuing standard DMD on task-specific data can adapt the training distribution, but does not explicitly favor outputs with better human preference. Direct reward optimization within DMD is possible, but non-differentiable rewards require differentiable surrogates for backpropagation. We propose Reward-Yield Distribution Matching Distillation (Ry-DMD). As illustrated in Fig. 4, Ry-DMD transforms the teacher distribution toward higher-reward outputs, such as clearer and more recognizable text, by constructing a reward-tilted target. The resulting target shift is then converted into a correction to the DMD teacher term. This preserves the overall DMD training framework while enabling the use of black-box objectives such as OCR, which we validate on VSR in Sec. 4.3; we further demonstrate its generality on text-to-image generation in Sec. 4.4. 3.3.1 Reward-Tilted Distribution Matching Our goal is to modify the teacher distribution toward higher-reward regions while preserving its original support. As a distribution-level motivation, we consider an exponential tilt of the clean teacher distribution. Let c denote the condition and pT,0(x0∣c)p_T,0(x_0 c) the clean teacher distribution. Given a scalar reward r(x0,c)r(x_0,c), we define the reward-tilted distribution as p0⋆(x0∣c)∝pT,0(x0∣c)exp(βr(x0,c)),p_0 (x_0 c) p_T,0(x_0 c) \! (β r(x_0,c) ), (1) where β>0β>0 controls the strength of reward tilting. This reweighting increases the relative probability of higher-reward samples while remaining within the support of the teacher distribution. As β→0β→ 0, p0⋆p_0 reduces to pT,0p_T,0, recovering the standard teacher-matching objective. Equivalently, Eq. 1 arises as the solution to a KL-regularized expected-reward objective; the derivation is provided in Appendix E.1. 3.3.2 Local Reward-Tilted Teacher Target The global target in Eq. 1 is not directly available during DMD training. Instead, at each noisy student state, we sample multiple teacher continuations and use their rewards to estimate how reward tilting would shift the local teacher target. During DMD training, the one-step generator predicts x0θ=Gθ(z,c)x_0^θ=G_θ(z,c), which is re-noised to xσ=(1−σ)x0θ+σϵ,ϵ∼(0,I).x_σ=(1-σ)x_0^θ+σε, ε (0,I). (2) Starting from xσx_σ, a stochastic K-step teacher rollout produces x0=T(xσ,c,η)x_0=T_T(x_σ,c;η), where η denotes the stochastic noise used during the rollout. These rollouts provide a Monte Carlo approximation of the conditional teacher endpoint distribution qTK(x0∣xσ,c)q_T^K(x_0 x_σ,c). Extending the reward-tilting principle in Eq. 1 to this conditional distribution, we define a reward-reweighted rollout distribution as qT,βK(x0∣xσ,c)∝qTK(x0∣xσ,c)exp(βr(x0,c)).q_T,β^K(x_0 x_σ,c) q_T^K(x_0 x_σ,c) \! (β r(x_0,c) ). (3) We estimate the resulting target shift using J independent teacher rollouts x0(j)j=1J\x_0^(j)\_j=1^J with rewards rj=r(x0(j),c)r_j=r(x_0^(j),c). Let πj _j denote the normalized reward weight of the j-th rollout, with πj∝exp(βaj) _j (β a_j) and aj=(rj−r¯)/(sr+ε)a_j=(r_j- r)/(s_r+ ), where rewards are standardized within each rollout group for robustness to reward scale. We then compute the original and reward-weighted teacher endpoint means as m^T=J−1∑j=1Jx0(j),m^β=∑j=1Jπjx0(j). m_T=J^-1 _j=1^Jx_0^(j), m_β= _j=1^J _jx_0^(j). (4) and use their difference δm=m^β−m^Tδ m= m_β- m_T (5) as the reward-induced shift of the local teacher target. Crucially, δmδ m is estimated entirely from teacher samples and scalar reward evaluations, without differentiating through the reward function. The normalized form of the local tilt, the Monte Carlo estimator, and additional interpretation are provided in Appendix E.2. 3.3.3 Reward-Yield DMD Update The reward-weighted rollouts tell us how the teacher endpoint target should move to favor higher-reward outputs. We incorporate this shift directly into the teacher term of DMD, while leaving the remaining distribution-matching procedure unchanged. Specifically, we convert the estimated endpoint displacement into a correction of the teacher term: sTRY=sT+λRYδmmax(σ,εσ),s_T^RY=s_T+ _RY δ m (σ, _σ), (6) where λRY _RY controls the reward-correction strength and εσ _σ stabilizes the update at small noise levels. The conversion from the endpoint displacement to the teacher correction is provided in Appendix E.3. 4 Experiments 4.1 Comparison with State-of-the-Art VSR Methods Table 1: Human preference study. Preference margins relative to Ours-Base. Method Fidelity Quality Overall RealViFormer −70%-70\% −72%-72\% −73%-73\% DOVE −68%-68\% −68%-68\% −70%-70\% DLoRAL −38%-38\% −28%-28\% −35%-35\% SeedVR2 −53%-53\% −55%-55\% −56%-56\% PS-SR −38%-38\% −38%-38\% −41%-41\% FlashVSR −15%-15\% −10%-10\% −15%-15\% Ours-Turbo −7%-7\% −5%-5\% −6%-6\% Ours-Base 0%0\% 0%0\% 0%0\% We evaluate on synthetic benchmarks, including YouHQ40 (Zhou et al., 2024), SPMCS (Tao et al., 2017), and RealisVideo-4K (Zhao et al., 2025), as well as the real-world VideoLQ (Chan et al., 2022b) and AI-generated AIGC35 benchmarks. Table 2 compares FlashVSR+ with representative VSR methods, including RealViFormer (Zhang and Yao, 2024), DOVE (Chen et al., 2026d), SeedVR2-3B (Wang et al., 2026), PS-SR (Wu et al., 2026), and FlashVSR (Zhuang et al., 2026). We report PSNR, SSIM, and LPIPS (Zhang et al., 2018) on datasets with ground truth, and MUSIQ (Ke et al., 2021), CLIPIQA (Wang et al., 2023), NIQE (Mittal et al., 2012), and DOVER (Wu et al., 2023a) on all datasets. FlashVSR+ achieves strong overall restoration quality, with qualitative results in Fig. 5 showing better preservation of text, fine textures, and natural details with fewer spurious artifacts. We further conduct a blinded pairwise user study following the GSB protocol used in (Wang et al., 2026). Ours-Base serves as the anchor, and each method pair is evaluated on 20 videos randomly sampled from RealisVideo-4K, VideoLQ, and AIGC35 benchmarks. Participants compare the two outputs in terms of visual fidelity, perceptual quality, and overall preference. Table 1 reports the resulting preference margins relative to Ours-Base. Ours-Base is preferred over all compared methods, while Ours-Turbo remains close despite its higher inference efficiency. Figure 5: Qualitative comparison with representative VSR methods. Zoomed regions highlight the reconstruction of text, fine textures, and facial details. Table 2: Quality comparison with state-of-the-art VSR methods under varying inference costs. Top-1/2/3 results are highlighted in red/orange/yellow. Dataset Metric RealViFormer DOVE SeedVR2-3B PS-SR FlashVSR Ours-Base Ours-Turbo YouHQ40 PSNR ↑ 23.50 20.91 23.34 22.35 22.53 23.64 23.51 SSIM ↑ 0.6051 0.5532 0.6206 0.5899 0.5787 0.6250 0.6097 LPIPS ↓ 0.3528 0.3514 0.2768 0.3122 0.2721 0.2698 0.2728 MUSIQ ↑ 60.36 60.82 57.17 66.20 65.91 64.71 63.91 CLIPIQA ↑ 0.3823 0.3964 0.3703 0.5063 0.4701 0.4257 0.4322 NIQE ↓ 3.490 4.581 4.213 3.72 3.385 3.560 3.737 DOVER ↑ 10.42 10.97 12.23 12.93 13.47 13.33 13.37 SPMCS PSNR ↑ 22.64 23.09 22.24 21.16 21.35 22.34 22.43 SSIM ↑ 0.5651 0.6109 0.5867 0.5452 0.5398 0.5714 0.5654 LPIPS ↓ 0.3368 0.2738 0.2521 0.3240 0.2680 0.2700 0.2647 MUSIQ ↑ 61.91 68.93 65.63 69.47 69.59 70.09 70.13 CLIPIQA ↑ 0.3416 0.4605 0.4633 0.4650 0.4618 0.4753 0.4546 NIQE ↓ 3.781 4.234 3.601 3.971 3.389 3.401 3.411 DOVER ↑ 7.428 8.352 8.462 8.907 9.250 9.676 9.811 Realis-4K PSNR ↑ 30.81 23.26 30.88 28.35 30.84 31.23 32.06 SSIM ↑ 0.8240 0.7841 0.8206 0.8517 0.8596 0.8639 0.8724 LPIPS ↓ 0.1631 0.213 0.1929 0.1970 0.1671 0.1639 0.1625 DOVER ↑ 11.35 12.60 12.93 12.20 13.27 13.43 13.30 VideoLQ NIQE ↓ 4.485 4.915 5.114 5.00 4.309 4.619 4.784 MUSIQ ↑ 49.40 45.79 38.70 45.34 49.70 47.64 46.93 CLIPIQA ↑ 0.3117 0.3009 0.233 0.3066 0.3140 0.2939 0.3013 DOVER ↑ 7.217 7.542 5.955 7.769 7.811 8.045 7.827 AIGC35 NIQE ↓ 4.983 5.173 4.816 4.923 4.888 4.851 5.014 MUSIQ ↑ 48.08 48.78 47.40 50.08 50.83 50.30 49.53 CLIPIQA ↑ 0.3871 0.4130 0.3945 0.4546 0.4058 0.4012 0.4001 DOVER ↑ 10.91 12.21 11.86 11.76 12.63 13.01 12.57 Figure 6: Efficiency comparison on 89-frame videos. (a) GPU time and peak memory on NVIDIA H200 equivalent GPU. (b) End-to-end wall-clock inference time on NVIDIA RTX 4090, including CPU KV-cache offloading and asynchronous prefetching overhead. (a) NVIDIA H200 equivalent GPU Method Backbone Params (B) GPU Time (s) ↓ Peak Memory (GB) ↓ 1080p 2K 4K 1080p 2K 4K DOVE 5.0 89.38 165.3 374.7 26.54 26.22 26.56 PS-SR 1.3 228.4 1123.1 1751.4 47.79 50.63 50.63 SeedVR2 3.0 51.46 157.8† 410.5† 111.21 49.57† 51.76† Ours-Base‡ 1.3 7.91 17.28 46.34 13.11 21.29 45.00 Ours-Turbo 1.3 6.63 12.31 30.59 8.72 13.23 26.33 (b) NVIDIA RTX 4090 † SeedVR2 full-frame inference OOMs so we use tiled inference. ‡ Ours-Base shares the same architecture and inference procedure as FlashVSR on H200 equivalent GPU. FlashVSR’s attention operators are unsupported on RTX 4090. Figure 7: Quality and efficiency of trained hybrid attention. The (3,1,25)(3,1,25) configuration achieves the best balance between DiT cost and restoration quality. We achieve substantial efficiency gains for high-resolution VSR. We evaluate computational efficiency in Figure 6 using 89-frame videos at 1080p, 2K, and 4K resolutions on NVIDIA H200 equivalent GPU and consumer RTX 4090 GPUs. We substantially reduce both inference latency and peak memory compared with the base model, enabling high-resolution restoration with much lower computational overhead. In particular, with additional memory optimizations, we enable 4K VSR inference on a single RTX 4090; implementation details are provided in Appendix B. 4.2 Efficient VSR with Hybrid Attention We evaluate hybrid attention both before and after adaptation. The training-free setting tests how much of the pretrained attention can be localized directly, while the adapted setting measures the resulting quality and efficiency after training the modified architecture. We first consider the training-free setting. We use (3,1,15)(3,1,15) as a conservative reference and vary one factor at a time. Since the attention pattern is modified without retraining, we only count clear structural failures as artifacts, including block or grid-shaped corruption and severe temporal ghosting; minor changes in sharpness or texture are not considered failures. The pretrained model tolerates (3,1,15)(3,1,15), but reducing the spatial or temporal context, or increasing the number of localized blocks, quickly introduces artifacts. In particular, (1,1,15)(1,1,15), (3,0,15)(3,0,15), and configurations with b≥20b≥ 20 fail under direct replacement, motivating explicit adaptation of the modified attention structure. We next adapt these hybrid-attention configurations. As shown in Fig. 7, adaptation substantially extends the usable range of hybrid attention. Increasing the number of local-attention blocks reduces latency and peak memory while preserving restoration quality up to (3,1,25)(3,1,25), whereas further localization to (3,1,30)(3,1,30) causes a clear quality drop. We therefore adopt (3,1,25)(3,1,25) as the default configuration. The more restrictive (1,1,15)(1,1,15) and (3,0,15)(3,0,15) configurations still exhibit visible artifacts after adaptation and are therefore excluded from the figure. 4.3 Reward-Guided VSR Post-Training Figure 8: Qualitative comparison of OCR-guided VSR post-training. Ry-DMD restores more recognizable text, while DMDR often introduces artifacts. We evaluate Ry-DMD on OCR-guided VSR post-training, a task requiring fine-grained visual fidelity (Yuan et al., 2026), using a synthetic text-focused dataset with annotated text ground truth; details of the data generation pipeline are provided in Appendix F. All methods start from the same pretrained one-step VSR model. Following the standard practice in Flow-GRPO (Liu et al., 2026), given target text y and the text y y recognized by PaddleOCR (Cui et al., 2025), we compute the OCR reward as rOCR=1−min(dLev(y^,y),|y|)|y|,r_OCR=1- (d_Lev( y,y),|y|)|y|, where both strings are stripped of spaces before evaluation. Table 3: Quantitative results of OCR-guided post-training. Method OCR ↑ PSNR ↑ LPIPS ↓ MUSIQ ↑ DOVER ↑ Baseline 0.624 23.94 0.246 75.35 9.79 TextSFT 0.656 22.02 0.295 76.41 10.58 DMDR 0.527 19.01 0.374 74.17 8.48 Ry-DMD 0.776 24.86 0.240 75.80 10.90 We compare four post-training strategies. Baseline denotes the original pretrained model without further adaptation. TextSFT continues standard DMD training on the text-focused dataset, testing whether task-specific data alone is sufficient. DMDR (Jiang et al., 2026) directly introduces an OCR optimization objective into DMD training; since the OCR reward is non-differentiable, we use a CTC-based differentiable surrogate and backpropagate through the OCR model. Instead, Ry-DMD directly uses the scalar OCR reward without requiring its gradient. As shown in Table 3, Ry-DMD achieves the strongest OCR improvement while preserving restoration quality. The qualitative results in Fig. 8 further show clearer text restoration with fewer artifacts. 4.4 Generalization to Text-to-Image Generation We further evaluate whether Ry-DMD generalizes beyond restoration by applying it to one-step SDXL text-to-image generation. We initialize the student from a one-step DMD2 model (Yin et al., 2024a) distilled from SDXL (Podell et al., 2024). Training uses OCR prompts from Flow-GRPO (Liu et al., 2026) together with a disjoint training split of DrawBench (Saharia et al., 2022). We optimize a combined reward consisting of OCR and PickScore. Ry-DMD relies only on scalar reward evaluations and does not backpropagate through either scorer. Table 4: Reward-guided text-to-image generation on SDXL. Model NFE ↓ OCR ↑ PickScore ↑ HPSv2 ↑ ImageReward ↑ Aesthetic ↑ SDXL 50 0.125 22.2 27.5 0.87 5.48 DMD2 4 0.098 22.5 29.2 0.83 5.47 One-step DMD2 1 0.087 22.0 26.8 0.67 5.39 Ry-DMD (ours) 1 0.241 22.7 29.6 1.02 5.44 We report PickScore (Kirstain et al., 2023), HPSv2 (Wu et al., 2023b), ImageReward (Xu et al., 2023), and the LAION aesthetic predictor (Schuhmann et al., 2022). As shown in Table 4, Ry-DMD improves the OCR score from 0.087 to 0.241, surpassing the 50-step teacher, while maintaining or improving complementary quality metrics. Ry-DMD also generalizes to four-step SD3.5-M, improving OCR accuracy from 0.501 to 0.824 without increasing NFE. As shown in Fig. 9, Ry-DMD improves text rendering while preserving scene content and overall visual quality. Additional results and hyperparameter analysis are provided in Appendices C and D. Figure 9: Qualitative comparison of Ry-DMD on 4-step SD3.5-M. Ry-DMD substantially improves text correctness and legibility while preserving the intended scene content. Beyond text rendering, the resulting images often exhibit cleaner structures, more coherent materials and lighting, and improved overall visual quality. 4.5 Rethinking DMD Scaling for One-Step VSR Figure 10: Ablation on DMD steps. To understand how to effectively scale up one-step VSR training, we first investigate whether allocating a larger optimization budget can consistently improve DMD performance. To reduce domain bias, we construct a fixed held-out evaluation set by sampling from multiple benchmarks, including AIGC, YouHQ, and VideoLQ, covering both synthetic and real-world content with diverse degradations. Intermediate checkpoints are evaluated using MUSIQ and DOVER. As shown in Fig. 10, both metrics improve rapidly during early DMD training. However, the gains do not persist: after an early peak, performance becomes non-monotonic and can decline with continued optimization. Simply increasing the number of DMD updates therefore does not reliably yield a better one-step VSR model. Table 5: Ablation on student initialization and teacher capacity. Init. Teacher MUSIQ ↑ DOVER ↑ Base 1.3B 45.87 8.512 Strong 1.3B 48.20 8.738 Base 14B 51.76 9.144 Strong 14B 52.78 9.417 We next examine two other components of the distillation setup: student initialization and teacher capacity. As shown in Table 5, both a stronger initialization and a larger teacher consistently improve the final student, with their combination performing best. These results indicate that effective DMD scaling depends on the overall distillation setup rather than optimization duration alone. 5 Conclusion We introduce FlashVSR+, combining hybrid spatiotemporal attention with Ry-DMD for efficient and reward-guided one-step VSR. The framework improves the quality efficiency trade-off and enables spatially untiled 4K inference on a single 24GB RTX 4090. Ry-DMD enables black-box reward optimization without reward backpropagation, with benefits demonstrated in both VSR and text-to-image generation. References Black et al. (2024) K. Black, M. Janner, Y. Du, I. Kostrikov, and S. Levine Training diffusion models with reinforcement learning. In International Conference on Learning Representations, Vol. 2024, p. 4965–4987. Cited by: §1, §2.3. Blau and Michaeli (2018) Y. Blau and T. Michaeli The perception-distortion tradeoff. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 6228–6237. Cited by: §1, §2.1. Chan et al. (2021) K. C. Chan, X. Wang, K. Yu, C. Dong, and C. C. Loy Basicvsr: the search for essential components in video super-resolution and beyond. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 4945–4954. Cited by: §2.1. Chan et al. (2022a) K. C. Chan, S. Zhou, X. Xu, and C. C. Loy Basicvsr++: improving video super-resolution with enhanced propagation and alignment. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 5962–5971. Cited by: §1, §2.1. Chan et al. (2022b) K. C. Chan, S. Zhou, X. Xu, and C. C. Loy Investigating tradeoffs in real-world video super-resolution. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 5952–5961. Cited by: §1, §4.1. Chen et al. (2026a) J. Chen, J. Yu, Y. Li, S. Xue, H. Liu, J. Xin, Y. Zhao, T. Ye, Z. Wu, Z. Wang, et al. SANA-video 2.0: hybrid linear attention with attention residuals for efficient video generation. arXiv preprint arXiv:2607.21553. Cited by: §2.2. Chen et al. (2026b) J. Chen, Y. Zhao, J. Yu, R. Chu, J. Chen, S. Yang, X. Wang, Y. Pan, Z. Daquan, H. Ling, et al. Sana-video: efficient video generation with block linear diffusion transformer. In International Conference on Learning Representations, Vol. 2026, p. 38973–38997. Cited by: §2.2. Chen et al. (2026c) J. Chen, J. Dong, L. Sun, Y. Yang, and J. Pan Stcdit: spatio-temporally consistent diffusion transformer for high-quality video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 38281–38290. Cited by: §2.1. Chen et al. (2026d) Z. Chen, Z. Zou, K. Zhang, X. Su, X. Yuan, Y. Guo, and Y. Zhang Dove: efficient one-step diffusion model for real-world video super-resolution. Advances in Neural Information Processing Systems 38, p. 85218–85237. Cited by: §2.1, §4.1. Chen et al. (2024) Z. Chen, F. Long, Z. Qiu, T. Yao, W. Zhou, J. Luo, and T. Mei Learning spatial adaptation and temporal coherence in diffusion models for video super-resolution. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 9232–9241. Cited by: §1, §1, §2.1. Chu et al. (2020) M. Chu, Y. Xie, J. Mayer, L. Leal-Taixé, and N. Thuerey Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics (TOG) 39 (4), p. 75–1. Cited by: §2.1. Clark et al. (2024) K. Clark, P. Vicol, K. Swersky, and D. Fleet Directly fine-tuning diffusion models on differentiable rewards. In International Conference on Learning Representations, Vol. 2024, p. 4793–4822. Cited by: §1, §2.3. Cui et al. (2025) C. Cui, T. Sun, M. Lin, T. Gao, Y. Zhang, J. Liu, X. Wang, Z. Zhang, C. Zhou, H. Liu, et al. Paddleocr 3.0 technical report. arXiv preprint arXiv:2507.05595. Cited by: §4.3. Dong et al. (2026) L. Dong, R. Guo, G. Bai, Z. Yuan, Y. Luo, and C. Zou Guiding distribution matching distillation with gradient-based reinforcement learning. arXiv preprint arXiv:2604.19009. Cited by: §2.3. Eyring et al. (2024) L. Eyring, S. Karthik, K. Roth, A. Dosovitskiy, and Z. Akata Reno: enhancing one-step text-to-image models through reward-based noise optimization. Advances in Neural Information Processing Systems 37, p. 125487–125519. Cited by: §2.3. Fan et al. (2023) Y. Fan, O. Watkins, Y. Du, H. Liu, M. Ryu, C. Boutilier, P. Abbeel, M. Ghavamzadeh, K. Lee, and K. Lee Dpok: reinforcement learning for fine-tuning text-to-image diffusion models. Advances in neural information processing systems 36, p. 79858–79885. Cited by: §1, §2.3. Huang et al. (2026) Y. Huang, X. Zhou, R. Wang, C. Zhang, J. Zhang, and T. Pang Reinforcing few-step generators via reward-tilted distribution matching. arXiv preprint arXiv:2605.26108. Cited by: §2.3. Jia et al. (2025) Z. Jia, Y. Nan, H. Zhao, and G. Liu Reward fine-tuning two-step diffusion models via learning differentiable latent-space surrogate reward. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 12912–12922. Cited by: §1, §2.3. Jiang et al. (2026) D. Jiang, D. Liu, Z. Wang, Q. Wu, L. Li, H. Li, X. Jin, C. Lu, Z. Li, M. Wang, et al. Distribution matching distillation meets reinforcement learning. In European Conference on Computer Vision, p. 281–299. Cited by: §2.3, §4.3. Ke et al. (2021) J. Ke, Q. Wang, Y. Wang, P. Milanfar, and F. Yang Musiq: multi-scale image quality transformer. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), p. 5128–5137. Cited by: §4.1. Kirstain et al. (2023) Y. Kirstain, A. Polyak, U. Singer, S. Matiana, J. Penna, and O. Levy Pick-a-pic: an open dataset of user preferences for text-to-image generation. Advances in neural information processing systems 36, p. 36652–36663. Cited by: §4.4. Li et al. (2025) X. Li, Y. Liu, S. Cao, Z. Chen, S. Zhuang, X. Chen, Y. He, Y. Wang, and Y. Qiao Diffvsr: revealing an effective recipe for taming robust video super-resolution against complex degradations. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV), p. 15319–15328. Cited by: §1, §2.1. Li et al. (2026) X. Li, M. Li, T. Cai, H. Xi, S. Yang, Y. Lin, L. Zhang, S. Yang, J. Hu, K. Peng, et al. Radial attention: (o(nlogn)) sparse attention with energy decay for long video generation. Advances in Neural Information Processing Systems 38, p. 16822–16852. Cited by: §2.2. Liang et al. (2022) J. Liang, Y. Fan, X. Xiang, R. Ranjan, E. Ilg, S. Green, J. Cao, K. Zhang, R. Timofte, and L. V. Gool Recurrent video restoration transformer with guided deformable attention. Advances in neural information processing systems 35, p. 378–393. Cited by: §2.1. Liu et al. (2026) J. Liu, G. Liu, J. Liang, Y. Li, J. Liu, X. Wang, P. Wan, D. Zhang, and W. Ouyang Flow-grpo: training flow matching models via online rl. Advances in neural information processing systems 38, p. 40783–40818. Cited by: §2.3, §4.3, §4.4. Liu et al. (2025) Y. Liu, J. Pan, Y. Li, Q. Dong, C. Zhu, Y. Guo, and F. Wang Ultravsr: achieving ultra-realistic video super-resolution with efficient one-step diffusion space. In Proceedings of the 33rd ACM International Conference on Multimedia, p. 7785–7794. Cited by: §2.1. Lv et al. (2026) Z. Lv, M. Xia, X. Wang, and K. K. Wong DUO-vsr: dual-stream distillation for one-step video super-resolution. arXiv preprint arXiv:2603.22271. Cited by: §1, §2.1. Mittal et al. (2012) A. Mittal, R. Soundararajan, and A. C. Bovik Making a “completely blind” image quality analyzer. IEEE Signal processing letters 20 (3), p. 209–212. Cited by: §4.1. Podell et al. (2024) D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. Müller, J. Penna, and R. Rombach Sdxl: improving latent diffusion models for high-resolution image synthesis. In International Conference on Learning Representations, Vol. 2024, p. 1862–1874. Cited by: §4.4. Rota et al. (2024) C. Rota, M. Buzzelli, and J. van de Weijer Enhancing perceptual quality in video super-resolution through temporally-consistent detail synthesis using diffusion models. In European Conference on Computer Vision, p. 36–53. Cited by: §2.1. Saharia et al. (2022) C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gontijo Lopes, B. Karagol Ayan, T. Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems 35, p. 36479–36494. Cited by: §4.4. Schuhmann et al. (2022) C. Schuhmann, R. Beaumont, R. Vencu, C. Gordon, R. Wightman, M. Cherti, T. Coombes, A. Katta, C. Mullis, M. Wortsman, et al. Laion-5b: an open large-scale dataset for training next generation image-text models. Advances in neural information processing systems 35, p. 25278–25294. Cited by: §4.4. Shen et al. (2025) X. Shen, C. Han, Y. Zhou, Y. Xie, Y. Gong, Q. Wang, Y. Wang, Y. Wang, P. Zhao, and J. Gu Draftattention: fast video diffusion via low-resolution attention guidance. arXiv preprint arXiv:2505.14708. Cited by: §2.2. Shi et al. (2022) S. Shi, J. Gu, L. Xie, X. Wang, Y. Yang, and C. Dong Rethinking alignment in video super-resolution transformers. Advances in Neural Information Processing Systems 35, p. 36081–36093. Cited by: §1, §2.1. Song et al. (2023) Y. Song, P. Dhariwal, M. Chen, and I. Sutskever Consistency models. arXiv preprint arXiv:2303.01469. Cited by: §1. Sun et al. (2026) Y. Sun, L. Sun, S. Liu, R. Wu, Z. Zhang, and L. Zhang One-step diffusion for detail-rich and temporally consistent video super-resolution. Advances in Neural Information Processing Systems 38, p. 172821–172841. Cited by: §1, §2.1. Tao et al. (2017) X. Tao, H. Gao, R. Liao, J. Wang, and J. Jia Detail-revealing deep video super-resolution. In 2017 IEEE International Conference on Computer Vision (ICCV), p. 4482–4490. Cited by: §4.1. Wallace et al. (2024) B. Wallace, M. Dang, R. Rafailov, L. Zhou, A. Lou, S. Purushwalkam, S. Ermon, C. Xiong, S. Joty, and N. Naik Diffusion model alignment using direct preference optimization. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 8228–8238. Cited by: §2.3. Wang et al. (2023) J. Wang, K. C. Chan, and C. C. Loy Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, Vol. 37, p. 2555–2563. Cited by: §4.1. Wang et al. (2026) J. Wang, S. Lin, Z. Lin, Y. Ren, M. Wei, Z. Yue, S. Zhou, H. Chen, Y. Zhao, C. Yang, et al. Seedvr2: one-step video restoration via diffusion adversarial post-training. In International Conference on Learning Representations, Vol. 2026, p. 40987–41009. Cited by: §2.1, §4.1. Wang et al. (2019) X. Wang, K. C. Chan, K. Yu, C. Dong, and C. C. Loy Edvr: video restoration with enhanced deformable convolutional networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), p. 1954–1963. Cited by: §2.1. Wu et al. (2026) A. Wu, Z. Qiu, T. Yao, and T. Mei PS-sr: pseudo-single-step video super-resolution via speculative diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 38218–38227. Cited by: §4.1. Wu et al. (2023a) H. Wu, E. Zhang, L. Liao, C. Chen, J. Hou, A. Wang, W. Sun, Q. Yan, and W. Lin Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), p. 20087–20097. Cited by: §4.1. Wu et al. (2023b) X. Wu, Y. Hao, K. Sun, Y. Chen, F. Zhu, R. Zhao, and H. Li Human preference score v2: a solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341. Cited by: §4.4. Xi et al. (2025) H. Xi, S. Yang, Y. Zhao, C. Xu, M. Li, X. Li, Y. Lin, H. Cai, J. Zhang, D. Li, et al. Sparse videogen: accelerating video diffusion transformers with spatial-temporal sparsity. arXiv preprint arXiv:2502.01776. Cited by: §2.2. Xie et al. (2025a) E. Xie, J. Chen, J. Chen, H. Cai, H. Tang, Y. Lin, Z. Zhang, M. Li, L. Zhu, Y. Lu, et al. SANA: efficient high-resolution text-to-image synthesis with linear diffusion transformers. In The Thirteenth International Conference on Learning Representations, Cited by: §2.2. Xie et al. (2025b) E. Xie, J. Chen, Y. Zhao, J. Yu, L. Zhu, C. Wu, Y. Lin, Z. Zhang, M. Li, J. Chen, et al. Sana 1.5: efficient scaling of training-time and inference-time compute in linear diffusion transformer. arXiv preprint arXiv:2501.18427. Cited by: §2.2. Xu et al. (2025a) H. Xu, Y. Liu, T. Li, R. Shen, B. Jiang, J. Peng, D. Luo, X. Hu, S. Yan, and H. Li Irpo: boosting image restoration via post-training grpo. arXiv preprint arXiv:2512.00814. Cited by: §2.1. Xu et al. (2023) J. Xu, X. Liu, Y. Wu, Y. Tong, Q. Li, M. Ding, J. Tang, and Y. Dong Imagereward: learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems 36, p. 15903–15935. Cited by: §4.4. Xu et al. (2025b) R. Xu, G. Xiao, H. Huang, J. Guo, and S. Han Xattention: block sparse attention with antidiagonal scoring. arXiv preprint arXiv:2503.16428. Cited by: §2.2. Xu et al. (2025c) Y. Xu, T. Park, R. Zhang, Y. Zhou, E. Shechtman, F. Liu, J. Huang, and D. Liu Videogigagan: towards detail-rich video super-resolution. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 2139–2149. Cited by: §1, §2.1. Yang et al. (2026) S. Yang, H. Xi, Y. Zhao, M. Li, J. Zhang, H. Cai, Y. Lin, X. Li, C. Xu, K. Peng, et al. Sparse videogen2: accelerate video generation with sparse attention via semantic-aware permutation. Advances in Neural Information Processing Systems 38, p. 96965–96991. Cited by: §2.2. Yi et al. (2026) Q. Yi, S. Li, R. Wu, L. Sun, Z. Zhang, and L. Zhang GDPO-sr: group direct preference optimization for one-step generative image super-resolution. arXiv preprint arXiv:2603.16769. Cited by: §2.1. Yin et al. (2024a) T. Yin, M. Gharbi, T. Park, R. Zhang, E. Shechtman, F. Durand, and W. T. Freeman Improved distribution matching distillation for fast image synthesis. Advances in neural information processing systems 37, p. 47455–47487. Cited by: §1, §4.4. Yin et al. (2024b) T. Yin, M. Gharbi, R. Zhang, E. Shechtman, F. Durand, W. T. Freeman, and T. Park One-step diffusion with distribution matching distillation. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 6613–6623. Cited by: §1. Yuan et al. (2026) Z. Yuan, R. Yao, C. Zheng, Y. Zhao, D. Dong, and M. Zhang Detached skip-links and R-probe: decoupling feature aggregation from gradient propagation for mllm ocr. In International Conference on Machine Learning, p. 152198–152218. Cited by: §4.3. Zhang et al. (2018) R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE/CVF conference on computer vision and pattern recognition, p. 586–595. Cited by: §4.1. Zhang and Yao (2024) Y. Zhang and A. Yao Realviformer: investigating attention for real-world video super-resolution. In European conference on computer vision, p. 412–428. Cited by: §4.1. Zhao et al. (2026) S. Zhao, X. Zhang, B. Chen, W. Li, Q. Xing, K. Zhang, Y. Wang, J. Li, L. Zhang, J. Zhang, et al. OARS: process-aware online alignment for generative real-world image super-resolution. arXiv preprint arXiv:2603.12811. Cited by: §2.1. Zhao et al. (2025) W. Zhao, J. Zhou, X. Zhu, W. Chen, X. Zhang, Z. Lei, and F. Wang Realisvsr: detail-enhanced diffusion for real-world 4k video super-resolution. arXiv preprint arXiv:2507.19138. Cited by: §4.1. Zheng et al. (2026) K. Zheng, H. Chen, H. Ye, H. Wang, Q. Zhang, K. Jiang, H. Su, S. Ermon, J. Zhu, and M. Liu Diffusionnft: online diffusion reinforcement with forward process. In International Conference on Learning Representations, Vol. 2026, p. 134129–134150. Cited by: §2.3. Zhou et al. (2024) S. Zhou, P. Yang, J. Wang, Y. Luo, and C. C. Loy Upscale-a-video: temporal-consistent diffusion model for real-world video super-resolution. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 2535–2545. Cited by: §1, §1, §2.1, §4.1. Zhu et al. (2026) Y. Zhu, X. Wang, S. Lathuilière, and V. Kalogeiton Diffusion reinforcement learning via centered reward distillation. arXiv preprint arXiv:2603.14128. Cited by: §2.3. Zhuang et al. (2026) J. Zhuang, S. Guo, X. Cai, X. Li, Y. Liu, C. Yuan, and T. Xue Flashvsr: towards real-time diffusion-based streaming video super resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 43482–43493. Cited by: §1, §1, §2.1, §3.1, §4.1. Appendix A Training Details In Stage 1, we train the 14B full-attention teacher for 4 days using LoRA with rank 384, AdamW with (β1,β2)=(0.9,0.999)( _1, _2)=(0.9,0.999), a learning rate of 2×10−52× 10^-5, and a batch size of 16. In Stage 2, we initialize the hybrid-attention student from pretrained FlashVSR stage-2 checkpoint and perform full-parameter training for 2 days using AdamW with (β1,β2)=(0.9,0.999)( _1, _2)=(0.9,0.999), a learning rate of 1×10−51× 10^-5, a batch size of 32, and EMA. In Stage 3, we perform full-parameter one-step distillation for 600 steps using the 14B teacher, AdamW with (β1,β2)=(0,0.999)( _1, _2)=(0,0.999), a learning rate of 1×10−51× 10^-5, a batch size of 16, and EMA. Training uses a filtered coreset together with the text-SFT augmentation. For Ry-DMD, we set β=0.1β=0.1, λRY=0.2 _RY=0.2, J=4J=4, and K=3K=3. Appendix B RTX 4090 Implementation The attention operators used in our H200 equivalent GPU implementation are not directly supported on RTX 4090 GPUs. We therefore implement a lightweight Triton adaptation based on FlashAttention-2 to support the same hybrid-attention computation on Ada GPUs. For 4K inference, the main memory bottleneck comes from the persistent KV cache of the retained global-attention layers. We selectively offload these KV caches to pinned CPU memory and asynchronously transfer them on a dedicated CUDA stream. Before the corresponding layer is executed in the next streaming chunk, the cache is prefetched back to GPU. This substantially reduces persistent GPU memory usage while overlapping most data transfers with computation. Importantly, this adaptation does not introduce spatial tiling: each frame is processed at its full spatial resolution, while only the KV-cache residency is changed. This enables spatially untiled 4K inference on a single 24GB RTX 4090. Appendix C Ry-DMD Hyperparameter Analysis We study the sensitivity of Ry-DMD to four key hyperparameters on SDXL: the reward temperature β, the correction weight λRY _RY, the number of teacher rollouts J, and the rollout depth K. Unless otherwise specified, we use β=0.1β=0.1, λRY=0.2 _RY=0.2, J=4J=4, and K=3K=3. We vary one parameter at a time while keeping the others fixed. As shown in Table 6, Ry-DMD is stable under moderate variations of its hyperparameters, while overly strong reward weighting or correction degrades performance. Table 6: Hyperparameter analysis of Ry-DMD on SDXL. We vary one hyperparameter at a time from the default configuration (β,λRY,J,K)=(0.1,0.2,4,3)(β, _RY,J,K)=(0.1,0.2,4,3). Setting β λRY _RY J K OCR ↑ ImageReward ↑ Default 0.1 0.2 4 3 0.241 1.02 β 0.05 – – – 0.212 0.98 0.2 – – – 0.236 1.01 0.4 – – – 0.203 0.92 λRY _RY – 0.05 – – 0.201 0.98 – 0.1 – – 0.237 1.01 – 0.4 – – 0.198 0.89 J – – 2 – 0.227 1.00 – – 8 – 0.248 1.04 K – – – 2 0.239 1.01 – – – 4 0.239 1.03 Appendix D Extending Ry-DMD to Few-Step Generation We provide additional quantitative results for four-step SD3.5-M with OCR-guided post-training. As shown in Table 7, Ry-DMD improves OCR accuracy from 0.501 to 0.824 without increasing NFE. Meanwhile, HPSv2, ImageReward, and Aesthetic scores all improve, indicating that better text rendering does not come at the expense of general image quality. Qualitative comparisons are provided in Fig. 9. Table 7: Ry-DMD on 4-step SD3.5-M. Ry-DMD substantially improves OCR accuracy while preserving the original 4-step sampling budget and improving general image-quality and preference metrics. Method NFE ↓ OCR ↑ HPSv2 ↑ ImageReward ↑ Aesthetic ↑ SD3.5-M (4-step) 4 0.501 30.3 1.02 5.52 + Ry-DMD 4 0.824 31.2 1.23 5.54 Appendix E Additional Derivations for Ry-DMD This section provides supporting derivations for Ry-DMD introduced in Section 3.3. We first show that the global reward-tilted target arises from a KL-regularized expected-reward objective. We then derive the local reward-tilted endpoint mean and its finite-rollout estimator, followed by the connection between the resulting endpoint shift and the DMD teacher correction. E.1 Derivation of the Reward-Tilted Target For a fixed condition c, consider the KL-regularized expected-reward objective (q)=x0∼q(⋅∣c)[r(x0,c)]−1βDKL(q(⋅∣c)∥pT,0(⋅∣c)),J(q)=E_x_0 q(· c) [r(x_0,c) ]- 1βD_KL (q(· c)\,\|\,p_T,0(· c) ), (7) subject to ∫q(x0∣c)dx0=1 q(x_0 c)\,dx_0=1. Introducing a Lagrange multiplier for the normalization constraint and setting the functional derivative with respect to q to zero gives logq⋆(x0∣c)pT,0(x0∣c)=βr(x0,c)+const. q (x_0 c)p_T,0(x_0 c)=β r(x_0,c)+const. (8) Therefore, p0⋆(x0∣c)=pT,0(x0∣c)exp(βr(x0,c))Z(c),Z(c)=pT,0(⋅∣c)[exp(βr(x0,c))],p_0 (x_0 c)= p_T,0(x_0 c) \! (β r(x_0,c) )Z(c), Z(c)=E_p_T,0(· c) [ \! (β r(x_0,c) ) ], (9) which recovers the reward-tilted target in Eq. 1. The exponential tilt changes the relative density of teacher-supported samples according to their rewards while preserving the support of the teacher distribution. In the limit β→0β→ 0, p0⋆(⋅∣c)→pT,0(⋅∣c)p_0 (· c)→ p_T,0(· c). E.2 Local Reward-Tilted Mean Estimation During DMD training, reward tilting is applied locally to the conditional endpoint distribution induced by stochastic teacher rollouts from the current noisy state. For fixed (xσ,c)(x_σ,c), let qTK(x0∣xσ,c)q_T^K(x_0 x_σ,c) denote the endpoint distribution induced by a stochastic K-step teacher sampler. The normalized form of Eq. 3 is qT,βK(x0∣xσ,c)=qTK(x0∣xσ,c)exp(βr(x0,c))Zβ(xσ,c),q_T,β^K(x_0 x_σ,c)= q_T^K(x_0 x_σ,c) \! (β r(x_0,c) )Z_β(x_σ,c), (10) where Zβ(xσ,c)=x0∼qTK(⋅∣xσ,c)[exp(βr(x0,c))].Z_β(x_σ,c)=E_x_0 q_T^K(· x_σ,c) [ \! (β r(x_0,c) ) ]. (11) The mean endpoint under the reward-tilted distribution is therefore mβ(xσ,c)=x0∼qTK(⋅∣xσ,c)[x0exp(βr(x0,c))]x0∼qTK(⋅∣xσ,c)[exp(βr(x0,c))],m_β(x_σ,c)= E_x_0 q_T^K(· x_σ,c) [x_0 \! (β r(x_0,c) ) ]E_x_0 q_T^K(· x_σ,c) [ \! (β r(x_0,c) ) ], (12) while the original conditional teacher mean is mT(xσ,c)=x0∼qTK(⋅∣xσ,c)[x0].m_T(x_σ,c)=E_x_0 q_T^K(· x_σ,c)[x_0]. (13) We denote the population-level displacement induced by reward tilting as Δm=mβ(xσ,c)−mT(xσ,c). m=m_β(x_σ,c)-m_T(x_σ,c). (14) In practice, we draw J independent teacher rollouts x0(j)∼i.i.d.qTK(⋅∣xσ,c),j=1,…,J,x_0^(j) i.i.d. q_T^K(· x_σ,c), j=1,…,J, (15) with rewards rj=r(x0(j),c)r_j=r(x_0^(j),c). Without reward standardization, the corresponding self-normalized Monte Carlo estimates are m^β=∑j=1Jπjx0(j),πj=exp(βrj)∑k=1Jexp(βrk), m_β= _j=1^J _jx_0^(j), _j= (β r_j) _k=1^J (β r_k), (16) and m^T=1J∑j=1Jx0(j). m_T= 1J _j=1^Jx_0^(j). (17) The finite-rollout estimate of the reward-induced endpoint shift is thus δm=m^β−m^T,δ m= m_β- m_T, (18) corresponding to Eq. 5. The expression above gives the unstandardized estimator of Δm m. In practice, we compute δmδ m using the group-standardized weights below, as in Eq. 5. Reward standardization. In the practical algorithm, rewards are standardized within each rollout group as aj=rj−r¯sr+ε,πj=exp(βaj)∑k=1Jexp(βak).a_j= r_j- rs_r+ , _j= (β a_j) _k=1^J (β a_k). (19) Centering by r¯ r does not affect the normalized exponential weights, since the common multiplicative factor cancels. Dividing by sr+εs_r+ is equivalent to using a rollout-dependent effective temperature βeff=βsr+ε, _eff= βs_r+ , (20) which reduces sensitivity to the numerical scale of different reward functions while preserving their within-group ordering. The standardized weights therefore implement a group-adaptive tilt, while the fixed-β population formulas above describe the unstandardized case. Small-β interpretation. For additional intuition, a first-order expansion of Eq. 12 gives Δm=βCovqTK(⋅∣xσ,c)(x0,r(x0,c))+(β2). m= _q_T^K(· x_σ,c) (x_0,r(x_0,c) )+O(β^2). (21) Thus, in the small-β regime, reward tilting moves the conditional teacher endpoint mean along variations in x0x_0 that covary positively with the reward. E.3 From Endpoint Shift to the DMD Teacher Correction We finally derive how the endpoint displacement estimated above translates into the reward correction in Eq. 6. In this section, sTs_T denotes the parameterized teacher term used in our DMD update. For σ>0σ>0, it is related to its associated clean-endpoint targetμT(xσ,c) _T(x_σ,c) by sT(xσ,c)=μT(xσ,c)−xσ.s_T(x_σ,c)= _T(x_σ,c)-x_σ. (22) Ry-DMD modifies the clean teacher target using the estimated reward-induced endpoint displacement: μTRY(xσ,c)=μT(xσ,c)+λRYδm, _T^RY(x_σ,c)= _T(x_σ,c)+ _RYδ m, (23) where λRY _RY controls the correction strength. Substituting this target into Eq. 22 yields sTRY(xσ,c) s_T^RY(x_σ,c) =μT(xσ,c)+λRYδm−xσ = _T(x_σ,c)+ _RYδ m-x_σ (24) =sT(xσ,c)+λRYδmσ. =s_T(x_σ,c)+ _RY δ mσ. Thus, under the above parameterization, the factor 1/σ1/σ follows directly from converting a clean-endpoint displacement into a displacement of the teacher term. For numerical stability at small noise levels, we use dRY=δmmax(σ,εσ),d_RY= δ m (σ, _σ), (25) which gives the practical update sTRY=sT+λRYdRY,s_T^RY=s_T+ _RYd_RY, (26) equivalent to Eq. 6. Thus, Ry-DMD leaves the remaining DMD optimization procedure unchanged and requires only sampled teacher endpoints and scalar reward evaluations, without differentiating through the reward function. Here, μT _T remains the original teacher target, and δmδ m is used as a rollout-estimated residual correction; the construction does not require μT=mT _T=m_T. Appendix F Synthetic Text-OCR VSR Dataset To study OCR-guided post-training, we construct a synthetic text-focused VSR dataset from the same video corpus used in our VSR pipeline. Synthetic text is rendered onto decoded high-quality frames to form the target video, and the corresponding low-quality input is generated with the standard VSR degradation pipeline. Since the text is procedurally generated, its content and spatial annotations are known exactly. Text rendering. Each sample contains one or two text instances sampled from a multilingual corpus covering English, Chinese, numeric strings, and several additional languages. We randomize font, size, color, outline, opacity, and shadow. Text remains temporally coherent across the video and follows simple motion patterns such as static placement, translation, scale/rotation changes, or partial motion outside the frame. Per-frame locations and visibility are retained as annotations. Filtering and OCR supervision. We discard invalid renderings with unsupported characters, insufficient size, or negligible visible area. For OCR reward computation, only sufficiently visible annotated text regions are evaluated, and a small number of valid frames are sampled from each video for efficiency. All compared post-training methods use the same source videos, rendered targets, degradation process, and filtering rules. Each sample additionally stores lightweight text metadata used for OCR supervision. A schematic example is shown below: ⬇ "source_path": "/videos/example.mp4", "gt_video": "[F, 3, H, W]", "lq_video": "[F, 3, H, W]", "text_metadata": "instances": [ "text": "EXAMPLE 42", "language": "en", "motion_type": "translate", "frames": [ "frame_index": 0, "bbox": [x0, y0, x1, y1], "visible_ratio": 1.0 ] ]