FlashVSR+ paper teaser comparing restored video detail against baselines, with 4K efficiency results — FlashVSR+ restores fine detail and natural textures while improving 4K throughput.

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× 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.

Three-stage training

Method Overview

A 14B full-attention teacher, a 1.3B hybrid-attention student, and reward-guided distillation form the FlashVSR+ pipeline.

Teacher

Train a 14B full-attention multi-step model with flow matching.

Hybrid student

Adapt a 1.3B backbone with local attention and sparse global blocks.

One-step model

Distill the student with distribution matching, pixel loss, and Ry-DMD.

Efficient high-resolution VSR

Hybrid Attention

Pretrained VSR attention is strongly concentrated around nearby space-time blocks. FlashVSR+ computes those interactions locally, while keeping periodic global blocks for distant context.

Attention locality map and local attention mass across DiT blocks before and after hybrid training — Attention analysis. Compact local neighborhoods receive a disproportionate share of attention; retained global blocks become less local after training.

Key observations

Primarily local

Just ~1% of valid key blocks lie in the local neighborhood, yet they receive over 20× their proportional share of attention mass.

Periodically global

After hybrid training, the retained global blocks take on more global context, with attention mass spread more broadly across space and time.

2× DiT speedup on 4K

Up to 2× faster DiT inference, with 45% lower peak GPU memory.

Reward-Yield Distribution Matching Distillation

Ry-DMD

Ry-DMD samples several teacher continuations from a noisy student state, scores their final outputs, and shifts the teacher target toward higher-reward samples. The reward only needs to return a scalar.

Core equations

01 / Reward-tilted target

p_0^\star(x_0\mid c)\propto p_{\mathrm T,0}(x_0\mid c)\exp\!\bigl(\beta r(x_0,c)\bigr)

Higher-reward teacher samples get more weight.

02 / Local target shift

\delta m=\sum_{j=1}^{J}\pi_j x_0^{(j)}-\frac{1}{J}\sum_{j=1}^{J}x_0^{(j)}

Compare reward-weighted and ordinary rollout means.

03 / Teacher correction

s_{\mathrm T}^{\mathrm{RY}}=s_{\mathrm T}+\lambda_{\mathrm{RY}}\frac{\delta m}{\max(\sigma,\varepsilon_\sigma)}

Inject the shift into DMD without differentiating through the reward.

Conceptual Ry-DMD diagram showing a teacher distribution shifted toward higher-reward samples — Ry-DMD concept: reward-weighted teacher rollouts shift the target distribution toward desirable outputs.

Results

Qualitative comparison of FlashVSR+ against representative VSR methods on text, parrot texture, and animated facial detail — Qualitative comparison with representative VSR methods. Zoomed regions highlight text, fine textures, and facial details.

Efficiency comparison

Inference on 89-frame videos. The table reports GPU time and peak memory on an NVIDIA H200 equivalent GPU; the chart reports end-to-end wall-clock time on RTX 4090.

(a) NVIDIA H200 equivalent GPU
Method	Params (B)	GPU time (s) ↓			Peak memory (GB) ↓
Method	Params (B)	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

RTX 4090 end-to-end wall-clock efficiency comparison across 1080p, 2K and 4K resolutions — (b) NVIDIA RTX 4090

† SeedVR2 uses tiled inference at 2K and 4K because full-frame inference runs out of memory. ‡ Ours-Base shares FlashVSR's architecture and inference procedure on the H200 equivalent GPU; FlashVSR's attention operators are unsupported on RTX 4090.

Generalization beyond VSR

Ry-DMD for Image Generation

On 4-step SD3.5-M, Ry-DMD raises OCR accuracy from 0.501 to 0.824 at the same four-step sampling budget. Text becomes more legible while the scene remains intact.

Comparison of baseline and Ry-DMD four-step SD3.5-M generations: candle, magnifying glass, and elevator sign with clearer text after Ry-DMD — Qualitative comparison on 4-step SD3.5-M. Top: baseline. Bottom: Ry-DMD. Prompts are shown in the figure.

Cite our paper

@article{zhuang2025flashvsr,
  title={FlashVSR: Towards Real-Time Diffusion-Based Streaming Video Super-Resolution},
  author={Zhuang, Junhao and Guo, Shi and Cai, Xin and Li, Xiaohui and Liu, Yihao and Yuan, Chun and Xue, Tianfan},
  journal={arXiv preprint arXiv:2510.12747},
  year={2025}
}

FlashVSR+

FlashVSR+

More detail.
More real.

4K restoration.
One RTX 4090.

Your local workstation.

Real World.

AIGC.

Revival.

FlashVSR+

Abstract

Method Overview

Teacher

Hybrid student

One-step model

Hybrid Attention

Ry-DMD

Results

Efficiency comparison

Ry-DMD for Image Generation

Cite our paper

FlashVSR+

FlashVSR+

More detail.More real.

4K restoration.One RTX 4090.

FlashVSR+

Abstract

Method Overview

Teacher

Hybrid student

One-step model

Hybrid Attention

Ry-DMD

Results

Efficiency comparison

Ry-DMD for Image Generation

Cite our paper

More detail.
More real.

4K restoration.
One RTX 4090.