FlashVSR: Towards Real-Time Diffusion-Based Streaming Video Super-Resolution

• 17 FPS at 768×1408 on one A100 delivers near real-time inference, with up to ∼12× speedup over prior diffusion-based VSR methods.FlashVSR also scales reliably to 1440p while producing detail-rich videos.Source: §1 · pp. 2–3, Figure 1 · p. 2
• FlashVSR uses a training-friendly three-stage distillation pipeline to build a one-step streaming VSR model.The stages progress from a full-attention image–video teacher to sparse-causal adaptation and then one-step distillation.Source: §1 · p. 2
• Locality-constrained sparse attention reduces redundant computation and aligns positional encoding ranges between training and inference for high-resolution generalization.It combines coarse top-k region selection with spatial local windows.Source: §1 · p. 2
• A tiny conditional decoder uses LR frames alongside latents to accelerate reconstruction while preserving visual quality.The decoder addresses the causal 3D VAE bottleneck and supports a more compact design.Source: §1 · p. 3
• VSR-120K provides 120k videos and 180k high-quality images for joint image–video training.The videos average more than 350 frames and are filtered through automated quality control.Source: §1 · p. 3

• Real-world VSR datasets address synthetic-degradation limitations but remain small or sensor-biased, motivating larger training resources.The related work describes paired consumer-device data as limited in scale and biased toward particular sensors.Source: §2.1 · p. 3
• Diffusion-based VSR improves restoration through mechanisms including optical-flow propagation, motion-aware objectives, and staged optimization.Representative methods include Upscale-A-Video, MGLD-VSR, and DiffVSR.Source: §2.1 · p. 3
• Streaming research extends diffusion models to long videos through block-wise sequential processing and causal attention.These approaches target sequences lasting minutes or longer, where streaming capability is essential for deployment.Source: §2.2 · p. 3
• Existing diffusion acceleration strategies include feature caching, one-step distillation, and sparse attention, with VSR methods such as DOVE and SeedVR2 achieving competitive results.These approaches reduce intermediate redundancy or compress iterative denoising into a single step.Source: §2.3 · pp. 3–4
• FlashVSR unifies one-step distillation, train-friendly streaming, locality-constrained sparse attention, and a tiny conditional decoder to target practical diffusion-based VSR.The design addresses efficiency, temporal scalability, and high-resolution generalization.Source: §2.3 · p. 4

3.1 OVERVIEW

• FlashVSR constructs VSR-120K and uses a three-stage pipeline: joint video–image training, causal sparse-attention adaptation, and distribution-matching one-step distillation.Images are treated as single-frame videos, while the Stage 1 teacher is adapted into a streaming student.Source: §3 · p. 4, §3.1 · p. 4, §3.2 · pp. 4–5

3.2 THREE-STAGE DISTILLATION PIPELINE

• The Stage 1 teacher uses full attention for joint image–video super-resolution training, with a segment mask separating image and video samples.A lightweight LR projection layer replaces the VAE encoder for LR inputs, and training uses flow matching.Source: §3.2 · pp. 4–5
• Stage 2 adapts the teacher to causal block-sparse attention, while Stage 3 distills it into a one-step model trained in parallel from LR frames and Gaussian noise.The Stage 3 objective combines distribution matching, flow matching, and pixel-space reconstruction losses.Source: §3.2 · p. 5

3.3 LOCALITY-CONSTRAINED SPARSE ATTENTION

• Locality-constrained attention aligns training and inference positional ranges, preventing repetition and blurring when the model processes ultra-high-resolution inputs.Each query is restricted to a spatial neighborhood, with boundary-preserved and boundary-truncated window rules.Source: §3.3 · p. 6, Figure 3 · p. 6

3.4 TINY CONDITIONAL DECODER

• The Tiny Conditional Decoder conditions reconstruction on LR frames and latents, reducing decoding complexity while preserving fine details.It is trained with pixel-level supervision and distillation from the original Wan decoder.Source: §3.4 · p. 6

Quantitative Comparisons

• FlashVSR consistently outperforms competing methods across datasets, especially on perceptual metrics including MUSIQ, CLIPIQA, and DOVER.The evaluation covers synthetic, real-world, and AIGC datasets.Source: §4.3 · p. 7, §4.2 · p. 7, Table 1 · p. 8

Qualitative Comparisons

• FlashVSR produces sharper, more detailed, and more natural reconstructions than baselines on VideoLQ and AIGC30.The visual comparisons highlight clearer hand textures and bookshelf details.Source: §4.3 · p. 7, Figure 5 · p. 9

Efficiency Analysis

• 136× and 114× faster than Upscale-A-Video and STAR, respectively, while remaining 11.8× faster than SeedVR2-3B and using 11.1 GB versus 52.9 GB peak memory.The comparison uses 101-frame videos at 768 × 1408 resolution.Source: §4.3 · p. 7, Table 2 · p. 8

Sparse Attention

• At 13.6% sparsity, sparse attention preserves nearly identical quality while reducing per-8-frame inference time from 1.105s to 0.355s, a 3.1× speedup.The comparison is against a full-attention baseline on REDS at 768 × 1408.Source: §4.4 · p. 8, Table 3 · p. 8

Locality-constrained Attention

• Both Boundary-Truncated and Boundary-Preserved locality-constrained attention variants outperform global attention across all metrics on ultra-high-resolution videos.Boundary-Truncated offers slightly higher perceptual quality, while Boundary-Preserved maintains competitive performance with better fidelity.Source: Table 5 · p. 9, §4.4 · p. 9

Tiny Conditional Decoder

• The tiny conditional decoder achieves nearly identical visual quality to the Wan decoder while reducing 101-frame decoding time from 11.13s to 1.60s, a ∼7× speedup.It also consistently outperforms the unconditional variant across PSNR, SSIM, and LPIPS.Source: Table 4 · p. 8, §4.4 · p. 9

5 CONCLUSION

• FlashVSR combines streaming distillation, locality-constrained sparse attention, and a tiny conditional decoder to deliver near-real-time, high-quality streaming VSR.The conclusion presents these components as the framework’s basis for efficiency and scalability.Source: §5 · p. 10

A.1 DATA SOURCES

• The source collection retains videos above 1080p and applies aesthetic and quality filtering to remove unsuitable defocus, low-contrast, and artifact-heavy samples.LAION-Aesthetic and MUSIQ predictors provide complementary filtering signals.Source: §A.1 · p. 16, §A.2 · p. 16
• Optical-flow-based motion filtering removes static or near-static segments so the dataset retains sufficient temporal diversity for VSR.RAFT estimates flow, and the flow-field L2 norm measures motion strength.Source: §A.3 · p. 16
• 120k high-quality video clips and 180k high-resolution images remain after filtering an initial collection of approximately 600k videos and 220k images.The dataset is intended as a large-scale resource for robust training.Source: §A.3 · p. 16

Causal LR Projection-In Layer

• The Causal LR Projection-In Layer compresses groups of four LR frames and projects their features into the DiT latent space for direct causal guidance.The resulting conditional embedding is added element-wise to patchified latent tokens.Source: §B.1 · pp. 16–17, Figure 6 · p. 17

Implementation Details

• Table 6 evaluates different KV-cache eviction strategies on the REDS dataset.The supplied passage identifies the table’s evaluation scope but does not report the strategy results.Source: Table 6 · p. 17

Causal LR Projection-In Layer

• The layer’s causal cache mechanism maintains consistency across clips during streaming inference with negligible overhead.Its design aligns with Wan’s 4× temporal and 8× spatial compression scheme.Source: §B.1 · pp. 16–17

Implementation Details

• A fixed prompt is used for every scene to avoid the computational overhead of generating scene-specific captions.This implementation choice supplies the same textual condition across streaming cases.Source: §B.2 · p. 17

KV-cache eviction strategies

• Head-wise importance-based KV eviction causes a clear performance drop, while the evaluated importance-based strategy provides no improvement over sliding-window eviction.The degradation is attributed to attention scores that do not reliably transfer between successive latents.Source: §B.3 · pp. 17–18

Stream VSR pipelines

• FlashVSR discards past predicted clips and conditions both training and inference on LR frames and noise latents.This design eliminates the train–inference gap while enabling efficient parallel training.Source: §B.4 · p. 18

Streaming inference

• At inference, FlashVSR consumes the current LR frame and noise in one step, then carries updated cached keys and values to the next frame.The cache stores recent latent context within a sliding window across all layers.Source: §B.4 · p. 19

Temporal consistency

• Early-layer cache states preserve structural and motion cues, whereas later-layer states carry cleaner high-frequency details that stabilize textures across frames.The complementary layer roles support temporal consistency during propagation.Source: §B.4 · p. 19

Additional visualizations

• FlashVSR achieves clearer frames and superior detail restoration than existing baselines on real-world and AIGC video enhancement.These qualitative comparisons are shown in the additional visualization results.Source: §C.1 · p. 19, Figure 9 · p. 20

High-resolution results

• At 1536 × 2688 resolution, Boundary-Preserved and Boundary-Truncated attention produce sharper, more stable frames than global attention.The variants align positional-encoding ranges between training and inference, avoiding repeated textures and blurring.Source: §C.1 · p. 19, Figure 10 · p. 21

Additional visualizations

• The paper provides a demo video illustrating FlashVSR’s super-resolution capability, although file-size compression may reduce its displayed quality.The original demo version is reported to have higher visual quality.Source: §C.1 · p. 19

User study

• FlashVSR with the tiny conditional decoder receives higher user preference than prior one-step VSR approaches and performs comparably to the original WAN decoder.The blind study covers 32 test sets and evaluates overall quality, video fidelity, and video quality.Source: §C.2 · p. 19, Table 7 · p. 22, Figure 11 · p. 22