Research map / Video / Video segmentation and tracking

SkipSR: Faster Super-Resolution with Token Skipping

NeurIPS 2026 · Video › Video segmentation and tracking · approach: Real-world video super-resolution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SkipSR, which uses a lightweight latent-space mask predictor to skip the diffusion transformer entirely on visually simple low-resolution patches and only refines complex patches, accelerating SeedVR/SeedVR2 video super-resolution by about 1.4-1.6x with comparable perceptual quality.

Problem

Diffusion-based video super-resolution is slow because the diffusion transformer processes every spatiotemporal patch uniformly, even though many video regions are low-detail and gain little from refinement.

Approach

Identify skippable low-detail patches from the low-resolution input with a learned predictor, remove them from the transformer sequence entirely (not just from attention), keep original RoPE positions for the remaining patches, and compose skipped patches with the refined output afterward.

Key results

• On VideoLQ, SkipSR-3B reaches 4.47 s/step (1.4x speedup) and skips 27.7% of tokens, with NIQE 4.112 and MUSIQ 54.175 close to SeedVR-3B (4.069, 57.41).
• On AIGC30, SkipSR-3B skips 39.6% of tokens for a 1.6x per-step speedup (3.85 s/step vs 6.41) and reports MUSIQ 65.14, above SeedVR-3B's 64.4.
• The paper claims up to 60% faster end-to-end latency on 720p videos, and reports a one-step SkipSR variant matching SeedVR2 quality in a user study with three experts.

Surprising findings

• Skipping only attention computation yielded about 23% speed loss relative to full skipping, so skipping the whole transformer is needed for the largest speedups.
• Synthetic heavily degraded benchmarks (e.g., YouHQ corrupted with white noise) yield almost no skippable patches, limiting acceleration.

Limitations

• No acceleration on videos with severe degradations where skippable patches are rare.
• Limited speedup on crowded scenes or camera jitter.
• User study uses only three experts and 25 samples.

Experimental setup

Datasets: VBench AIGC-30 VideoLQ YouHQ-40 SPMCS UDM10 REDS30

Benchmarks: VideoLQ AIGC30 SPMCS UDM10 REDS30 YouHQ40

Metrics: NIQE MUSIQ CLIP-IQA DOVER BRISQUE TFLOPs s/step PSNR SSIM LPIPS

Baselines: SeedVR-3B SeedVR-7B SeedVR2-3B SeedVR2-7B RealViformer MGLD-VSR STAR DOVE

Builds on: SeedVR SeedVR2 APT RoPE MMDiT

Closest papers by meaning

• SkipDiff: Adaptive Skip Diffusion Model for High-Fidelity Perceptual Image Super-resolution (AAAI 2024 · similarity 0.91)
• DOVE: Efficient One-Step Diffusion Model for Real-World Video Super-Resolution (NeurIPS 2025 · similarity 0.90)
• Taming Diffusion Prior for Image Super-Resolution with Domain Shift SDEs (NeurIPS 2024 · similarity 0.88)
• One-Step Diffusion for Detail-Rich and Temporally Consistent Video Super-Resolution (NeurIPS 2025 · similarity 0.88)
• Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling (CVPR 2025 · similarity 0.88)
• SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.88)