FlashVSR

Use this repo skill when a task involves the official OpenImagingLab FlashVSR implementation, diffsynth FlashVSR pipelines, one-step streaming video super-resolution, Locality-Constrained Sparse Attention (LCSA), the block_sparse_attn backend, v1/v1.1 model bundles, or A100-oriented inference. It is an operating guide for inference, not a training, dataset, or third-party ComfyUI skill.

Route by task

• Install, diagnose, or validate weights/backend: read setup-and-weights. Use its read-only environment and weight checkers before loading checkpoints.
• Prepare inputs or run a restoration job: read inference, then its API and workflow references. Choose full, tiny, or tiny-long deliberately.
• Check whether this skill matches a changed checkout: read repo-provenance.md before refreshing.
• Cross-cutting failures: read troubleshooting.md, then the nearest sub-skill troubleshooting reference.

Public runtime contract

FlashVSR is primarily a 4x video restoration pipeline. The official conditioning tensor is RGB bfloat16 in [-1, 1] with shape [1, 3, F, H, W]. Prepare the source with bicubic 4x scaling, center-crop each spatial dimension to a positive multiple of 128, and use a streaming frame count F = 8n + 1. The expected model output is [3, F-4, H, W]; it is converted to RGB uint8 frames and encoded as an MP4 at the source FPS.

The LCSA backend is not optional for the official masked streaming path. Before expensive inference, require:

python -c "from block_sparse_attn import block_sparse_attn_func; print('LCSA import OK')"

A CUDA-enabled PyTorch import alone is not sufficient. The target profile verified for this skill is Python 3.11, PyTorch 2.6.0+cu124, CUDA 12.4, and NVIDIA A100 SM80. Other GPUs or CPU execution are not equivalent verification.

Version and route rules

• Prefer v1.1 for new runs; keep all DiT, projection, decoder, and VAE files from one version-atomic model directory.
• Use full when the Wan VAE route fits memory; use tiled=True when full VAE decode is the bottleneck.
• Use tiny for ordinary clips with the conditional decoder.
• Use tiny-long when GPU input/output memory is the limiting factor; it keeps prepared LQ slices and decoded chunks on CPU, but total host memory can still grow with duration.
• Start with one step, cfg_scale=1.0, if_buffer=True, kv_ratio=3.0, local_range=11, color_fix=True, and a sparse ratio of 2.0. Recompute topk_ratio = sparse_ratio * 768 * 1280 / (H * W) for every geometry.

Installation boundary

The generated skill is self-contained runtime guidance and safe diagnostics; it never downloads weights or depends on the original checkout at runtime. Install the public diffsynth package and its pinned CUDA requirements according to install.md, build Block-Sparse-Attention separately, and package the version-matched projection, TCDecoder, and positive-context support in the application that executes inference. Do not treat third-party integrations as the official path.

Verification boundary

Static API, geometry, frame-plan, weight-contract, and diagnostic-script checks are complete. Native LCSA kernel, real checkpoint loading, full/tiny/tiny-long GPU inference, OOM recovery, and MP4 reopen checks must remain explicitly reported as native candidates until executed with real weights. Read the final verification report under the configured test artifact directory for the exact status; this production run does not import the skill into the managed router.