PedroMarinhoDev/ReCaVSR-ComfyUI

Nodes on GitHub · The original weights · ReCaVSR's code · Paper

ReCaVSR (USTC)'s transformer as an int8 ConvRot file for ComfyUI. ReCaVSR is a one-step streaming diffusion video super-resolution model built on Wan2.2 TI2V-5B: it upscales a clip 8 frames at a time, one denoising step per block, and decodes each block as it goes. With the ComfyUI-ReCaVSR nodes it runs on ComfyUI's own memory management.

_{The input (bicubic 4x) against ReCaVSR at 4x, crops at 1:1. Left: a rainy street, 960×480 to 3840×1920, input on top. Right: a portrait, 360×360 to 1440×1440, input on the left. bf16, fast kernels, FlashDecoder. Clips from ReCaVSR's real-world demo set.}

Files

For ComfyUI/models/recavsr, next to prompt.safetensors and flashdecoder.safetensors from kopper/ReCaVSR:

File	Size
transformer_int8_convrot	6.27 GB	this repository: the transformer 1.8× faster, 47 to 48 dB from bf16, about 4 GB less VRAM
transformer (bf16)	10.6 GB	kopper/ReCaVSR, as released

Use

1. Install ComfyUI-ReCaVSR: clone it into ComfyUI/custom_nodes, pip install -r requirements.txt, and restart ComfyUI 0.44 or newer.

2. Download the files:

   cd ComfyUI/models
   hf download PedroMarinhoDev/ReCaVSR-ComfyUI recavsr/transformer_int8_convrot.safetensors --local-dir .
   hf download kopper/ReCaVSR prompt.safetensors flashdecoder.safetensors --local-dir recavsr
   

3. Open Templates → ComfyUI-ReCaVSR → recavsr_video_4x, pick transformer_int8_convrot.safetensors in ReCaVSR Model Loader, load a video and run.

Measured

The ReCaVSR Upscale Video node at 4x with fast kernels and FlashDecoder, models in RAM and kernels built (ComfyUI 0.44.0, PyTorch 2.14.1, sage attention, dynamic VRAM). PSNR is against bf16 with exact kernels, over every frame of the node's output.

Clip	Card	bf16	int8 ConvRot	int8 against bf16
Rainy street, 149 frames, 960×480 → 3840×1920	RTX 4090	49 s	36 s	48.0 dB
	RTX 3090	111 s	67 s
Portrait, 180 frames, 360×360 → 1440×1440	RTX 4090	17 s	15 s	47.2 dB
	RTX 3090	33 s	23 s

The transformer alone runs 1.75 to 1.8 times as fast on both cards; the whole node gains less, because the FlashDecoder and the frame copies don't change. Peak VRAM at 1440×1440: 11.7 GB against 15.7 GB. The full tables, with the exact kernels and the Wan 2.2 VAE decoder, are in the GitHub README.

What changed from the release

transformer_int8_convrot.safetensors is kopper/ReCaVSR's transformer.safetensors converted by the pack's tools/convert_int8.py. The self-attention, cross-attention query and output, and feed-forward weights of the 30 blocks are rotated by a 256-wide block Hadamard matrix (ConvRot) and rounded to int8 with one scale per output row, in ComfyUI's int8_tensorwise format with each layer's comfy_quant metadata, so ComfyUI's mixed-precision ops load it. The patch embedding, LQ projection, conditioning, cross-attention keys and values (computed once per video), norms and head stay bf16. Mean relative weight error 0.96%. The int8 GEMMs need an NVIDIA GPU of compute capability 7.5 or newer (RTX 20 series up); elsewhere ComfyUI dequantizes the weights, which saves memory but not time.

License

kopper/ReCaVSR states no licence for its weights; the authors' README says the weights are subject to their respective licences. ReCaVSR's code is Apache-2.0, and the transformer is Wan2.2 TI2V-5B (Apache-2.0) with ReCaVSR's adapters merged in. This file is a conversion of those weights and follows their terms; the ComfyUI-ReCaVSR nodes are GPL-3.0. See NOTICE.

@misc{wang2026recavsronestepstreamingdiffusion,
  title={ReCaVSR: One-Step Streaming Diffusion Video Super-Resolution with Recycled Latents and Learned Cache Routing},
  author={Xijun Wang and Xin Li and Suhang Yao and Zirui Lang and Bingchen Li and Zhibo Chen},
  year={2026},
  eprint={2609.37831},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2609.37831}
}