An INT8-ConvRot version of the transformer of kopper/ReCaVSR, a one-step streaming video super-resolution model built on Wan2.2-TI2V-5B and trained for 4×. ReCaVSR is by Xijun Wang, Xin Li, Suhang Yao, Zirui Lang, Bingchen Li and Zhibo Chen (University of Science and Technology of China): paper, code. Both ReCaVSR and Wan2.2-TI2V-5B are under the Apache-2.0 license, and so is this file.
This repository also holds ComfyUI-ReCaVSR, the ComfyUI nodes that run it, and ReCaVSR's other two files, unchanged, so everything installs from here. The same nodes are part of ComfyUI-Lumen (its Upscale ×4).
In the ComfyUI folder:
hf download paralaif/ReCaVSR-comfy-int8-convrot --include "ComfyUI-ReCaVSR/*" --local-dir .
mv ComfyUI-ReCaVSR custom_nodes/
hf download paralaif/ReCaVSR-comfy-int8-convrot recavsr_transformer_int8_convrot.safetensors prompt.safetensors --local-dir models/diffusion_models/recavsr
hf download paralaif/ReCaVSR-comfy-int8-convrot flashdecoder.safetensors --local-dir models/vae/recavsr
Restart ComfyUI and open Templates › ComfyUI-ReCaVSR › recavsr_video_4x: Load Video → ReCaVSR Upscale Video → Save Video, with ReCaVSR Loader feeding the model. The video's audio and frame rate are kept. See ComfyUI-ReCaVSR/README.md for the nodes' inputs, memory use and the converter.
| File | Format | Size | Content |
|---|---|---|---|
recavsr_transformer_int8_convrot.safetensors |
INT8-ConvRot | 5.84 GiB | The transformer, 240 linear layers quantized |
prompt.safetensors |
BF16 | 4 MB | ReCaVSR's fixed prompt embedding, unchanged from kopper/ReCaVSR |
flashdecoder.safetensors |
FP32 | 218 MiB | ReCaVSR's LR-conditioned FlashDecoder, unchanged from kopper/ReCaVSR |
ComfyUI-ReCaVSR/ |
The ComfyUI nodes (MIT; the parts ported from ReCaVSR keep Apache-2.0, see its NOTICE) |
Changes from transformer.safetensors (BF16, 9.88 GiB): in each of the 30 blocks, the self-attention q/k/v/out, the cross-attention q/out and the two feed-forward weights are stored in INT8 with a scale per output row, after a 256-wide block Hadamard rotation (ComfyUI's int8_tensorwise format with convrot); activations are quantized at runtime. Everything else is unchanged BF16: the cross-attention keys and values (computed once per video), the LQ and recycled-latent projections, the patch embedding, conditioning, norms and head.
It was made from the BF16 file with Comfy Kitchen's INT8 ConvRot weight quantizer, using ComfyUI-ReCaVSR/tools/convert_int8.py.
recavsr_transformer_int8_convrot.safetensors 6.3 GBflashdecoder.safetensors 229 MBprompt.safetensors 4 MBkopper/ReCaVSR