# [REPOST + UPDATE] ReCaVSR - A one-step streaming video super-resolution built on Wan2.2 - ComfyUI custom nodes — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/6db15d4827aee346

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-11-11 11:31 ingested 2026-11-12 02:40

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xtfgdc/repost_update_recavsr_a_onestep_streaming_video/)

## retained text



```
Reddit filters deleted my previous post, probably because of a link I had posted, so I'm reposting it without that link.

https://preview.redd.it/8eork6ufsy0i1.png?width=1146&format=png&auto=webp&s=e9f2944a6f1898aaf3f83fd433a4d27c04e5f2dd

I ported ReCaVSR, a one-step streaming video super-resolution, built on Wan2.2 TI2V-5B to native ComfyUI nodes.

\- 4× or 2×, one diffusion step per block of 8 frames. Its own FlashDecoder decodes each block as it goes, so a longer clip costs time, not memory.

\- 149 frames, 960×480 → 3840×1920: 36 s on an RTX 4090, 67 s on a 3090 (int8 file, fast kernels)

\- int8 ConvRot transformer: about 1.8× faster, 4 GB less VRAM, 47–48 dB from bf16

\- 1440×1440 output runs on an 8 GB card; ComfyUI streams the weights that don't fit

\- Exact kernels match the reference implementation bit for bit. The Wan 2.2 VAE is there as an alternative decoder: sharper, but 25–40× slower to decode.

It's a diffusion upscaler, so fine detail is generated rather than recovered. Tested on Linux with an RTX 3090 and 4090, ComfyUI 0.44.  


**Nodes:** [https://github.com/PedroMarinhoDev/ComfyUI-ReCaVSR](https://github.com/PedroMarinhoDev/ComfyUI-ReCaVSR)  
**ComfyUI-Manager:** search "ReCaVSR video super-resolution", or run comfy node install comfyui-recavsr  


**int8 weights:** [https://huggingface.co/PedroMarinhoDev/ReCaVSR-ComfyUI](https://huggingface.co/PedroMarinhoDev/ReCaVSR-ComfyUI)  
**bf16 weights:** [https://huggingface.co/kopper/ReCaVSR](https://huggingface.co/kopper/ReCaVSR)

All credit for the model to Xijun Wang et al. at USTC. Feedback welcome.  


**UPDATE:** ReCaVSR can now upscale videos of any length, at output sizes up to about 5.5K on a 24 GB card (if you have the patience).

This full 3:58 clip (7,153 frames) was upscaled 2×, 1920×1080 → 3840×2160 (4K), in about 25 minutes on a 4090 (4.65 frames/s).

https://reddit.com/link/1xtfgdc/video/bkw8n3jtxy0i1/player

The node now streams frames in and out, so a longer clip doesn’t use more RAM or VRAM. When the output is too big for the card, it chunks the input projection and offloads the KV cache automatically. It also encodes on NVENC, which otherwise sits idle while the model runs.  


**More examples:**  
  
1344×768 → 5376×3072 (4×)

https://reddit.com/link/1xtfgdc/video/5rg5h783xy0i1/player

1344×768 → 5376×3072 (4×)

https://reddit.com/link/1xtfgdc/video/koobfm74xy0i1/player

488×830 → 976×1660 (2×)

https://reddit.com/link/1xtfgdc/video/8ddkzot5xy0i1/player

360×360 → 1440×1440 (4×)

https://reddit.com/link/1xtfgdc/video/rkqm3yu6xy0i1/player


```



document fetched 2026-11-12 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xtfgdc/repost_update_recavsr_a_onestep_streaming_video
