ReCaVSR one-step streaming 4x video super-resolution ported to ComfyUI
LatentSpacer · reddit · 2026-11-10
Developer LatentSpacer ported ReCaVSR — a one-step streaming video super-resolution built on Wan2.2 TI2V-5B by USTC researchers — to native ComfyUI nodes.
- 2×/4× upscaling with one diffusion step per block of 8 frames; a built-in FlashDecoder decodes each block on the fly, so longer clips cost time, not memory.
- Benchmarks: 149 frames at 960×480 → 3840×1920 in 36 s on an RTX 4090 and 67 s on a 3090 (int8, fast kernels).
- int8 ConvRot transformer is 1.8× faster with 4 GB less VRAM while staying 47–48 dB close to bf16; 1440×1440 output fits an 8 GB card via weight streaming.
- Kernels match the reference implementation bit for bit; the Wan 2.2 VAE is available as an alternative decoder — sharper but 25–40× slower to decode.
The author notes it's a diffusion upscaler, so fine detail is generated rather than recovered. Nodes and int8/bf16 weights are open-sourced.
More from Infra
- Semiconductor selloff looked overdone: order books and hyperscaler spend tell a different story — BenBajarin · 2026-11-11
- KubeRay joins CNCF as the standard way to run Ray workloads on Kubernetes — PyTorch · 2026-11-11
- Hugging Face launches Local AI 301 course on quantization — huggingface · 2026-11-11
- PyTorch megakernel runs 816B MoE on one TPU v7 host at 433 tok/s, beats GB200 on cost — PyTorch · 2026-11-11
- Can local AI kill the Akamai thesis? NPUs ship in a third of new devices, but hybrid wins for now — pdamodaran · 2026-11-11
- AWS official guide: build a Bedrock invocation observability data lake in five stages — AWS ML Blog · 2026-11-11