Comfy-Org Ships RelayVSR to Upscale Video 4x Without Running Full Diffusion
Comfy-Org repackaged RelayVSR as int8 ComfyUI-ready safetensors for 2× and 4× video upscaling Uses sparse Wan2.2 5B keyframes every 15 frames plus a lightweight Flash decoder per frame Hits 12.59 FPS at 4096×3072 with the LightX2V-optimized variant on a single GPU Native 2× variant targets
Takeaways
- Comfy-Org repackaged RelayVSR as int8 ComfyUI-ready safetensors for 2× and 4× video upscaling
- Uses sparse Wan2.2 5B keyframes every 15 frames plus a lightweight Flash decoder per frame
- Hits 12.59 FPS at 4096×3072 with the LightX2V-optimized variant on a single GPU
- Native 2× variant targets AI-generated video post-processing without amplifying temporal flicker
- Apache-2.0 licensed with open inference code and a research paper
- Small model is 4×-only; arbitrary scales require the full generative reference model
Comfy-Org has packaged RelayVSR weights for direct use in ComfyUI, including native 2× and 4× video upscaling checkpoints. RelayVSR runs a large generative model on sparse keyframes, then uses a smaller network to reconstruct the remaining frames. The design reduces inference cost while retaining the generated detail and temporal context that lightweight per-frame upscalers often lose.
Both checkpoints use int8 quantization and carry an Apache-2.0 license. They target developers who need practical post-processing for generated video or real-world footage without running a large diffusion model on every frame.
Files, paths, and pairings
| Checkpoint | Scale | Intended use |
|---|---|---|
| relayvsr_x4_int8_convrot.safetensors | 4× | High-magnification restoration and upscaling |
| relayvsr_x2_lightx2v_int8_convrot.safetensors | 2× | Lower-cost post-processing for generated video |
A ComfyUI workflow needs the selected RelayVSR checkpoint and its matching Flash VAE decoder. The model uses no text encoder, which removes one component from the loading and inference path.
- Copy the RelayVSR checkpoint into
models/diffusion_models/. - Copy the matching Flash VAE decoder into
models/vae/. - Load the checkpoint and decoder with the same scale in the RelayVSR workflow.
- Rescan the model directories or restart ComfyUI if the files do not appear.
One heavy pass every 15 frames
The method comes from the RelayVSR paper by Xijun Wang and collaborators. Its Sparse Generative Relay mechanism sends occasional keyframes through a large generative model, which produces reference latents. These latents are compressed internal representations containing texture, structure, and other visual features that guide reconstruction.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
