Skip to main content
RelayVSR upscales real footage in one pass. A one-step Wan2.2 5B model writes reference latents every 15th frame, and a small Flash decoder fills the frames between them from those latents and the original low-quality clip. ComfyUI support landed in ComfyUI #17511. Frame rate and audio pass through untouched, so only the picture is rescaled. The scale factor comes from the checkpoint rather than a widget: a 2x model only works with the 2x VAE, and mixing pairs fails.
Make sure your ComfyUI is updated.Workflows in this guide can be found in the Workflow Templates. If you can’t find them in the template, your ComfyUI may be outdated.If nodes are missing when loading a workflow, possible reasons:
  1. You are not using the latest ComfyUI version (Nightly version)
  2. Some nodes failed to import at startup

​
Models

Download the pair for the scale you want and place each file in the folder shown on its card.

Diffusion Model (2x)

relayvsr_x2_lightx2v_int8_convrot.safetensors → ComfyUI/models/diffusion_models/

VAE (2x)

relayvsr_flash_vae_x2_fp32.safetensors → ComfyUI/models/vae/

Diffusion Model (4x)

relayvsr_x4_int8_convrot.safetensors → ComfyUI/models/diffusion_models/

VAE (4x)

relayvsr_flash_vae_x4_fp32.safetensors → ComfyUI/models/vae/
The 2x pair was retrained by LightX2V for the step after video generation, so a 768 x 1344 clip from MiniMax-H3 comes out at 1536 x 2688. The 4x pair is the original RelayVSR scale.

​
Video Upscale: RelayVSR 2x

Upscale a clip 2x with the workflow below. The upscaler runs inside a subgraph, so the canvas only shows the video going in and coming out. RelayVSR 2x video upscale workflow preview

Download Workflow

Download JSON or search “RelayVSR 2x” in Template Library
Input materials

utility_relayvsr_video_x2_upscale_input.mp4

LoadVideo · utility_relayvsr_video_x2_upscale_input.mp4
  1. Load your clip in the Load Video node. The template trims it to the first 5 seconds with a Trim Video node, so raise that duration for longer clips or upscale long footage in passes.
  2. Run the workflow. The result is saved by Save Video under ComfyUI/output/video/.
For 4x, swap the two loader values in the subgraph: UNETLoader to relayvsr_x4_int8_convrot.safetensors and VAELoader to relayvsr_flash_vae_x4_fp32.safetensors. Nothing else changes. If a pass runs out of memory, trim the clip and upscale it in several passes.