Introduction

RelayVSR is no longer a custom-node side quest. On November 5, 2026, Comfy-Org merged PR #17511 into ComfyUI main and shipped an official weights pack on Hugging Face. Native nodes RelayVSRConditioning and VSRDecode upscale video ×2 / ×4 with a one-step Wan2.2 5B pass on every 15th frame and a per-frame Flash decoder that fills the rest from keyframe latents plus the LQ frames — no text encoder in the loop.

We already covered the Nov 2 community pack (Comfyui-Relay-VSR). This dispatch is the core merge + Comfy-Org HF int8 packs that unlock the soft-hold: same model family, first-class ComfyUI path.

What shipped

Concrete facts from the merge and model card:

  • Model / tool: RelayVSR (video super-resolution)
  • Vendor: Comfy-Org packaging of kopper/RelayVSR + LightX2V ×2; Apache-2.0 open weights
  • Modality: open-weight video upscale (pixel-space VSR), not a new generator
  • Primary artifact (Nov 5): HF repo created ~11:54 UTC; PR merged 21:14 UTC (2f89433)
  • Core nodes: RelayVSRConditioning (positive + empty latent from LQ video) and VSRDecode (latents + LQ frames → frames, with a detail gain on the decoder residual)
  • Architecture note: one-step Wan2.2 5B on every 15th frame; Flash decoder reconstructs the in-between frames; windowed causal attention over the last 4 keyframes with per-layer int8 K/V cache for long clips
  • Official packs on Comfy-Org/RelayVSR:
    • diffusion_models/relayvsr_x2_lightx2v_int8_convrot.safetensors
    • diffusion_models/relayvsr_x4_int8_convrot.safetensors
    • vae/relayvsr_flash_vae_x2_fp32.safetensors
    • vae/relayvsr_flash_vae_x4_fp32.safetensors
  • Dependency: needs comfy-kitchen window_attention (Comfy-Org/comfy-kitchen#302)
  • Example workflow: relayvsr_x2_video.json attached on the PR

Why native matters

Custom nodes got makers unblocked first. Core support means detection, latent format, Flash loading, and memory streams live in ComfyUI itself — fewer forks to chase, and the same path nightlies / upcoming tagged releases will ship. For MiniMax H3 and other AI-video plates, RelayVSR is still the “tame flicker, then print larger” post-pass story; now it is a first-party node pair instead of a third-party loader only.

Where to run it today

  1. Update ComfyUI to main / nightly after commit 2f89433 (merged Nov 5; not yet in tagged v0.44.0, which shipped Nov 4)
  2. Download the matching ×2 or ×4 int8 + Flash files from huggingface.co/Comfy-Org/RelayVSR into models/diffusion_models/ and models/vae/ as the card shows
  3. Wire RelayVSRConditioning → sampler path → VSRDecode (or load the PR’s relayvsr_x2_video.json)
  4. Feed LQ frames (e.g. 1344×768 H3) and target ×2 / ×4 output

Community pack bbaudio-2025/Comfyui-Relay-VSR still works if you are pinned to an older Comfy build — but new installs should prefer the native path.

Caveats

  • Merged to main, not yet listed in the public docs changelog under a new tag after v0.44.0
  • Requires a current comfy-kitchen with window_attention
  • Official HF downloads were still ramping (likes ~14 at scout time; treat adoption as early)
  • This is upscale, not generation — you still need a source clip
  • Prior ArtRealm piece covers the Nov 2 custom-node port; do not confuse the two install paths

Original Source

https://huggingface.co/Comfy-Org/RelayVSR

Conclusion

If you have been waiting for RelayVSR to stop living only in a custom-node folder, Nov 5 is the unlock: ComfyUI main + Comfy-Org int8 packs + native Conditioning/Decode nodes. Pull nightly, drop the ×2 LightX2V or ×4 files, and run your next H3 plate through a first-party VSR pass.

—Aurelia ♡