RelayVSR 2x AI video upscaling in ComfyUI, built from the official template — native nodes since ComfyUI 0.45.0.

Settings: RelayVSR x2 model (relayvsr_x2_lightx2v_int8_convrot.safetensors, 4.81 GB) with the RelayVSR Flash VAE (relayvsr_flash_vae_x2_fp32.safetensors); 1 step, CFG 1, euler sampler, simple scheduler. The model paints reference latents for sparse keyframes (one every 15 frames) and the Flash decoder carries them across every frame, so long clips are restored in a streaming pass. Frame rate and audio pass through — only the picture is scaled. For 4x, swap in relayvsr_x4_int8_convrot with the x4 Flash VAE — nothing else changes.

Needs a real GPU — if you run out of memory, trim the input and split it into several passes. Free official template: import the JSON in ComfyUI and run.

The full RelayVSR pack — 2x and 4x presets, 12 settings presets and guides — is on Whop: https://whop.com/videolab-073d/relayvsr-video-upscale-2x-4x-ai-upscaling-prores-comfyui/