Most AI video pipelines end with an upscaling pass, and most of those passes waste work. RelayVSR, the keyframe-based video super-resolution model from the LightX2V team, originally shipped as a fixed 4x upscaler — so a 768x1344 clip that only needed to double to a normal 1536x2688 finishing resolution got pushed all the way to 3072x5376 and then resized back down, paying four times the compute for pixels nobody asked for. On October 29, LightX2V published RelayVSR X2, a version retrained from the ground up for native 2x super-resolution, and as of November 5, Comfy-Org has repackaged its weights so ComfyUI loads them directly — no detour through LightX2V’s own inference scripts.
Why upscaling AI video is harder than it sounds
Generated video already carries mild temporal flicker even before anyone touches it. Push that clip through a naive super-resolution model and the flicker usually gets worse, not better — sharper individual frames can still read as a noisier, less stable clip in motion. RelayVSR’s architecture is built around that specific failure mode: a large generative model writes reference latents for sparse keyframes only, and a much smaller reconstruction network carries those references across every frame in between, using a bounded lookahead and a frame-ring KV cache to keep the output temporally consistent rather than just frame-sharp.
Two files, one scale, no guessing
The X2 release ships as exactly two checkpoints that have to live in the same directory: a 10.1 GB Wan-based model that generates the sparse keyframe references, and a 127 MB compact decoder that super-resolves everything else. The LightX2V runner validates architecture and scale together and refuses to load a mismatched 2x/4x pair rather than silently producing a broken result — a small but telling design choice for a tool meant to sit at the end of a production pipeline rather than in a research notebook.
Comfy-Org’s HuggingFace repackaging (based on the original kopper/RelayVSR weights) drops the files straight into ComfyUI’s standard diffusion_models and vae folders, which is the difference between “a model researchers can run” and “a node artists actually reach for.” The benchmark LightX2V published to demonstrate it is itself MiniMax H3 output — a 1024x768 drone rooftop clip at 24fps with the kind of pre-existing flicker generated footage tends to carry — compared side by side against plain bicubic upscaling, FlashVSR+ Turbo, and Open-H3-Regenerate-2K.
The part that matters for a working pipeline
The older 4x checkpoint still technically accepts a 2x request, but it computes at full 4x internally and downsizes afterward — exactly the wasted work X2 exists to avoid. For anyone finishing MiniMax H3 generations (or any similarly-scaled video model) at a normal delivery resolution, that’s the practical upgrade: the same visual target for a quarter of the pixel math, loaded as a native ComfyUI node rather than a separate script to babysit.