GetWarpedNoiseFromVideoHunyuan

HunyuanVideo was the model that made local video generation feel real at the end of 2024 - 13 billion parameters, gloriously uncensored, and hungry enough for VRAM to make people sweat. If you're running it in ComfyUI, this node is the quickest way to apply the Go-with-the-Flow trick to it: instead of starting your sampler from random noise, you start from noise that has been warped along the optical flow of a real video, so the model begins with motion structure already baked into the latent. Kijai built the dedicated Hunyuan variant right after the CogVideoX one because the same trick transfers with just a change of latent layout.

Same family, same pipeline as the rest of the pack: a RAFT optical-flow model measures where each pixel moves between consecutive frames of your reference video, and a seeded random-noise field is dragged along that motion frame by frame while keeping its gaussian statistics intact. Only the motion is extracted - the reference frames themselves never reach the sampler.

How it works here

The Hunyuan variant fixes what HunyuanVideo cares about: 16 noise channels, BTCHW latent shape (the tooltip on the generic node says it plainly: HunyuanVideo wants BTCHW), and the 4x temporal compression of its VAE. Give it num_frames and it computes (num_frames - 1) / 4 + 1 latent frames for you - 49 frames becomes 13 latents, same as CogVideoX's arithmetic but in Hunyuan's channel layout.

The inputs that matter

• num_frames (default 49) - target video length; the noise is interpolated to match.
• noise_downtemp_interp - how warped noise is collapsed from video frames to latent frames: nearest (default), blend, blend_norm, randn (throws the warp away), or disabled.
• degradation (default 0.5) - blends warped noise toward fresh random noise. Same hedge Kijai ships on the CogVideoX variant: at 0.5 you keep half the warp structure, and at 1.0 the effect is fully gone.
• seed - seeds the initial noise field. Same video + same seed = same noise.

Two small notes on what this node doesn't have. There's no output_device input, unlike its siblings - latents are always returned on CPU, which is fine for piping into a sampler and saves VRAM on a model that's already heavy. And there's no optical_flows output: only noise (LATENT) and visualization (IMAGE, a normalized preview of the noise field so you can check the warp is tracking your motion). The node skips computing the flow visualization, which makes it marginally faster.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-VideoNoiseWarp

Or search VideoNoiseWarp in ComfyUI Manager. Dependency is just einops; the RAFT optical-flow model downloads itself on first run, so have internet for that one execution. You'll pair the noise output with your HunyuanVideo sampler setup - Kijai's ComfyUI-HunyuanVideoWrapper is the usual companion if you don't already have a working pipeline.

Realistic expectations

The whole pack is marked WORK IN PROGRESS and has been dormant since spring 2025 - an experiment that works, not a maintained library. Honest framing on the model side too: HunyuanVideo's moment as the local video model largely passed to Wan as the ecosystem moved on, per the KB's own history, so you're probably on this node because you already have a Hunyuan pipeline you like, not because it's the cutting edge. If your output looks like ordinary random noise, check degradation - at 1.0 the warp is fully disabled. And the RAFT pass runs at full video resolution over every frame pair, so keep reference clips short and reasonably sized or expect a slow, VRAM-hungry first stage.