VOIDWarpedNoiseSource

VOID is Netflix's open video-inpainting model - the one the community got excited about in early 2026 for removing objects and cleaning up shots without re-rendering the whole frame. It works in two passes. Pass 1 does the actual inpainting. Pass 2 re-samples the result to fix temporal consistency, and to do that it starts from noise that's been warped along the optical flow of your video - so the noise itself moves with the footage instead of flickering like static. That warped noise is produced by the VOIDWarpedNoise node, which returns it as a LATENT.

The problem: SamplerCustomAdvanced doesn't accept a LATENT in its noise slot. It wants a NOISE object. This node is the five-line adapter that bridges the type mismatch - it takes the warped-noise latent and wraps it into a NOISE source you can plug straight into the sampler.

How it works

The mechanism is embarrassingly simple, which is the point of a plumbing node: it converts the latent into a Noise_FromLatent wrapper and hands it back. The real substance lives in the surrounding VOID graph, not here. Crucially, this node is designed to pair with the VOIDSampler node - a DDIM sampler that skips the standard KSampler's noise scaling. VOID was trained with a diffusers scheduler that expects noise at a completely different scale (the standard KSampler multiplies noise by roughly 4500×, which is incompatible with VOID). Use this NOISE with a normal KSampler and you'll get a wrong, blown-out result; the VOID graph routes around that deliberately.

Inputs and outputs

• warped_noise (LATENT) - straight from VOIDWarpedNoise's output. The tooltip is explicit: "Warped noise latent from VOIDWarpedNoise." Don't improvise with a random latent here.

Output: one NOISE object, into SamplerCustomAdvanced's noise input, alongside the VOIDSampler on the sampler slot.

How you get it

Ships with ComfyUI core - no install, no Manager. VOID support landed in ComfyUI in May 2026 (two months after Netflix released the weights, which the community immediately started wiring up with third-party nodes before core support arrived). If your ComfyUI is current, the whole VOID node family is there.

The model files

The VOID checkpoint - a CogVideoX-Fun-1.5-based inpainting model - goes in your checkpoints/diffusion_models folders like any model. The optical-flow part is the sneaky one: VOIDWarpedNoise uses RAFT-large (raft_large.pth), and ComfyUI never downloads it. You must place it in models/optical_flow/ yourself. Skip that and the warped-noise chain won't run, and the error doesn't always say why.

Where people get burned

Wrong sampler pairing is the big one - this NOISE only makes sense with VOIDSampler + SamplerCustomAdvanced. The other trap is frame counts: VOIDWarpedNoise rounds your requested length down so the latent's temporal dimension stays even (a patch_size_t=2 requirement), because an odd count lets a phantom wrapped frame corrupt the end of your clip - visible as jitter or subjects "disappearing" near the end. So if your output is a couple of frames shorter than you asked, that's intentional and it beats the alternative. If you notice end-of-clip jitter anyway, check that you're feeding this node the warped noise and not a fresh random one.