LatentWalkNoise

The obscure one. LatentWalkNoise doesn't touch prompts and doesn't touch image latents - it walks the noise that the sampler starts from. Same prompt, same conditioning, same seed family, but the diffusion path starts at a different point and gradually drifts from your start noise to your end noise. If LatentWalkConditional is the beginner-friendly version of prompt travel, this is the one for people who want the sampling journey to vary instead.

Why it needs a second pack

Here's the honest catch, straight from the README: samplers don't support batches or lists of noise vectors. So the node can't just hand you a batch and let KSampler chew through it. Instead you need BadCafeCode's execution-inversion-demo installed, which lets you build for-loops inside ComfyUI. That's a real extra dependency - bigger than the pack's own requirements - and it's the main reason most people never bother with this node. The conditional walk gets you 90% of the visible animation effect with a tenth of the plumbing.

How it works

You feed it two start/end noise objects - the NOISE type that comes out of the advanced sampler's noise nodes, like Noise_RandomNoise. It interpolates between them along the same travel modes as the rest of the pack (linear, circle, quadratic, cubic, quartic, geometric, hinge, norm), with factor only mattering for hinge and norm and reflect flipping the walk into a ping-pong A→B→A.

Two technical details in the source are worth knowing. First, there's no blend input - the only blending mode is add, hardcoded, because the other blend math doesn't work on noise. Second, the noise vectors get rescaled to a specific linalg norm after blending. That's not pedantry: diffusion noise needs a particular magnitude to behave, and the README notes that without the rescaling "the overall noise becomes too small - sampling doesn't work as expected." So the node quietly fixes a problem you'd otherwise hit in the first frame.

The outputs are where the future-proofing shows. NOISE_BATCH is a batch of noise vectors for "in case samplers support it in the future" - as of now, nothing consumes it. ACCUMULATION is what you actually wire into the execution-inversion loop, feeding the accumulated noise through each iteration of the for-loop that drives your animation.

Who should actually use it

Honestly? People doing serious prompt-travel work who've hit the limits of conditioning-space interpolation and want the sampling path itself to be part of the animation - think frame-to-frame coherence where the noise evolves smoothly rather than jumping. For everyone else, LatentWalkConditional gives you a nearly identical visible effect and doesn't demand you install an experimental loop framework. This node is the pack's most capable and most fiddly piece.

Installing it

The pack itself is trivial - it's the Latent Walk suite (rnbwdsh/ComfyUI-LatentWalk), a fork of Nicholas McCarthy's TravelSuite:

cd ComfyUI/custom_nodes
git clone https://github.com/rnbwdsh/ComfyUI-LatentWalk

Restart ComfyUI, then separately install execution-inversion-demo (via ComfyUI Manager or git clone https://github.com/BadCafeCode/execution-inversion-demo). Dependencies are torch, numpy, Pillow and scipy - all standard, no model files. The README's example animations run on flux.1-schnell GGUF, so it's current-model compatible. Just budget for the fact that this one needs friends.