Wan Latent Preview (VAE Utils)

This one gives you a cheap, fast look at what's inside a Wan latent without paying for a full VAE decode. Latent in, image out - but the image is an approximation, a rough sketch of the real thing, meant for your eyes while you're debugging rather than for saving. Think of it as a peephole into the middle of a Wan pipeline.

Why you'd reach for it

Wan generations are slow and the VAE decode is not free, so you really don't want to run a full decode just to answer "is my latent even in the right ballpark?" That's the gap this fills. You wire it partway through a graph, get a quick low-fidelity render of the current latent, and decide whether it's worth continuing before you commit the compute to a proper decode.

The reason base ComfyUI doesn't just do this for you is a latent-space problem. ComfyUI's built-in live previews lean on fast approximate decoders that are keyed to specific latent formats - the ones it grew up on. Wan doesn't use those. It's a 3D causal VAE with a 16-channel latent that also compresses time, which is a different animal from the 4-channel SD/SDXL latents the stock preview machinery was tuned around. Point the generic preview at a Wan latent and you get mush or nothing useful. This node knows the Wan latent's shape, so its approximation actually resembles the frame you're generating.

How it works

It's a lightweight approximate decode, not the real one. Instead of running the full VAE - the expensive, accurate path - it does a fast linear-ish mapping from the Wan latent's channels straight to RGB. That's the same trick the "latent2rgb" style previews use everywhere in ComfyUI: cheap, instant, and good enough to recognize composition, color, and gross motion. What it won't give you is fine detail or true colors, because the whole point is that it skips the network that would produce those. If the preview looks a little flat or smeary, that's expected - it's a proxy, and the real decode is the source of truth.

Inputs and outputs

Refreshingly minimal:

• samples - the Wan LATENT you want to eyeball. That's the only input.

The output is an IMAGE, which means you wire it into a Preview Image node (or Save Image if you really want to keep the rough version) and look at it. Because it's a normal image output, you can also chain it into anything that takes an image - but honestly its job ends at "let me see what I've got."

Installing it

The pack is dependency-free - nothing beyond what ComfyUI already ships with. Install through ComfyUI Manager by searching ComfyUI-VAE-Utils, or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/spacepxl/ComfyUI-VAE-Utils

Restart ComfyUI. No model file is needed for this node - the pack's only download belongs to its Latent Upscale node, and this preview node doesn't touch it.

Common issues

The main thing to get straight is expectations: don't judge your output on this. It's approximate by design, so a preview that looks soft or off-color is not evidence your generation is broken - decode it properly before you conclude anything. This is a debugging instrument, not a render.

Second, it's a Wan preview. Hand it a latent from an SDXL or Flux workflow and the channel mapping it uses won't match that latent's format, so the picture will be meaningless. Keep it on the ecosystem it was built for. And if you're pulling latents out of Kijai's WanVideoWrapper, be aware those come normalized differently than native Wan latents - the pack ships a separate Scale/Unscale Latents node precisely for that mismatch, and a preview that comes out looking like noise is a good hint you're feeding it a latent that hasn't been un-normalized yet. spacepxl builds a lot of the low-level Wan plumbing, and this little previewer is the kind of quality-of-life tool that only exists because someone was neck-deep in Wan latents and got tired of waiting on full decodes to check their work.