H3 TAE Decode (draft preview)

H3 TAE Decode is the draft-preview node: it runs H3's video latents through a 9.3 MB tiny autoencoder instead of the full video VAE, turning a decode that takes about a minute per shot into roughly two seconds. Drafts smear texture - that's the trade - but composition, motion, framing and layout all read clearly. It's built for seed hunts and batch triage, and it is emphatically not for finals.

The name is a deliberate echo of the classic TAE (tiny autoencoder) pattern for video models: a cheap decoder that exists so you can look at a latent before you pay for the real one. If you've ever queued six seeds at 243 frames to check a pose, you know the pain this removes. You check the draft, find the good seed, and only then run the real decode.

How it works

You feed it the sampler's video latents and it decodes them chunk by chunk through the tiny decoder, returning plain image frames. What you get is honest about being draft quality:

• Per-frame 2D - there's no temporal smoothing, so motion will look less fluid than the real VAE's output.
• The TAE's own output resolution - not the render resolution, so expect smaller frames.

The required inputs are samples (the sampler's video_latents output, or any [B,24,T,H,W] H3 latent) and tae_name (a dropdown that defaults to taeh3.safetensors). The single output is frames (IMAGE), which you can wire to a preview node or a video writer for a quick look.

The node's relationship with the file it loads is deliberately strict: if the TAE layout doesn't match what it reconstructs, it raises rather than guessing - the comment says a changed file upstream means "this node needs updating rather than guessing." You won't get silently wrong frames.

Installing it

The node ships with the ComfyUI-H3-Multishot pack:

cd ComfyUI/custom_nodes
git clone https://github.com/jlucasmcrell/ComfyUI-H3-Multishot

Or search H3 Multishot in ComfyUI Manager. Requires ComfyUI v0.30.0+. The one extra download is the TAE file itself - 9.3 MB, from Kijai's MiniMax-H3-TAE repo - and it must land in ComfyUI/models/vae_approx/ as taeh3.safetensors. Miss that and the node tells you exactly where to put it.

Gotchas

The whole node is a trade, so the gotcha is the trade. A TAE decode is not a preview of final quality - texture is smeared and there's no temporal smoothing, so a frame that looks muddy in draft can be fine in the real render, and a composition that looks fine in draft can fall apart once the full VAE's detail comes in. Use it to pick seeds and check blocking, not to judge skin or hair. And remember the file: installing the pack gets you the node, but not the 9.3 MB decoder it loads, so the first run on a fresh machine throws until you drop taeh3.safetensors in vae_approx.