H3 Studio · Live Preview (TAEH3)

H3's real decode path is slow on purpose: it runs the full video VAE, decodes a handful of temporal frames, and only then does the pack pick the best still. Running that every few steps just to peek at progress is a great way to double your wall-clock time. H3StudioTAEH3Preview (H3 Studio · Live Preview (TAEH3)) is the lightweight escape hatch - a tiny approximate decoder that shows you what's forming while sampling runs, so the full VAE only gets used once, at the end.

What it actually is

It's an optional inline preview for a model, not a renderer and not a replacement for final decoding. You drop it into the model path, flip enabled on, and it attaches a sampler wrapper to a clone of your model that decodes cheap preview frames on CPU as denoising progresses, ships them to the frontend as JPEGs, and leaves the actual generation path untouched. Your final output still goes through the full H3 video VAE exactly as before - the preview is a peek, not the picture.

How it works

The implementation uses ComfyUI's public sampler-wrapper and websocket APIs (no KJNodes required), targeting Kijai's Apache-2.0 taeh3.safetensors decoder checkpoint. When you enable it, the node looks for that file in the vae_approx model roots. Find it and it clones your model, attaches the wrapper, and starts streaming CPU-decoded previews every preview_every_n_steps. Don't find it and here's the nice part: it logs a warning and continues generation anyway - previews are optional, and the pack treats them that way rather than killing your run over a missing convenience file.

The inputs that matter

• enabled - off by default. This is the master switch; the rest of the inputs only matter when it's on.
• tiny_vae - an enum with one real option, taeh3.safetensors. If more checkpoints appear in your vae_approx folder they'll show up here.
• preview_every_n_steps - default 1 (every step). Crank this up if the CPU decode is stealing time from sampling; every 4–5 steps is plenty to watch a composition form.
• max_resolution - default 768, the cap the preview gets downscaled to. Bigger previews cost more CPU; you don't need 1024 to judge a pose.
• jpeg_quality - default 90. Purely cosmetic for the preview stream.

The single output is the patched model, which you feed forward exactly where the original model was going.

Install

Part of the H3 Studio pack - ComfyUI Manager (search ComfyUI-MiniMax-H3-Studio) or:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/thaakeno/ComfyUI-MiniMax-H3-Studio.git
cd ComfyUI-MiniMax-H3-Studio
python -m pip install -r requirements.txt

Restart and hard-refresh. Then grab the optional preview model - it's not part of the core H3 stack and isn't auto-downloaded:

• taeh3.safetensors from Kijai/MiniMax-H3-TAE → ComfyUI/models/vae_approx/

Where people get burned

The most common "problem" is actually by design: nothing appears and you assume it's broken. Check the ComfyUI console for the H3 Studio TAEH3 preview disabled warning - that means the checkpoint isn't in the right folder, and generation quietly continued without previews. Also keep expectations calibrated: TAEH3 is an approximate decoder, so previews can look softer or noisier than the final full-VAE result - don't kill a run because the preview looked rough at step 3. And if the CPU preview is visibly slowing your sampling, that's what preview_every_n_steps is for. It's an alpha pack on the classic frontend; Nodes 2.0 support is still pending.