MiniMax H3 Still Decode (EXP/T8)

MiniMax H3 is a video model, but sometimes all you want out of it is one good frame. This node is the cheap way to get it: it decodes only the video half of an H3 AV latent and hands you a single candidate still, no full-clip decode, no audio involved. If you've been running H3 as a still-image generator through one of the experimental Ref2VA "still" workflows, this is your extract-and-pick node.

Why it exists: H3's latent is a joint AV thing - video and audio live in one nested tensor - so the normal AV decode path is doing a lot of work you don't need when your target is a thumbnail, a reference image, or a frame to feed back into the next shot. MiniMaxH3StillDecodeT8 walks straight past the audio and decodes video only, then lets you choose which frame survives.

The inputs that matter

You feed it the AV latent straight off the sampler and the video VAE (the same one your normal decode uses - don't hand it the audio VAE, it wants the video stream). The frame_selection combo does the picking: middle, first, last, or index. middle is the sensible default and what you'll use 90% of the time; reach for index when you know exactly which frame you want, in which case set frame_index (0–3599) to go with it. Nothing else needs touching.

What comes out

Three sockets:

• image - the single chosen frame, ready to wire into anything that eats an IMAGE.
• candidate_frames - the batch of frames it decoded before choosing, so you can eyeball alternatives or feed them to a preview node.
• report_json - the pack's standard machine-readable report, useful when you're scripting this in a batch and want the decode stats.

How to install

This is part of the MiniMax H3 Audio T8 pack, so you install the pack, not the node. In ComfyUI Manager, search "MiniMax H3 Audio T8" and hit install, or run comfy node install minimax-h3-audio-t8. Prefer the terminal:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8

Then restart ComfyUI. The pack has no mandatory pip dependencies and won't download H3 weights for you - bring your own H3 DiT, CLIP and VAE, matching what you generated the latent with.

Where people get burned

The classic mistake is feeding it a latent from a task that didn't actually produce frames - H3's "still" contracts can be configured down to one frame, so candidate_frames may be a batch of one and that's fine, not a bug. And remember the geometry rules that apply to everything in this pack: canvases are multiples of 32, and the still route is capped around 1920×1088. Also worth noting: H3's local weights sit under the MiniMax H3 Community License, which excludes the US, EU, UK and Korea - check you're licensed to run them before you build a workflow around this. Once you are, this node is the closest thing H3 has to a "download this frame" button.