MiniMax H3 AV Latent Info / 音视频潜空间信息

MiniMax H3 hands you a latent that is secretly two tensors: a video stream ([B, 24, latent_t, H/16, W/16]) and an audio stream ([B, 32, 2, audio_t]), packed into one NestedTensor. Most nodes treat it as an opaque blob. When you're handed a workflow you didn't build - or a latent you saved from an earlier run - you have no way to see what's actually inside without just sampling it and hoping.

Latent Info is the inspection tool for exactly that situation. One latent input, three outputs: report, frames, and video_tokens.

What the report tells you

The report output is a human-readable block you can drop into any text preview node (ShowText, PreviewText, whatever you already use):

• both tensor shapes, dtype, and device
• the pixel resolution implied by the latent grid
• the frame count and duration at 24 fps
• the video token count - the number that actually determines render time
• batch size

And it flags problems rather than silently passing them along. If latent_t is off H3's 17k+5 grid it reports frames UNKNOWN instead of guessing. If the audio length doesn't match what the frame count implies, it prints a MISMATCH line. If the video channels aren't the expected 24, you get a WARNING. Those three cases are exactly the "this latent came from somewhere strange" situations that would otherwise blow up mid-sampling.

The machine-readable outputs are there for building logic: frames (0 if the grid is invalid) and video_tokens, so you could route a latent to different downstream paths based on its size.

When you'd actually use it

This is a debug-first node, and that's fine. The pack positions it as the thing you reach for when inspecting unfamiliar workflows before sampling - and that's the honest use case. If someone shares an H3 graph and you're not sure whether the latent feeding your sampler is a normal run or something that went through a half-finished edit, run it through Latent Info first. It also pairs naturally with MiniMax H3 Split AV Latent when you want to see the streams individually.

It doesn't modify anything. No inputs beyond the latent, no settings to fiddle with - pure reportage.

Install

Part of the MiniMax-H3 Tools pack, which needs ComfyUI 0.30.0+. Install via ComfyUI Manager by searching "MiniMax H3 Tools," or:

cd ComfyUI/custom_nodes
git clone https://github.com/Rinne414/ComfyUI-MiniMaxH3-Tools.git

Restart ComfyUI, find MiniMax H3 AV Latent Info under MiniMax H3. No extra Python dependencies, no model files. Keep in mind the pack-wide caveat: stock SaveLatent/LoadLatent can't round-trip an H3 AV latent, so if you're inspecting one, do it in the same graph that produced it.