MiniMax H3 Latent Inspect

Every H3 latent looks like one blob in the node graph, and the blob is lying. It's a NestedTensor wrapping two separate tensors - video and audio - and half the traps in this pack come from forgetting that. MiniMaxH3LatentInspect is the debugger you drop in when something upstream is producing the wrong shape or dtype and you need to see it. It's the smallest node in the pack and possibly the most useful one to keep around.

What it does

Feed it a samples LATENT and it walks the payload, printing a report of what's actually inside:

type: NestedTensor
  [0] shape=(1, 24, 32, 104, 58) dtype=torch.bfloat16
  [1] shape=(1, 32, 2, 178) dtype=torch.bfloat16
other keys: ['noise_mask']

Two outputs: samples, passed straight through untouched (so you can insert it into a chain without changing the data flow), and report, a STRING with those lines. As an output node it also prints the report to the console, which is where it earns its keep - you can read the shapes without hunting through a string output on a busy canvas.

Why you'll actually reach for it

The audio latent is 4-D - same rank as an image latent - and everything in this pack warns you not to resize "whatever is 4-D or 5-D" or you stretch the soundtrack. But you can't act on that advice until you can see which tensor is which. Inspect tells you instantly: the video tensor is (B, C, T, H, W), 5-D, and the audio is (B, 32, 2, L), 4-D. If the numbers look wrong, you've found the bug before it costs you a 40-minute encode.

A few things it makes obvious:

• dtype drift. If a node has quietly cast your latent to fp32, you'll see it here before it bloats a save or changes sampler behaviour.
• Missing noise_mask. If you expected a denoise mask and it isn't in other keys, your partial-denoise wiring is wrong and Inspect will show it before the sampler silently ignores you.
• Unexpected frame counts. A video latent that should hold T = 32 holding T = 60 explains a lot of downstream confusion.

Install

Part of the ten-node pack, so it comes with the rest:

cd ComfyUI/custom_nodes
git clone https://github.com/panghea/ComfyUI-MiniMax-H3-Inpaint-Tools
cd ComfyUI-MiniMax-H3-Inpaint-Tools && pip install -r requirements.txt

Or through ComfyUI Manager, search MiniMax H3 Inpaint Tools, restart, and it's under MiniMax H3/latent. It has no other inputs and no dependencies beyond the pack's minimax-h3-latent-core, so it's about as frictionless a debugging node as this ecosystem offers.