H3 Infinite - Latent Info

The entire Herrgotts H3 Infinite Continuation Suite runs on one deceptively fragile assumption: that the thing flowing through your graph is a complete audio+video H3 latent, not just a video latent, not a half-decoded tensor, not a stale save from an old chain. H3ContinuousLatentInfo is the node you drop into the middle of a workflow the moment that assumption is in doubt. It's not glamorous - it takes one latent and prints what it is.

That single input, one output info (a human-readable STRING), plus three ints: frame_count (pixel frames at 24 fps), video_steps (the raw video latent temporal dimension), and audio_steps (the audio latent tick count). Because H3's video and audio latents have different temporal granularity - the default 39-frame context corresponds to 12 video-latent steps but 65 audio-latent ticks at the 40 Hz audio rate - these numbers are not interchangeable, and this node is how you learn to read them.

When you actually reach for it

• Before a continuation. You're about to feed previous_latent into a Continue node. A quick Latent Info check confirms the frame count matches what the analyzer's handover metadata expects - if they disagree, the Continue node will refuse to run (frame-count mismatch) and now you know why.
• After a save/load round-trip. Saved AV latents carry metadata, and a load from the wrong clip index can give you a latent that looks right. Checking frame_count against the workflow's expectations catches it in seconds.
• Diagnosing a bad seam. If the stitcher is trimming the wrong number of frames, video_steps vs frame_count tells you whether you're thinking in latent steps or pixel frames - they differ by roughly a factor of 12 for H3, and plenty of "the seam is doubled" bugs are really "I trimmed in the wrong units."

Install

cd ComfyUI/custom_nodes
git clone https://github.com/HerrgottMargott/Herrgotts-H3-Infinite-Continuation-Suite.git

Restart ComfyUI, reload the UI. No pip deps, no bundled weights, and unlike every other node in this pack it doesn't even need the H3 checkpoints loaded - it just introspects the latent tensor. If you want to add a quick sanity check to a workflow without any generation cost, this is the cheapest node in the suite.

Gotchas

The output info string is the truth: "H3 AV latent | N frames @ 24fps | video (...shape) | audio (...shape)". If frame_count looks like a nonsense number, the latent isn't what you think it is - don't push it through Continue, because the frame-count guard in the v1.4 path is strict by design. This node doesn't fix anything; it just stops you from spending an hour debugging the wrong tensor.