H3 Infinite - Load AV Latent

MiniMax H3 clips take a while to generate, and a long chain takes forever. If your ComfyUI crashes on clip 4, you do not want to re-run clips 1–3 to get the latent history back. That's what H3ContinuousLoadLatent is for: it loads a previously saved complete AV latent - plus its handover metadata - straight back into the graph, so you can pick the chain up exactly where you left off.

The input set is tiny and worth understanding precisely:

• latent_path - the specific safetensors file or a folder, absolute or relative to ComfyUI/output. Default h3_continuous.
• clip_index - "Clip to continue FROM. Generating clip 2: load 1." This is the subtle one and the source of most user error. The Save node stores clip N in the N-th slot, and you load the previous clip's number: to generate Clip 2 you load 1. 0 = newest file, which the tooltip honestly warns is "not retry-safe" - don't use it in a workflow you plan to re-run, because a newer file will silently change what you continue from.

Outputs: latent (the AV latent into your Continue node), resolved_path (what it actually opened - useful when you're unsure if the relative path resolved where you thought), latent_info (the same kind of frame/shape diagnostic the Latent Info node gives you), and handover - the saved H3_CONTINUOUS_HANDOVER metadata that the v1.4 Continue node requires. This is the key wiring: v1.4 Continue won't run without handover metadata, and in a load-based workflow this output is where it comes from. It's also why Save must have had the handover attached in the first place.

Install

Part of the Herrgotts-H3-Infinite-Continuation-Suite pack:

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

Restart ComfyUI, reload. No pip deps, no bundled weights; the H3 models come from Comfy-Org's MiniMax-H3 repo.

Gotchas

Clip indices are your mental model - this node is the reason the suite keeps telling you to keep sequential indices (Clip 1 → 1, Clip 2 → 2). If you loaded a latent whose frame count disagrees with the handover metadata's frame count, the v1.4 Continue node will hard-refuse with a mismatch error rather than silently producing garbage - that's a feature, and the fix is re-saving the source clip with matching metadata. And again: when it's time to generate clip N, you load clip N−1. Get that backwards and you'll be continuing from your own future.