H3 Continuation · Load AV Checkpoint

Continuation chains have a restart problem. You render clips 1 through 4, clip 4 is off, and now you'd like to re-render clip 5 without regenerating everything upstream. If the chain only exists in your graph's memory, you can't - you fix the bug and rerun from the top, burning 40 minutes per attempt.

This node is the escape hatch. Point it at the folder where Save AV Checkpoint has been writing latents, tell it which clip you want to continue from, and it gives you the full AV latent back, plus whatever handover metadata the file was saved with.

Inputs

Two widgets, both simple and both with a sharp edge.

latent_path is a specific .safetensors file or a folder - absolute, or relative to ComfyUI's output directory. The default is h3_continuous, matching the pack's own prefix convention.

clip_index is "the clip to continue FROM." Generating clip 2 means loading 1. That reads awkwardly for the first thirty seconds and then makes complete sense, because a continuation chain is defined by what it carries forward.

Set it to 0 and you get the newest file in the folder - with the tooltip's own warning attached: not retry-safe. "Newest" changes the moment ComfyUI writes anything, so a workflow that worked yesterday can quietly load a different clip today. Use explicit indices for anything you care about reproducing.

Outputs

• latent (LATENT) - the nested H3 AV latent, ready to feed Continue's previous_latent.
• resolved_path (STRING) - the file it actually opened. Worth a glance the first time you run a chain; it's also how you discover that clip 0 loaded something older than you assumed.
• latent_info (STRING) - a readable summary: frame count, video and audio shapes, saved head context, visible frames, technical padding, and the handover state (phase-aligned tail, legacy tail, freeze detected or not).
• handover (H3_CONTINUOUS_HANDOVER) - the typed metadata wire, ready for Continue's handover input or the Stitch node.

The mechanism worth knowing

The metadata isn't in a sidecar file. It's embedded in the safetensors header itself, as a JSON blob alongside the tensors - including the handover analysis, the saved head-context count, the visible frame count and the technical padding. That's why the loader can hand you a self-describing clip with no bookkeeping: the file knows how long its visible portion is and where the inherited history sits.

It's also why a metadata mismatch is survivable rather than fatal. If the embedded handover can't be parsed, the node logs a warning and returns available: false along with the latent's own frame count - you still get usable media, just without the cut recommendation. And it refuses outright if the file doesn't contain both video and audio tensors, which is a good sanity check that you're loading an H3 AV latent and not something else with a similar name.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

Then restart ComfyUI and hard-refresh the browser (Ctrl+Shift+R) - the pack has frontend pieces, and a stale tab is the usual explanation for anything looking wrong after an update. Keep one copy of the repo in custom_nodes; the README warns that duplicates produce stale node definitions.

No extra dependencies for this one beyond the pack and safetensors. No models needed to load a latent, though you'll need the H3 stack (model, text encoder, video VAE, audio VAE) before anything you load it into can actually generate.

The obvious workflow

Keep clip_index values aligned with reality. Clip 1 is index 1, clip 2 is index 2 - the same slot numbering the save node uses, so re-rendering clip 3 overwrites slot 3 and nothing else. When a chain goes wrong, you re-run only the tail.

One honest caveat: this is a mid-2026 pack, and H3's own weights carry a licence that excludes the US, EU, UK and South Korea from its grant - outputs included. That has nothing to do with loading a file, but it's the kind of thing you'd rather notice now than after a client delivery.