H3 Continuous - Continue from Latent (Legacy v1.0)

The v1.0 Continue node is the same direct video+audio latent continuation as the v0.x original, with the one quality-of-life fix that made the whole suite usable: you now type duration in seconds, not frames. 10 seconds in, and the node snaps it up to H3's 17k+5 temporal grid - 243 frames, which is 10.125 s at the native 24 fps. That sounds trivial until you've built a 7-clip chain where every node wanted a frame count and you had to do the math in your head.

Class H3ContinuousContinueV1 is the "Clip 2+" node: previous clip's latent goes in one side, a prompt and optional last frame go in the other, and out come the positive conditioning and the new latent for the sampler. Same phase-aligned direct-latent handover as v0.x - it carries real latent history forward rather than reconstructing it, which is what keeps motion and native audio coherent across the seam.

The pieces that matter

• previous_latent - the sampler output of the last accepted clip. This is the source of the carried history.
• duration - seconds at 24 fps. Default 10, snaps upward to the H3 grid.
• handover_mode - AUTO (default) consumes freeze-analysis metadata from the previous clip's saved latent; MANUAL falls back to manual_landing_tail_frames (34). Auto is the intended workflow.
• alignment_mode - phase_aligned_extended is the recommended v1.0 handover. phase_aware and legacy_17 exist only to reproduce older workflows or A/B diagnostics; legacy_17 requires a multiple-of-17 tail.

Outputs to remember: actual_head_context_frames is the number of reused frames that the stitcher must trim, ignored_tail_frames tells you how much of the dead FL2VA tail was skipped, and handover_info is a string dump for when things look wrong. positive and latent go straight to your H3 sampler.

Optional inputs: handover (the analyzer's metadata, if you're not reading it from the saved latent), last_frame, and reference_image for endpoint guidance.

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 and reload the browser. Nothing fancy to install - the pack extends ComfyUI's native MiniMax H3 support and ships no weights; the H3 checkpoints, video/audio VAEs, and Qwen 3 VL encoder come from the Comfy-Org MiniMax-H3 repo. That download is the expensive part, not this node.

Gotchas

Duration snapping is upward, always - a requested 5.0 s can come out as 124 frames (5.17 s) because that's the nearest legal 17k+5 length. That's H3's grid, not a bug in this node. And be careful with alignment_mode when reproducing old workflows: legacy_17 is strict about multiples of 17 and will error if you hand it 34 frames when the layout math disagrees. For new projects, the v1.4 masked Continue is the recommended path; v1.0 is the stable, boring workhorse that old graphs still run on.