H3 Infinite - Continue from Latent v1.2

Class H3ContinuousContinueV11 is the "H3 Infinite - Continue from Latent v1.2" node, and it's the version that made the suite's long-form promise real: the freeze-aware handover from the v1.2 analyzer now actually gets consumed, lock metadata or no-lock fallback, and the continuation runs phase-aligned direct video+audio latent carryover. For a beginner this is the node to study, because the v1.2 generation nodes are the cleanest example of the whole idea: clip 1 gets made with a Start node, saved, analyzed, then every clip after that is one Continue node.

It looks almost identical to v1.0 - same duration in seconds, same handover_mode, alignment_mode, manual_landing_tail_frames, ref_image_size, same optional handover, last_frame, reference_image. The difference is in what AUTO handover now means. The v1.2 analyzer (class ...AnalyzeHandoverV11) produces richer metadata: when it detects a freeze it marks the safe endpoint, and when it can't detect one it records a no-lock fallback that excludes freeze_hold - 1 ending frames. This Continue node understands both cases, so the "previous clip was still moving and I cut into real motion" failure mode got a lot rarer.

How to use it

Wire up clip, vae, and prompt as you would any H3 gen. Plug previous_latent from the previous clip's sampler output. Set duration in seconds (default 10, snapped up to the 17k+5 grid). Decide on a last_frame if you want an endpoint/quality reset - that's the recommended FL2VA-style workflow. The positive and latent outputs go to the sampler, and actual_head_context_frames goes to your stitch/seam nodes so they can strip the duplicated head.

Inputs a beginner actually adjusts:

• handover_mode - AUTO (default) uses the analyzer metadata; MANUAL uses manual_landing_tail_frames.
• alignment_mode - keep phase_aligned_extended. phase_aware/legacy_17 are for reproducing old behavior only.
• context_frames - minimum carried history, default 22; phase alignment can extend it backward.

Install

From 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 the UI. No extra pip deps - the pack wraps ComfyUI's native H3 support - and no bundled weights. The H3 transformer, VAEs, and Qwen 3 VL encoder come from Comfy-Org MiniMax-H3; with the pruned int8 FL2VA transformer plus the ~32B Qwen 3 VL, budget for a large download and real VRAM.

Gotchas

v1.2 is where the manual Save/Load latent workflow got standardized - clip index 1, 2, 3 - so keep indices sequential or the Load node grabs the wrong clip. And note the direct-latent path still carries a duplicated head that the stitching must remove; that's actual_head_context_frames, and forgetting to wire it into the stitcher is the classic cause of doubled-motion seams. It's a good node and it still works, but for new projects the v1.4 masked continuation is the recommended direction - v1.2 is the reference implementation that made the suite credible.