H3 Infinite - Auto Handover — Masked AV v1.4

The v1.4 Auto Handover analyzer is the node that makes v1.4's entire philosophy work, and it's worth understanding even if you never touch a knob, because it's now the single source of truth for where every seam lands. Feed it the decoded full render of the accepted H3 clip, and it finds the safe continuation boundary: it detects the FL2VA freeze/unstable landing tail, excludes it, and then snaps that visual cutoff backward to the latest exact Masked-AV video boundary - the 39-frame (or larger) shared AV grid. Both Stitch Ready (the visible end of the previous clip) and the next clip's protected video head end at that same frame. One boundary, two uses, no mismatch. That's the whole upgrade from older versions, where the render cut and the continuation source could disagree.

The consequence you'll actually feel: the previous visible video and the next protected context now end at the same exact frame, so the stitched seam is internally consistent. And because the boundary is discrete, that final snap can move the handover up to ~16 rendered frames earlier than the detector's ideal visual cutoff. That's deliberate - the author live-tested recovering those grid-lost frames and got visible motion/alignment errors, so the release prioritizes a clean seam over squeezing back a few frames.

Inputs

• images - decoded full rendered frames from the accepted clip.
• preset - Balanced (default; validated detector with freeze_hold 8 and fixed 3-frame safety), Motion Safe (same detector, 6-frame safety), or Custom.
• context_frames - fixed to "39" in v1.4, the minimum exact Masked-AV boundary. The tooltip is worth quoting: audio tail carryover is independent and does not change this video handover.
• Custom-only advanced controls - analysis_window, freeze_hold, safety_margin, analysis_size, the primary/secondary thresholds, and the robust gates (min_final_match_percent, max_consecutive_final_outliers, etc.), plus safety_mode (fixed recommended; adaptive is legacy).

Outputs

handover (the metadata both Continue v1.4 and the stitcher consume), analysis_info (STRING), freeze_detected, freeze_start_frame, ideal_handover_end_frame, phase_aligned_handover_end_frame, ignored_tail_frames, and confidence. The three frame ints + confidence are the debug trio when a seam is off - and if ignored_tail_frames is huge, your clip had a long landing and the snap is doing exactly its job.

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 pip deps, no bundled weights - H3 models/VAEs from Comfy-Org's MiniMax-H3 repo. And the v1.4 caveat applies: this node's metadata is consumed by the masked path, which needs a current ComfyUI with PR #15375 native H3 AV-mask support.

Gotchas

Feed it the render of the accepted clip, matching the latent you're about to save - if the analyzer's frame count disagrees with the saved latent's frame count, v1.4 Continue will refuse the handover rather than guess. And for dialogue, remember the video boundary here doesn't constrain audio: the Continue node's audio_tail_carryover can independently extend protection through the previous clip's full tail, so don't be confused when the analyzer's video cut and the audible audio seam don't line up. That's the feature.