H3 Infinite - Continue — Native Masked AV v1.4

This is the flagship, and the one you should build new chains with. H3ContinuousContinueV14 changes how continuation works at a mechanical level: instead of reconstructing the previous clip's context through guide/keyframe rows (the v1.3 way), it copies the previous video+audio latent directly into the new target and protects it in place with ComfyUI's native per-stream denoise masks - video mask 0 means "preserve these frames," mask 1 means "generate the rest." No decode, no re-encode, no regeneration of the seam. That's the "Native Masked AV" in the display name, and it fixes the two things that historically made long H3 chains janky: brightness mismatch at the seam and dialogue endings getting cut off because the picture had to cut early.

The catch, and read it before anything else: v1.4 requires a current ComfyUI build with native MiniMax H3 AV-mask support from PR #15375. The node probes the live runtime rather than trusting a version string, and it deliberately ships no compatibility shim. If it reports Masked AV unavailable, update ComfyUI, fully restart, and hard-reload the browser. Older ComfyUI can still use the registered v1.3 nodes - just not this one.

The three settings that changed everything

• masked_context_frames - the protected previous-video context, default 39 (~1.625 s at 24 fps, exactly 65 audio-latent ticks at H3's 40 Hz audio rate). Legal exact joint AV boundaries are 39/90/141/192/…; 39 is the recommended starting point, the bigger ones eat your generation budget.
• duration_mode - Net New Content is now default: it picks the nearest legal H3 17k+5 total so ~your requested duration remains as newly generated video after the 39-frame head is removed. "Total Generation" keeps the old meaning (duration counts the whole latent). Net New Content is a convenience, not a speed hack - it samples a longer total, so it costs more time and VRAM.
• audio_tail_carryover - Full Previous Tail is default and is the dialogue-saver: video must cut at the freeze-safe boundary, but audio can stay protected through the previous clip's actual remaining tail, so a word doesn't die because the picture needed a cut. Match Video Handover reproduces the old behavior for A/B testing. audio_feather_ticks should stay at 0, especially for speech.

Required: clip, vae, previous_latent, prompt, width, height, duration, ref_image_size. Optional: handover (mandatory in practice - v1.4 refuses to run without the previous clip's H3_CONTINUOUS_HANDOVER metadata), last_frame, qwen_reference_1. Outputs: positive, latent, actual_head_context_frames (wire this to the stitcher!), ignored_tail_frames, handover_info, picture_map.

Install

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

or ComfyUI Manager → search the pack title. Restart, reload. No pip deps, no bundled weights - the H3 FL2VA transformer, video/audio VAEs, and Qwen 3 VL come from Comfy-Org's MiniMax-H3 repo (that's the multi-GB download). Optional SageAttention/KJNodes for speed, explicitly not required - and if you hit OOMs, disable SageAttention first when troubleshooting.

Gotchas

The 39-frame protected head still exists internally; Net New Content compensates by sampling longer, so it's slower, not free. Dialogue tail carryover can only preserve audio that was actually generated - if the source clip itself ends mid-word, that continuation still has to be synthesized. And the author's own advice: for best results use Full Previous Tail + audio_feather_ticks=0, keep a Last Frame connected as a quality reset, and don't mention the protected context in your prompt - it's already in the target latent.