H3 Infinite - Stitch Saved Chain — Masked AV v1.4

H3 Infinite - Stitch Saved Chain - Masked AV v1.4 is the v1.4 sibling of the suite's memory-bounded saved-chain stitcher, and it does the same job: take the clip_00001.safetensors-style latents you saved clip-by-clip, decode one at a time with the MiniMax H3 VAEs, and write an H.264/AAC MP4 without ever holding the whole chain in memory. If you're running the v1.4 workflows, this is the node the "Stitch Saved Chain" example points at.

What's different from v1.2

The stitcher's mechanics are basically unchanged, but its defaults and expectations now match the v1.4 native Masked AV worldview:

• Safe Tail Bridge is off (max_safe_tail_bridge_frames = 0). In v1.2, the stitcher bridged a couple of frames that latent phase alignment had discarded. v1.4 doesn't need that crutch - Auto Handover snaps to one exact shared AV boundary, so both the visible clip and the next protected head end at the same place. Keep it at 0.
• Luminance matching is marked legacy/diagnostic. Native Masked AV solved the brightness seam the suite used to see; luminance_match stays off unless you're deliberately A/B-testing an older seam.
• Extended audio survives the head trim naturally. Because v1.4 protects the previous clip's audio tail independently of the video, that carried-over dialogue is already sitting in the decoded next clip - the normal video-head trim doesn't throw it away, and the stitcher preserves it automatically.

The rest of the inputs are the familiar set: video_vae and audio_vae, latent_prefix (default h3_continuous/clip), first_clip/last_clip (0 = auto-use the highest numbered clip), filename_prefix, video_crossfade_frames (4), audio_crossfade_ms (15), and crf (18). Outputs are video_path and stitch_info.

Compatibility is the quietly nice part

The author explicitly kept the older paths alive. Early experimental v1.4 saved metadata that used a different protected-source offset still stitches via a retained compatibility path, and the v1.2 guide-based chains from older projects still load. So you're not locked into finishing a project with the exact node version you started it with - upgrading to v1.4 mid-project shouldn't strand your saves.

The things that bite, same as before

• PyAV is required - lazily imported and normally provided by ComfyUI, but a bare install will hit a clear error if it's missing.
• Clips must agree: matching resolution (even dimensions for H.264), matching audio sample rate, mono or stereo only.
• Sequential indices matter - last_clip = 0 discovers the highest existing clip, but gaps in the middle of the range will still fail.
• Current ComfyUI required for v1.4 generation: this stitcher reads the latents and VAEs, but the v1.4 Continue nodes upstream need ComfyUI's native MiniMax H3 AV-mask support (PR #15375). If you're on an older build, the v1.2 nodes remain registered and usable.

Install the whole pack through ComfyUI Manager (search Herrgotts-H3-Infinite-Continuation-Suite) or:

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

Restart after. No pack-specific pip packages, no bundled weights - the MiniMax H3 models (pruned FL2VA checkpoint, Qwen refs, video and audio VAEs) come from Comfy-Org's repo as usual.