H3 Infinite - Stitch Saved Chain v1.2

Every H3 clip tops out at a few seconds, so a "long" video is really a stack of clips. H3 Infinite - Stitch Saved Chain v1.2 is the node that assembles that stack at the very end, from latents you saved along the way. Its superpower is right in the name: it's memory-bounded. It decodes one saved AV latent at a time and encodes straight to MP4, so your VRAM (and RAM) usage stays flat no matter whether you're stitching 3 clips or 30. No giant all-clips-in-memory batch, no decoding an entire project just to join it.

The workflow it expects is the classic "generate clip by clip" pattern: make each clip with the Start/Continue nodes, save each one with H3 Infinite - Save AV Latent keeping sequential indices (clip_00001.safetensors, clip_00002.safetensors, …), then run this node once at the end. It reads each saved full video+audio latent, decodes it with the MiniMax H3 VAEs, and writes an H.264/AAC MP4 directly to your output folder.

The inputs that matter

• video_vae / audio_vae - the MiniMax H3 Video VAE and Audio VAE used to decode the saved latents. Wire the same ones you generated with.
• latent_prefix - where the saved clips live, relative to ComfyUI's output folder. Default h3_continuous/clip matches the naming in the example workflows.
• first_clip / last_clip - the range to stitch. Set last_clip to 0 and it automatically uses the highest numbered clip it finds for that prefix, which is handy mid-project.
• filename_prefix - where the finished MP4 goes (video/Herrgotts_H3_Infinite_Stitched by default).
• video_crossfade_frames (4 recommended) and audio_crossfade_ms (15 recommended) - the seam smoothing. Short context-aligned video blend plus a brief audio de-click crossfade so joins don't pop.
• crf - H.264 quality, 18 is the high-quality default. Lower = bigger/lossless-er, higher = smaller.
• max_safe_tail_bridge_frames (2 recommended) - this is the v1.2 seam-repair knob. H3's latent phase alignment sometimes discards a couple of perfectly good rendered frames at a boundary; Safe Tail Bridge reuses up to N detector-approved frames from the previous clip and skips the same number of early frames in the next one, instead of just dropping them.
• luminance_match and friends - an experimental fallback only. Off by default. Only reach for it if a persistent brightness seam shows up at joins; the Safe Tail Bridge path is the intended default.

What it hands you back

Two outputs: video_path (the full path to the written MP4) and stitch_info (a summary string describing what it did, including per-clip trims and seam choices).

Where people get burned

• It needs PyAV. The node imports it lazily, and it's normally present because it's part of current ComfyUI's video/audio stack - but if you're on a stripped-down install you'll get a "PyAV is required" error. That's not a pack bug; install/update PyAV.
• Everything must match across clips. Same resolution (and even dimensions - H.264 requires it), same audio sample rate, and audio must be mono or stereo. Mix resolutions mid-chain and it errors rather than rescale for you.
• Ambiguous prefixes. If you saved chains into overlapping folders, it refuses to guess and tells you to use the exact latent_prefix.
• Don't restart mid-range. Keep indices sequential; last_clip = 0 is a convenience for "up to the highest", not a fix for skipped numbers.

Install the whole pack via 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

then restart ComfyUI. The pack has no model weights - the MiniMax H3 files come from Comfy-Org's repo - and no extra pip dependencies beyond what ComfyUI already ships.