MiniMax H3 Motion Context Archive Stitcher

Here's a workflow you might have met outside this pack: render a long H3 piece as a sequence of motion-continuation clips, each saved as an h3_motion_context_av_v1 safetensor, then stitch them back together. MiniMax H3 Motion Context Archive Reel is Creator Palette's final assembly for exactly those archives - the ones NikoDemon80's ecosystem and noEmbryo's H3 Motion Context Clip Stitcher produce. It decodes each clip once and dissolves the shared Motion Context head across the seam, in picture and audio.

It's an MiniMax/internal node, and it exists because Creator Palette deliberately doesn't reinvent the wheel for generated seams - it streams its own reel to disk - but archives made outside the pack still deserve a first-class door back in. The boundary algorithm is adapted from noEmbryo's MIT-licensed stitcher.py: a linear video dissolve plus an equal-power audio crossfade over the chosen H3 context length, then the whole thing lands on the same reel type the rest of the pack uses.

How it works

Point it at a folder of numbered clip_*.safetensors files and it:

1. Matches files against pattern, filters by first_clip/last_clip, and verifies there are no gaps - it refuses to stitch across a missing clip rather than let the timeline silently lie.
2. Decodes each clip once through your video_vae and audio_vae.
3. Overlaps consecutive clips by context_length frames (that's the shared Motion Context head) and dissolves them: linear video blend, equal-power audio crossfade.
4. Streams the result through the pack's spill/reel path - so the finished movie is never concatenated into one giant IMAGE tensor, the same memory discipline as the rest of the pack.

The report output gives you a per-seam audit (overlap, what each clip contributes, audio seconds), ending with the totals.

The inputs that matter

• video_vae, audio_vae - required; how each archived clip gets decoded.
• folder - default h3_context. An absolute path, or relative to ComfyUI/output (that's where the noEmbryo/NikoDemon80 tooling typically writes).
• pattern - default clip_*.safetensors.
• first_clip / last_clip - number range; last_clip 0 means "through the end."
• context_length - the overlap, and the one knob that matters. Options are 5, 22, 39, 56 frames - the H3 context lengths - default 22. Longer overlap = smoother motion carry but more of each clip gets folded into the dissolve. If your seams pop, go up; if faces smear during transitions, come down.
• fps - 24 default, the output rate.

Outputs: reel (ready for H3 Save to mux) and report (the STRING audit).

What it's not

It's not a motion-generator - that's the tool that made the archives. It's the assembly half, and it's here so you can keep using the archive tooling you already have while finishing the piece in Creator Palette's output pipeline. If it can't find the folder, it tells you the exact path it tried and whether you gave it an absolute one or one relative to output/ - the folder resolution is the first thing to check when it errors. Install is the shared pack install: ComfyUI Manager → search "H3 Creator Palette", or git clone https://github.com/z3rofeels/ComfyUI-H3-Creator-Palette into custom_nodes/, restart, hard-refresh. ComfyUI 0.34.0+; no extra Python deps (safetensors and torchaudio are ComfyUI's own).