MiniMax H3 Clip Bin Saver (Media Pool)

Long-form H3 work generates a pile of clips, and the whole problem this pack solves is that the next clip you make depends on choosing the right previous one. MiniMaxClipBinSaver ("MiniMax H3 Clip Bin Saver (Media Pool)") is the archive side of that: it takes a finished segment and files it into a per-project "clip bin" with a preview card, a star rating, a shot tag, the prompt you used, and the family tree of which clip it was continued from. Later, the Clip Bin Picker shows you those clips as a gallery so you can continue from the take you actually like - instead of decoding five files by name to remember which run was the good one.

This is a media pool, not a video stitcher. It doesn't concatenate anything or make your final film; it organizes the raw material - latents and previews - so a human (or the Auto chaining modes) can make informed choices.

What you must feed it

Three required inputs plus the latent:

• latent - the sampler's full joint AV latent. Same rule as everywhere in this pack: the continuation context must be the complete, untrimmed latent with both video and audio streams. Save it even though you'll trim pixels for delivery.
• project_name - which bin this clip belongs to (e.g. Default_Project, sci-fi_short). Projects are fully isolated from each other, which is how you keep several stories from cross-contaminating.
• shot_tag - a free-text label ("hero turns", "Take 2", "wide empty shot"). Leave Auto (自动编号) and it auto-numbers as Shot 1, Shot 2, … based on how many clips are already in the project.
• rating - 1–5 stars (default 4). This is the quality signal the Picker's rating filter later uses to skip the duds, so be honest - marking everything 5 stars defeats the one workflow feature that keeps garbage out of your continuation chain.

The optional slots that make it good

• images - decoded frames of the delivered segment (connect your trimmed output). With this connected the saver extracts the real first and last frames and builds the side-by-side preview card.
• audio - the segment's audio, muxed in when it encodes a video.
• prompt - the positive prompt for the segment, stored to meta.json so you can recall what each shot was for.
• parent_clip_id - wire the previous Picker's clip_id here and the bin records parent-child lineage, so branching and retakes have a traceable history.
• video_file_name - connect the Filenames output of a VHS_VideoCombine (or type an MP4 name) and the saver will copy that finished video into the clip's asset folder. The wildcard input accepts VHS's list format too.
• save_video (default on) - with this on, the saver archives an MP4 for each clip: it copies the associated video if you gave one, or encodes one from your images if you didn't.

Outputs: clip_id (a timestamped ID like clip_20260908_… - keep this, it's the lineage handle), preview_image (the card), and bin_path (where the clip folder lives).

Where clips go, and the ffmpeg catch

Storage is under your ComfyUI output dir: output/minimax_h3_bins/<project>/<clip_id>/, holding the latent safetensors, preview.png, meta.json, and optionally video.mp4. That self-contained folder is the "self-sufficient asset package" idea - latent plus finished AV in one place.

The grounded gotcha: that auto-encoding shells out to ffmpeg on your PATH, not an imageio fallback. If ffmpeg isn't installed/found, the node logs a warning and silently skips video encoding - your clip is still archived, but you get no MP4, which means the Picker gallery loses its hover-play and audio playback for that clip. On Windows/portable installs that don't add ffmpeg to PATH, check this first when videos go missing.

Install

ComfyUI Manager (search "MiniMaxH3 PrefixStream"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/knoic/ComfyUI-MiniMaxH3-PrefixStream.git
cd ComfyUI-MiniMaxH3-PrefixStream
pip install -r requirements.txt

Restart, Ctrl+F5. Dependencies are just torch/safetensors/pillow. Remember the H3 weights themselves are separately licensed - the MiniMax H3 Community License excludes US/EU/UK/Korea - and this node does nothing about that, but also adds no license of its own beyond the pack's MIT.