MiniMax H3 Save AV Latent (Standalone)

Video generation is slow enough that you don't want to babysit it - and long-form H3 work is a lot of generation. MiniMaxSaveLatent ("MiniMax H3 Save AV Latent (Standalone)") is the pack's answer to "my ComfyUI restarted / I ran out of VRAM / it's 2am and I want to continue this chain tomorrow." It writes the sampler's complete joint audiovisual latent out to a safetensors file on disk, so the next session can load it and continue exactly where the last one stopped - no re-running the earlier clips.

This is the "standalone" alternative to the fancier Clip Bin system. The Clip Bin Saver archives a latent plus preview cards, ratings, tags and lineage for gallery-based picking. SaveLatent is the lean version: one file, one purpose, zero dependencies beyond what ComfyUI already ships. If you're working a single linear story and don't need a media pool, this is the node you actually reach for.

What it writes and where

One key thing to internalize: you save the full joint AV latent - the LATENT coming straight out of the sampler (video and audio packed together, in H3's NestedTensor form), not the trimmed pixel output. That full latent is the only thing rich enough to seed the next clip's protected head with audio intact. The node splits it into video and audio tensors and stores both in a single .safetensors file, with metadata stamped in (format: minimax_h3_av_latent, frame_count, clip_index, and the video/audio shapes). Files land in your ComfyUI output directory.

Inputs worth knowing

• latent (required) - the sampler's joint AV latent. The node will refuse a latent with no video stream.
• filename_prefix - default minimax_h3/clip. The folder part is created under the output dir, so minimax_h3/clip saves to output/minimax_h3/clip_*.safetensors.
• clip_index - the slot system. Set it to 1, 2, 3… and you get clip_00001.safetensors, clip_00002.safetensors, etc. - fixed slots that make "clip 2 was the good one" a real, addressable fact. The tooltip calls 0 "auto-incrementing," but read the code before you trust that: 0 actually writes to a single clip_temp.safetensors that the next save overwrites. It pairs with the Load node's "0 = latest file" logic for quick continuation, but for durable multi-clip chains use explicit numbered slots.

Outputs are saved_path (the absolute path written) and latent_info (a human-readable summary - frame count, video shape, audio shape). Both are strings; latent_info is handy for logging or feeding a text display.

Install

Part of the pack - 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. Requirements are just torch, safetensors and pillow - all already present in a working ComfyUI, and the node explicitly degrades to torch.save if safetensors were somehow missing.

Two grounded gotchas. First, this saves latents, and latents are tied to the model that made them - they're not a portable video format, and the safetensors won't mean anything to a non-H3 workflow. Second, a 10+ second H3 clip's latent is not tiny; budget disk space accordingly and don't point filename_prefix at your ComfyUI install directory by accident, because "output dir relative" is the intended home.