MiniMax H3 Native AV Latent Checkpoint Save / 原生音画Latent检查点保存 (Advanced EXP/T8)

Here's the mental model to get right before touching this node: "checkpoint" here does not mean "resume my sampler after it crashed mid-diffusion." MiniMax H3 generates video and audio together as one joint latent, and this node saves that completed joint AV latent to disk as a .h3latent.safetensors file, then verifies it. It's a way to park a finished latent so you can reload it in a later session, do the decode again, or inspect it - not a "continue from step 5" button. The pack says so bluntly, and it's worth believing: default behavior writes nothing, and it never claims to be a diffusion-internal resume.

What it does well is paranoia. The file is a no-pickle safetensors (safe against the pickle-execution attacks you don't think about until they matter), written atomically under the ComfyUI output directory, and never overwrites an existing file. It records an exact content manifest - tensor shapes, dtypes, masks, checkpoint ID - and computes a whole-file SHA-256. verify_after_write is on by default, so the temp file is reloaded and checked against its own manifest before it's atomically placed.

Inputs

• av_latent - the joint H3 audio/video latent to save.
• filename_prefix - a relative prefix; a unique suffix is always added. Absolute paths and .. traversal are rejected, which is the kind of guardrail that keeps you out of trouble.
• checkpoint_id - a logical identity that must match when you reload. Treat it as the "name" of this checkpoint.
• confirm_save - the safety gate, false by default. With it off, the node computes the exact content manifest and writes nothing. You flip this on deliberately after reviewing what you're about to write. This is the node's way of saying "are you sure."
• verify_after_write - keep on. It's the atomic-placement verification pass.
• hash_chunk_megabytes - just a CPU hash chunk bound; changes memory/speed only, never the digest.

Outputs

status, checkpoint_path (the relative path to hand to the Load node), file_sha256, manifest_json, report_json, plus the av_latent passed through unchanged.

Install and usage

Shared pack install: ComfyUI Manager → search MiniMax H3 Audio T8, or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into ComfyUI/custom_nodes/, restart. No pip extras, no auto model downloads - bring your own H3 weights.

The workflow rhythm is: finish a segment or a long-video section, save its latent with this node, later reload with the Load node and decode. Files land under output/MiniMaxH3/latent_checkpoints by default. Because it never overwrites and defaults to not-writing, nothing breaks silently - the failure mode is "nothing was saved because I left confirm_save off," which is exactly what you want when you're experimenting. It's Experimental, so don't build your entire pipeline on it yet, but if you've ever lost a 20-minute H3 render to a mid-chain queue reset, this is the node that parks the result safely.