MiniMax H3 Native Latent Resume Manifest / 原生Latent恢复清单 (Advanced EXP/T8)

Long-running H3 jobs die. Power blips, drivers crash, you close the laptop lid mid-render. The painless recovery fantasy is "just load my saved latent and keep going" - and the reason that's a fantasy is that you have no idea whether the file on disk is actually the latent you think it is. MiniMaxH3NativeLatentResumeManifestT8Advanced is the node that answers that question with math instead of hope.

It builds a chunked, exact SHA-256 manifest of one complete H3 AV latent - video tensor, audio tensor, optional mask, metadata, all of it - tagged with a logical checkpoint_id. Save the manifest text somewhere. Later, after you've reloaded the latent, paste the manifest back in and the node recomputes and compares. Any mismatch in content, shape, dtype, mask, or checkpoint ID and the default mismatch_policy: error fails closed: you do not get a "resume_verified" you can trust, you get an error and a report.

What it is and isn't

The honest framing from the author is the important part. This node verifies. It does not save your latent - that's the pack's separate Native Latent Checkpoint Save/Load node. And critically, it is not diffusion-process checkpointing: it doesn't restore sampler state, sigma schedules, the queue, or CUDA state. "不冒充扩散内部断点恢复" - it won't pretend to be an in-diffusion resume. If you want to resume mid-sampling at exact step boundaries, that's the Dual-Clock NFE Checkpoint + Resume node, which has its own machinery (and its own contract compiler). This manifest node lives one level up: it certifies that the completed latent you're about to build on is the genuine article.

The inputs that matter

• av_latent - the complete H3 nested AV latent you want to certify.
• checkpoint_id - a logical name, timeline_checkpoint by default, that must match between the baseline and the verification.
• expected_manifest_json - leave empty to create a baseline; paste the saved manifest to verify a reloaded latent.
• mismatch_policy - error (safe default) or report_only. The tooltip is a gem: report_only exposes a MISMATCH report "without allowing it to be mistaken for a verified checkpoint." Use it to debug, never to declare victory.
• hash_chunk_megabytes - 8 by default. It bounds how much each temporary CPU hash chunk holds; it affects speed and memory, not the resulting hash. So a mismatch is never a chunk-size artifact.

Outputs: status, resume_verified (a real BOOLEAN you can gate on), content_sha256, and manifest_json.

Where it fits in a workflow

Wire it after the checkpoint Save/Load node in the pack's Long Video examples. First run: empty expected_manifest_json, grab manifest_json, tuck it into a text file or note node. After a reload: paste it in, and either read resume_verified or let the error policy stop you before you feed a corrupt latent into another 20-minute render. Chunked hashing means the hasher streams over CPU chunks, so it won't spike VRAM on a big latent either.

Install

This one ships inside the MiniMax H3 Audio T8 pack (Manager search "MiniMax H3 Audio T8"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8

Restart, then load examples/workflows/04-long-video/2026-08-23_H3_Native_Latent_Resume_Manifest_Advanced_EXP.json. No extra pip dependencies, no model downloads - this is pure latent bookkeeping, and it's one of the few nodes in the pack you can run on a machine with no H3 weights at all. It won't make your renders crash-proof, but it turns "I hope that file is right" into "I know."