MiniMax H3 Native AV Latent Checkpoint Load / 原生音画Latent检查点加载 (Advanced EXP/T8)

Pair this with the Native Latent Checkpoint Save node and you've got the H3 version of "save my work, come back later." MiniMaxH3NativeLatentCheckpointLoadT8Advanced reloads one no-pickle .h3latent.safetensors file from the ComfyUI output store and, before it hands you the latent, verifies it. Read the description once and you'll see the ethos: it checks the embedded exact-content manifest, and optionally demands an independently saved manifest and the whole-file SHA-256. Loads that don't match fail closed.

What it restores matters: a completed H3 joint audio/video latent - video and audio as one nested latent, the way H3's joint AV transformer produces it - plus any optional mask and supported metadata. What it is not: a diffusion resume. The pack's README is explicit that this node doesn't pretend to restore NFE state or continue an interrupted sample. It's "I saved this finished result and I want it back," not "resume generation from step 7."

Inputs

• checkpoint_path - the relative .h3latent.safetensors path returned by the Save node. You can paste it after a restart, or wire the Save output straight in within the same graph.
• expected_manifest_json - optional, but this is where the paranoia lives. If you supply it, any latent/mask/metadata/dtype/shape/checkpoint-ID mismatch fails closed instead of loading.
• expected_file_sha256 - optional whole-file hash from Save. Use it to detect a file where both the payload and the embedded manifest were replaced. This is the "someone swapped my file" tripwire.
• hash_chunk_megabytes - CPU hash chunk size; changes memory/speed only, not the digest.

Outputs

av_latent (the restored joint latent, ready for decode or re-encode), plus a whole dashboard of verification results: status, resume_verified (boolean), checkpoint_id, content_sha256, file_sha256, manifest_json, and report_json. If you're automating a long pipeline, resume_verified is the signal you gate on.

Install and honest expectations

Same pack install as always: 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 model downloads - you already have the models from the rest of your H3 setup.

Where this earns its keep: long-video work where a big latent is the expensive output of many minutes of sampling. Save it, reload it next session, decode when you have time. The gotcha is expectations - people see "checkpoint" and expect resume-capable sampling, then get frustrated that this only restores completed latents. It does that one job, verifiably. Use it for exactly that and it's great; use it to skip work and it'll just reload the finished thing you already had. Experimental label applies, so treat the file format as pack-specific rather than something the rest of the ecosystem reads.