MiniMax H3 SPEED Model + VAE Fingerprint / 模型与VAE指纹 (Advanced)

"model_v2_final_fixed.safetensors" is a promise, not a fact. Filenames lie - people rename files, two folders hold different checkpoints with the same name, and a spectrum profile fit on one H3 checkpoint has no business being reused on another. MiniMaxH3SPEEDModelVAEFingerprintT8Advanced settles the question with math: it streams the complete diffusion checkpoint and video VAE files and produces binding SHA-256 fingerprints. The whole SPEED calibration chain (Harvester → Accumulate → Finalize → Plan) wants these fingerprints so a spectrum profile is provably bound to the exact weights it was measured on.

How it works

You pick a checkpoint_name from the diffusion_models folder list and a video_vae_name from the VAE folder list - no typing paths, just dropdowns - and it SHA-256s the complete file bytes of each. The outputs are checkpoint_fingerprint, vae_fingerprint, and a report_json that records names, byte sizes, fingerprints, and confirms gpu_model_loaded: false.

Two design choices are worth knowing before you use it:

• It does not load the weights onto the GPU. Streaming the bytes through a hash is CPU + disk work. That means it's safe to run anywhere, but the first execution is genuinely disk-I/O heavy - hashing ~42GB of model plus a VAE takes a real while. ComfyUI may cache the result for unchanged inputs, so it's a one-time cost per file.
• The fingerprint covers the whole file. That's stricter (and slower) than a header-only check, but it means any byte-level change to the weights changes the hash. For binding a spectrum profile, that's exactly the property you want.

Where this plugs into the SPEED flow

The Harvester and Accumulate nodes each take checkpoint_fingerprint and vae_fingerprint inputs. The honest default is unrecorded - and the tooltip spells out the author's position: "Record a checkpoint SHA/header fingerprint; filenames alone are not proof." So the intended loop is: run this node once, paste its two fingerprints into your Harvester/Accumulate runs, and now every spectrum measurement is bound to those exact weights. If you swap models, the fingerprints change, and the downstream dataset refuses to pretend the old measurements still apply.

Installing

Standard pack install: ComfyUI Manager → "MiniMax H3 Audio T8", or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into custom_nodes, restart. No pip deps; you pick the files from folders you already have. Practical tips: run it once before a big calibration session (not in the middle), write the fingerprints somewhere you can reuse them across graphs, and don't panic when the first run sits at "calculating" for minutes - it's hashing tens of gigabytes, and the second run will be nearly instant. This is bookkeeping dressed as a node, but it's the bookkeeping that keeps SPEED's claims honest.