MiniMax H3 SPEED Spectrum Dataset Accumulate / 频谱数据集累积 (Advanced)

One clip's spectrum is a curiosity. A hundred independent clips' spectrum is a dataset - and that's the bar the pack sets before it will let a spectrum profile authorize SPEED's delta-optimal transitions. MiniMaxH3SPEEDSpectrumDatasetAccumulateT8Advanced is the node that gets you from 1 to 100 without blowing up your machine: it accumulates exact per-clip H3 spatial power statistics across separate ComfyUI executions, storing only the statistics, never the source latents or CUDA tensors. You can feed it one clip per run, restart ComfyUI between runs, and keep building the same dataset.

How it works

Each execution takes one video_latent (again: the separated [B,24,T,H,W] video latent, never the joint AV latent) and folds its spatial power statistics into an accumulated spectrum_dataset object. The guardrails are where the design shows:

• batch_id - a unique, stable ID for the actual source batch you're measuring. Repeats are rejected. Run the same batch twice and it fails closed rather than double-counting your data. This is the anti-cheating mechanism: you can't inflate your clip count by re-running the same latent.
• task_family, checkpoint_fingerprint, vae_fingerprint - the contracts that keep the dataset homogeneous. "Model, VAE, task and latent contracts must remain identical" is the node's own rule; change your checkpoint mid-dataset and it refuses to keep accumulating, because a spectrum dataset mixing two different models is meaningless.
• max_temporal_samples (32) - how many frames per clip get sampled.
• previous_dataset (optional) - feed the accumulated dataset back in to continue it across runs, which is the whole point: run → accumulate → save the dataset via the File node → restart → load it back → accumulate more.

The optional dataset_provenance_json and source_entry_json inputs let you attach reviewed provenance (where the clips came from, their manifest-bound hashes) so the "independence" claim has evidence behind it rather than being vibes.

The workflow shape

This is a batch job, not a one-shot. Realistically: you generate a batch of H3 clips, separate their video latents, accumulate them one per run (or per clip), then use the File node to persist the dataset, then start the next batch. It's deliberately tedious, and that's the point - dataset building should be tedious enough that you notice when you're about to cheat it. The author would rather the whole delta-optimal SPEED path be unused than used with garbage data.

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; H3 weights on you. Pair it with the Fingerprint node (to fill the contract hashes honestly), the File node (to persist), and the Finalize node (to fit the profile once you've crossed the 100-clip line). Expect duplicate-batch-ID errors in normal use - that's the guard doing its job, not a bug.