MiniMax H3 SPEED Spectrum Dataset File / 频谱数据集文件 (Advanced)

The Accumulate node builds your H3 spectrum dataset in memory, but in-memory data dies with the process - and dataset building spans many ComfyUI restarts by design. MiniMaxH3SPEEDSpectrumDatasetFileT8Advanced is the persistence layer: it loads an accumulated spectrum_dataset from disk or explicitly saves it back, under the ComfyUI output directory, atomically. It stores no source latent or video (only the accumulated statistics), it requires confirmation before writing, and it refuses silent overwrites and unsafe paths. The author's file-handling philosophy, in one node: explicit, auditable, recoverable.

How it works

The mode combo (default load) picks the direction:

• load - reads dataset_name from the output directory and returns the spectrum_dataset object, ready to feed back into Accumulate for another batch or into Finalize once you're done.
• save - writes the connected spectrum_dataset to output/MiniMaxH3/... (or wherever the pack's output layout puts it), returning file_path and a report_json with what actually landed on disk.

For save you need two honest-to-goodness confirmation knobs: overwrite (false by default - existing files are not clobbered silently) and confirm_write (false by default - you must explicitly flip it, the pack's standard double-consent for anything that writes). The node is an output node, so it terminates a branch and shows you the result path. Outputs: spectrum_dataset (pass-through, so save can sit in a line), file_path, and report_json.

Why "atomic" matters

The description promises atomic saves: write to a temp location, verify, then move into place. In practice that means a crash mid-save can't leave you with a half-written JSON that reads back as a corrupt dataset - which matters enormously for a thing you've spent dozens of runs accumulating. Same story on load: if the file is missing or the format is wrong, you get an explicit error rather than a silently empty dataset that would then get accidentally re-accumulated and "confirmed" with bad data.

The honest limits

This node handles the dataset container, not the profile. If you load a dataset and feed it to Finalize, Finalize is what turns statistics into a spectrum_profile - and Finalize enforces its own 100-clip/quality bars. Don't expect this node to bless weak data; its job is narrower, and it does it strictly. Also worth knowing: because save writes under the output directory, cleanup of old datasets is your job - the pack never deletes anything for you.

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; purely file bookkeeping, no model, no GPU. Typical lifecycle: Accumulate → this (save, confirm) → restart → this (load) → Accumulate more → ... → Finalize. The double-confirm on save will annoy you exactly once, and then you'll be grateful the pack made it impossible to accidentally destroy a week of accumulation.