FireRedAudio 多 Take 试听评审板 · T8star-Aix

Show someone eight candidate takes with filenames like take-001.wav, take-002.wav and their "blind" review is rigged from the start - people anchor on the first file, the seed numbers, or whatever order they loaded. TakeReviewBoard fixes the anchoring problem for real: it presents up to 8 candidates as stable, anonymized A/B/C labels in a shuffled order, hides seeds and source names entirely, gives you a native player for each, and records 1–5 ratings and notes. Then you explicitly adopt one take, and only a second run produces it as the chosen output. The review writes a new manifest; the source audio is never touched.

How it works

It takes an audio_batch and shows its successful items as native players in a stable shuffled order - the shuffle is derived from a digest of the manifest + each file's hash + line ID, so the same batch always shows the same anonymous arrangement (no reshuffling between reviews of the same set). Ratings and notes are JSON keyed by line ID ({"take-001": 5}, {"take-001": "natural"}), validated against the batch's actual line IDs - an unknown ID raises instead of silently vanishing. Unknown ratings or out-of-1–5 scores are rejected.

The inputs: audio_batch, selected_position (default 0), optional selected_line_id, ratings_json, notes_json, review_name, subfolder, preview_limit. selected_line_id takes priority over position when both are set.

The run-twice behavior - read this twice

First run with defaults, nothing is adopted. selected_position 0 is "blind listen only": all adopted flags stay false, deliberately. You listen, rate, note, then set the position (or line ID) of the take you're adopting and run again - the node validates your choice, writes the review manifest, and only then does 已采用 Take come out. The candidates keep their opaque anonymous filenames; the seed lives only in the audit manifest. It's a small ceremony, and it's the entire point: the author built the default so a first run can't accidentally auto-adopt candidate #1.

Outputs: 已采用 Take (AUDIO), 带评审记录 AudioBatch (the batch with ratings/notes attached), 已采用 line ID, 评审 Manifest 路径, 评审报告.

Where it fits

It's the human checkpoint at the end of any multi-take flow: SeedAudition's 全部候选, ReferenceCandidates' 全部候选, or a CreativeCandidatePool's anonymous candidates all feed straight in. The 24_seed_blind_review and 28_creative_line_candidates examples show the pattern - candidates in, blind listen, explicit adopt, adopted take onward. It's the only node in this pack where "run it twice" is correct behavior, and forgetting that is the classic first-use confusion.

Installing

Pack-wide install (ComfyUI Manager search comfyui-fireredaudio-T8, or clone + python scripts\setup_runtime.py). No model needed - it plays files and writes manifests. The one trap beyond the run-twice flow: rating JSON with a stray space or wrong case in a line ID will error out with "unknown line ID" - copy line IDs from the manifest rather than retyping them.