DiffusionGemma Advertisement Audio Audition & Lock

DiffusionGemmaAdvertisementAudioCandidateSelector is the music picker with teeth. In the advertisement stack, MiniMax Music 3 (or your upload) produces a few candidate soundtracks, and this node runs them through decoded-audio QC, picks a usable excerpt, and then locks the result by waveform hash so the rest of the pipeline can never silently swap in a different song. The "audition" in the display name is doing real work: this is the point where you listen, choose, and commit.

It's a content-aware version of the music-video pack's audio selector, adapted for ads. The key difference is the explicit content contract: soundtrack_contract_json says whether the campaign wants Instrumental, Vocal, or Auto, and the QC respects it. Instrumental music does not need a vocal proxy; vocal music does. So an instrumental candidate won't get failed for lacking a voice, and a vocal candidate can't slip through without one.

The mechanism, roughly: you can hand it up to four candidate_N AUDIO inputs and a candidate_count (1–4), then selection_mode decides the behavior. auto_select runs the QC heuristics - measured tempo, onset pressure, tonal consistency, clipping, silence, artifact detection - and picks. Or you lock explicitly: lock_candidate_1 through lock_candidate_4, or lock_by_hash against a locked_waveform_sha256 you already approved. expected_bpm (0 = don't enforce) and excerpt_duration_seconds (30 by default) bound the choice, and minimum_score (default 0.52) is the weighted numeric floor - importantly, just the floor: hard QC failures like clipping or silence block regardless of score. source_policy tells it whether the candidates are generated (full musical QC) or an uploaded track (technical-integrity checks plus the vocal-content contract).

Outputs are where the locking shows: selected_audio (the exact excerpt you approved), suggested_start_seconds, waveform_sha256 (the lock), plus director_report_json and audit_report_json with the evidence, and the universal status/ready pair. In the full workflow the selected report also feeds the DiffusionGemma Director so H3 camera and body motion follow the measured excerpt rather than the requested metadata.

The hard rule to internalize: this node fails closed. A failed lock blocks rather than falling back to some other song, because for an ad the soundtrack is contractual - the mixer, assembler, and QA gate all verify that the audio they receive still hashes to the waveform you locked here. If you hear a better take, don't fight the lock; re-run the selector with the new candidate and lock that.

The practical workflow is: generate your two to four candidates, run once in auto mode, look at the reports, then listen - the pack is explicit that QC is a lightweight signal heuristic, not a substitute for ears - and lock your chosen candidate and its excerpt start for repeatable production. Where people stumble: forgetting to connect soundtrack_contract_json and then being confused why a generated instrumental failed vocal checks, or re-running with a different seed and wondering why downstream won't accept the new audio - because it now hashes differently than the lock. Install is the pack install (ComfyUI Manager, search "DiffusionGemma Prompt Builder", or clone into custom_nodes); the workflow around it wants ComfyUI 0.33.1+ and the Music 3/H3 models, but this selector itself only decodes and measures audio.