Curating takes

Why you need this

Generation is fast; listening is not. A model can hand you eight versions of the Lumen Clash lobby loop before you have finished hearing the first one, and most will be close but wrong in some small way. Curating is stage 3 of the pipeline, generate and choose: producing takes in batches, judging them against a fixed checklist, and writing down enough about each that the keeper can be found and made again.

The idea

A take is one generated clip. A batch is several takes made from one request, each with its own seed (the number that sets the starting noise, covered in Seeds and sampling settings). Curating has three parts.

1. Batch. ACE-Step 1.5's batch_size setting makes 1 to 8 takes at once; the inference docs give a default of 2. A larger batch costs more memory, and the Gradio guide says the interface clamps batch size to what your memory tier allows. Its AutoGen option starts the next batch, same settings with new seeds, while you listen to the current one, and Apply These Settings to UI copies a batch's settings back into the input fields so you can iterate on a good one.

ACE-Step 1.5 can also score takes: the Gradio guide describes a perplexity-based quality score, and the tutorial a lyrics alignment score that rates how well the singing lines up with the lyrics. Use scores to decide what to hear first, never to decide what to ship. A score cannot hear a loop seam or a drum fill in the wrong place.

2. Listen with a checklist. The same questions, in the same order, for every take. Without a list you judge each take against the last one you heard, and drift.

Check	How to listen	What failure sounds like
Artifacts	Headphones, full volume range	Metallic swirl, warbling, crackle on hits
Tempo drift	Tap along, or loop 8 bars against a click at the set BPM	Beats slide early or late against the click
Low end	Small speakers and headphones	Bass and kick blur into one boomy smear
Prompt match	Read the prompt, then listen	Named instruments missing, wrong mood
Vocals (songs only)	Read the lyrics along	Wrong or mumbled words, skipped lines
Ending	The last 5 s	Cut mid-note, or a fade that never finishes
Loopability	Loop the best 8 bars many times	A jump in energy, key, or texture at the seam
Clipping	Peak meter and ears	Peaks at 0 dBFS, crackle on the loudest moments

Two habits keep the list honest. Match loudness before comparing: a louder take sounds better for being louder, so set every finalist to the same integrated loudness first (see Decibels and loudness). Listen on the speakers players will use. A card game is often played on a phone or laptop; small speakers drop most of the bass, so a take whose groove lives in the bass line can fall apart there. Do one pass on headphones for artifacts and one on a phone speaker for the real experience.

3. Log it. A take log is a table with one row per take, written as you listen, not afterwards.

Field	Why
Take ID	Matches the file name
Date	Model versions change; the date anchors the row
Model and version	acestep-v15-sft is a different take than acestep-v15-turbo at the same seed
Prompt	The prompt text, or the name of a prompt file kept beside the log
Seed	The one number that makes the take repeatable
Settings	Steps, guidance, duration, and anything else not at its default
Verdict	keep, maybe, or reject
Notes	Which checklist item failed, with a time ("drift at 0:31")

Name files so they sort. Put the most general part first, use an ISO date (year-month-day sorts correctly), and pad numbers with zeros so take 10 sorts after take 09. A pattern: lumen-lobby_2026-10-02_t03_s4129.flac. Take ID and seed in the name mean a file that strays from its folder still points back to its row.

Worked example

Illustration only: these takes, seeds, and verdicts show the method, not a record of real runs.

The Lumen Clash lobby loop. One request, made as one batch of 8 on ACE-Step 1.5: model acestep-v15-sft, base lobby prompt from Prompting for music, bpm 120, duration 48 (24 bars of 4/4 at 2 s per bar), inference_steps 50, guidance_scale 7.0, seeds 4127 to 4134.

Take	Seed	Pass 1 (headphones)	Pass 2 (checklist)	Pass 3 (loop test, phone)	Verdict
t01	4127	ok	pads warble at 0:22		reject
t02	4128	ok	clean, mood right	seam jumps in energy	maybe
t03	4129	ok	clean, mood right	bars 3 to 10 loop clean, bass survives	keep
t04	4130	piano missing			reject
t05	4131	ok	tempo drifts late from 0:31		reject
t06	4132	ok	clean	thin on phone, groove gone	maybe
t07	4133	ends mid-note			reject
t08	4134	ok	crackle on snare		reject

How the keeper was chosen:

1. Pass 1, all 8 takes, headphones. Anything wrong in the first listen goes: t04 (named instrument missing) and t07 (bad ending). Six remain.
2. Pass 2, checklist, at matched loudness. t01, t05, and t08 each fail one item, noted with a time. Three remain.
3. Pass 3, the best 8 bars looped on a phone speaker. t02's seam jumps, t06 loses its groove without bass. t03 holds.

t03 is the keeper. Its row gives everything needed to remake it: model, prompt, seed 4129, settings. t02 and t06 stay as "maybe": if t03 later fails in the game, they are the next listen, not a new batch.

Listening time: pass 1 is 8 takes of 48 s (384 s), pass 2 is 6 takes of 48 s (288 s), and pass 3 loops an 8-bar, 16 s section 8 times for each of 3 takes (384 s). Total 1,056 s, or 17.6 minutes, for one lobby loop.

In a game's audio pipeline

This is the end of stage 3, generate and choose. Settings come from Seeds and sampling settings; a near-miss take can be repaired with Variations and audio-to-audio instead of discarded. The keeper moves to stage 4 (Seamless music loops). Its log row becomes the start of the provenance record in stage 5 (Licensing and provenance).

Common mistakes

• Judging by the first 10 seconds. The problems live later: drift at 0:31, a bad ending. You ship a loop that stumbles every time it comes round.
• Comparing at different loudness. The louder take wins every time. In the game, after loudness matching, it is no better than the others.
• Writing the log afterwards. Seeds get swapped between rows, and the remake of "t03" is a different song.
• Trusting scores alone. A high-scoring take with a bad seam passes the screen and fails in the game.
• Only listening on studio headphones. The bass that made the take on headphones vanishes on the phone the player uses.
• Names that do not sort. take1, take2, take10 sorts as 1, 10, 2, and the keeper gets overwritten by a later "take3".

Cost

Generation is the small part. A batch of 8 runs in one request, limited by memory: each extra take in a batch needs its own working memory, so lower the batch on a smaller machine rather than the quality settings. Listening is the large part: 8 takes of 48 s is 384 s, 6.4 minutes, for a single pass, and the three-pass method above is 17.6 minutes. Disk is modest: 8 takes as 48 kHz, 16-bit stereo WAV are 73.7 MB (1 MB = 1,000,000 bytes); ACE-Step 1.5's default output format, FLAC, is lossless and smaller. Delete rejects once the log row is written; keep the log forever.

Going further

• Variations and audio-to-audio: fixing a "maybe" instead of rolling a new batch.
• Licensing and provenance: turning the log row into a provenance record.
• Decibels and loudness: why loudness matching comes before judging.
• The ACE-Step 1.5 tutorial's section on random factors, batches, and scoring.
• Try it: write your listening checklist on one card and use it unchanged for a whole batch.