loop-produce
/loop-produce
Renderuje gotową pętlę dźwiękową przy lokalnej instalacji ACE-Step z zatwierdzonego briefu i eksportuje ją do pliku WAV w projekcie.
- Autor
- alexandrehebert ↗
- Repozytorium
- soundlines ↗
- Licencja
- Apache-2.0
- Ostatnia zmiana repozytorium
- 13.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 8
- Skrypty
- 6
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 13.11.2026, bez odwołań do wycofanych modeli i API.
Apache-2.0, otwarta: można używać i zmieniać, z zachowaniem licencji i informacji o zmianach.
- 6 skryptów w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: scripts/ace-server.sh, scripts/arrange-loop.py, scripts/brief.mjs, scripts/extract-loop.py, scripts/plan-codes.py, scripts/render-takes.mjs
Co robi i kiedy
Renderuje gotową pętlę dźwiękową przy lokalnej instalacji ACE-Step z zatwierdzonego briefu i eksportuje ją do pliku WAV w projekcie.
Argumenty
/loop-produce Brief slug to render, e.g. "turntable-hip-hop", optionally "draft" or extra seeds
Oryginalny opis (EN)
Render a pro-quality seamless genre loop with the local ACE-Step 1.5 install from an approved brief (docs/loops/briefs/<slug>.json): server lifecycle, two-phase 4B-LM planning + XL SFT DiT rendering at 80 steps, take screening, grid-snapped loop extraction and export to public/loops/<slug>.wav. Use when generating, re-rendering, retaking or drafting an audio loop for a genre, or when ACE-Step renders fail, run out of memory or use the wrong model.
Jak działa
- Environment
- Procedure (final quality)
- Draft profile (optional, brief iteration only)
- Source-conditioned arrangement profile
- Rules
Podgląd SKILL.md
---
name: loop-produce
description: 'Render a pro-quality seamless genre loop with the local ACE-Step 1.5 install from an approved brief (docs/loops/briefs/<slug>.json): server lifecycle, two-phase 4B-LM planning + XL SFT DiT rendering at 80 steps, take screening, grid-snapped loop extraction and export to public/loops/<slug>.wav. Use when generating, re-rendering, retaking or drafting an audio loop for a genre, or when ACE-Step renders fail, run out of memory or use the wrong model.'
argument-hint: 'Brief slug to render, e.g. "turntable-hip-hop", optionally "draft" or extra seeds'
---
# Produce a loop with ACE-Step
You turn an approved brief into a shipped WAV loop plus its selection report. You do not write or edit
briefs (`loop-brief`), set quality thresholds (`loop-validate`) or decide whether it ships (`loop-review`).
Read [references/ace-step.md](./references/ace-step.md) once: it holds the verified configuration and
every failure mode met so far.
## Environment
```sh
export ACESTEP_HOME=${ACESTEP_HOME:-$HOME/Tools/ACE-Step-1.5} # default install; contains checkpoints/ and .venv/
export LOOP_STAGE_ROOT=$HOME/.cache/soundlines-loops # takes live here, never in the repo or /tmp
PY=$ACESTEP_HOME/.venv/bin/python # has acestep, mlx, librosa, soundfile, matplotlib
S=.github/skills/loop-produce/scripts
STAGE=$LOOP_STAGE_ROOT/<slug>
```
Check `ls $ACESTEP_HOME/checkpoints` lists `acestep-v15-xl-sft`, `acestep-5Hz-lm-4B`, `vae`. If
ACESTEP_HOME is under `/tmp` or `/private/tmp`, warn the user: macOS purges it on reboot (≈45 GB).
## Procedure (final quality)
1. **Lint the brief** — `node .github/skills/loop-brief/scripts/check-brief.mjs docs/loops/briefs/<slug>.json`.
On FAIL stop and hand back to `loop-brief`; never patch the prompt yourself.
2. **Plan codes (LM phase, no server running)** — `$S/ace-server.sh stop`, then
`cd $ACESTEP_HOME && $PY <repo>/$S/plan-codes.py <repo>/docs/loops/briefs/<slug>.json $STAGE`
(~1 min load + ~20 s/seed). Codes are keyed to a hash of prompt/bpm/meter/key/duration; a changed
brief invalidates them automatically.
3. **Render takes (DiT phase)** — `$S/ace-server.sh start final` (≈2 min, verifies the XL model is
loaded), then `node $S/render-takes.mjs docs/loops/briefs/<slug>.json $STAGE final`. ≈7 min per
seed (80 steps × ~5 s on 25 GB M-series); five seeds ≈ 40 min. Run it as an async/background
command and wait for completion; it resumes and skips finished seeds if interrupted. Use a fresh
stage after changing the render request. The renderer rejects stale WAVs anywhere in the takes folder.
4. **Screen takes** — have `loop-validate` run take mode on `$STAGE/takes/*.wav` with `--png-dir $STAGE/png`.
Move FAIL takes to `$STAGE/rejected/` (with the reason) so the extractor ignores them. If fewer
than two takes survive, render extra seeds (`render-takes.mjs … final <seed> …` after planning them).
5. **Extract** — `$PY $S/extract-loop.py docs/loops/briefs/<slug>.json $STAGE/takes <asset> docs/loops/briefs/<slug>.selection.json`
where `<asset>` is the brief's `asset` or `public/loops/<slug>.wav`.
6. **Stop the server** — `$S/ace-server.sh stop` (it holds ~22 GB).
7. **Hand off** to `loop-validate` (loop mode on the asset) and then `loop-review`, passing `$STAGE`.
## Draft profile (optional, brief iteration only)
`$S/ace-server.sh start draft` then `render-takes.mjs … draft` renders turbo 8-step takes with the
in-server 1.7B planner in ~30 s each into `$STAGE/drafts/`. Use drafts only to hear whether a new brief
is in the right territory; never extract a shipped asset from drafts. No `plan-codes.py` needed.
## Source-conditioned arrangement profile
For another arrangement of an original Soundlines loop, the brief includes
`arrangement: { source: "public/loops/<slug>.wav", strength: 0.75 }`.
Use `arrange-loop.py prepare <brief> <stage>` to repeat the original loop for source conditioning,
then `render-takes.mjs <brief> <stage> arrangement <seed ...>` with the final XL server.
Cover mode uses source-audio conditioning and automatically bypasses LM planning; never claim
4B-planned-code provenance for this profile. Keep XL SFT, 80 steps, guidance 7, ADG and shift 1.
After take screening, use `arrange-loop.py extract <brief> <stage> <asset> <report>`.
Extraction compares only source-aligned full cycles and preserves the exact source sample count.
The report records source checksum, source frame count and time-aligned chroma comparison.
Run the unchanged loop gate and compare both versions by ear before approval: matching tempo/key
or a high chroma score alone does not prove the melody is preserved.
## Rules
- Never ship from `acestep-v15-turbo` or `acestep-v15-base`; both shipped profiles use XL SFT,
80 steps, guidance 7, ADG, shift 1. Text-to-music requires 4B LM codes; source-conditioned
arrangements require recorded original-source provenance. Do not lower steps to save time.
- Do not run the LM and XL DiT in the same process or with the server up during planning (MPS OOM).
- Keep the loop WAV (PCM16). Never transcode loops to AAC/MP3: encoder padding breaks the seamless wrap.
- Only the asset and `<slug>.selection.json` enter the repo; takes, codes, PNGs and logs stay in `$STAGE`.
- One server, one job at a time (`workers=1`, in-memory queue); never start a second instance.
- Stop processes by the PID file (`ace-server.sh stop`), never by name.
Pliki w paczce (8)
- SKILL.md 5 KB
- references/ace-step.md 6 KB
- scripts/ace-server.sh 3 KB
- scripts/arrange-loop.py 6 KB
- scripts/brief.mjs 1 KB
- scripts/extract-loop.py 5 KB
- scripts/plan-codes.py 3 KB
- scripts/render-takes.mjs 7 KB
Wersje
W bazie nie ma innych skilli o nazwie loop-produce.
Migawka bazy z 13.11.2026 · wersja paczki a5b203cf2dbb (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100loop-brief
Pisze i poprawia brief produkcyjny pętli dźwiękowej dla gatunku muzycznego i sprawdza go lintem przed renderowaniem. - 100ace-step
Generuje, przerabia i rozszerza muzykę modelem ACE Step przez CLI RunComfy, z tagami stylu i tekstami w wielu językach. - 95neuraldeep-music
Generuje muzykę z tekstu modelem ACE-Step w NeuralDeep i pobiera pliki WAV, z zatwierdzeniem kosztu każdego utworu. - 100audio-loop-editing
Tworzy płynne pętle audio i sprawdza granice montażu oraz przesunięcie fazy. - 90ce-produce
Tworzy gotowy do publikacji szkic z zatwierdzonego briefu treści w sześcioetapowym potoku, na polecenie napisania posta lub wersji roboczej briefu. - 87ai-music
Generuje muzykę AI przez CLI runcomfy, wybierając model: ElevenLabs dla wokalu w wysokiej jakości albo ACE Step dla tańszych kompozycji.