Music with ComfyUI

Call a user-defined ComfyUI server (local or remote) to generate music / audio with AceStep Audio 1.5 (text + optional lyrics → MP3). The user may specify any of theme, music length, style, and lyrics (one or more). If no style or lyrics are given explicitly, the agent derives them from the user's request — and from any image the user provides (see Image → Music).

• T2M (Text → Music) → AceStep Audio 1.5 (the only model here). A music description (style/genre tags) plus optional lyrics → an MP3.

The workflow expects several model files on the server (see Models and MODEL_URL.txt):

• Diffusion model: acestep_v1.5_xl_turbo_bf16.safetensors
• CLIP (dual): qwen_0.6b_ace15.safetensors + qwen_4b_ace15.safetensors
• VAE: ace_1.5_vae.safetensors

config.json at the skill root holds every default; any value can be overridden by an env var (see the Data & Privacy table there).

Image → Music — Deriving Style and Lyrics

The agent reads the image; the workflow reads only text. AceStep 1.5's graph has no image input — the sole text node is TextEncodeAceStepAudio1.5 (tags + lyrics). When the user sends an image and does not also specify a full style prompt and/or lyrics, let the agent derive the tags and lyrics, then hand only that text to the script:

• Use the image to infer a suitable style / mood / genre for --tags (and, if the song has words, draft lyrics in the right --language).
• Pass the user's theme/genre on top of what the image implies; never fabricate constraints the user didn't ask for.
• No image is ever uploaded. The reference image is used for vibe extraction only — it is never uploaded to, or sent to, any external destination.

If the user does give explicit style and/or lyrics, honour those first and use the image only as a mood reference.

When to use this section:

• Image arrives → follow the rules in references/image-first.md
• User explicitly asks for a style/genre, or writes lyrics → use them verbatim (cleaned up as needed)

When to Use (Recognizing the Request)

This skill fires whenever the user asks to write, compose, arrange, or make a song / melody / music track, in any language — even if no genre or length is given. Common trigger phrases include, but are not limited to:

• English: "make me a song", "make up a song", "compose", "compose a song/track", "write a song", "arrange a song", "write some music", "make a melody", "produce a track", "write me a tune"
• 中文: "写歌", "作曲", "编一首歌", "编一个歌", "编个曲", "编曲", "做首歌", "帮写首歌", "生成一首歌", "来首原创"
• 其他语言: Japanese (曲を作る、歌を歌う、作曲する), Korean (곡 만들기, 작곡), Spanish (hacer una canción, componer, escribir una canción), etc.

The signal is the intent to create new music (write / compose / arrange / make / generate a song/melody/track). Match it loosely across languages; if the intent is clearly "create a new song", use this skill.

• Generate music from a description / theme → music (--tags)
• The user sends an image, then asks for a song (within reason) → image-first mode, see references/image-first.md

When NOT to use this skill

This skill is for generation, not post-processing of existing audio. Do not use it (and do not look here) for:

• Converting / transcribing / resampling / normalizing an audio file the user already has (that is plain audio processing — use an audio tool).
• Adjusting volume, trimming, splicing, or remixing an existing file.
• Adding lyrics/subtitles to a pre-existing recording (unless the user wants a new song sung).

Data & Privacy (read before running)

• What is transmitted: the music tags/lyrics and the workflow JSON go to the configured endpoint for every run. No image is ever sent — AceStep 1.5 has no image input; if the user provides one, the agent visual-reads it and turns it into tags/lyrics before calling the script.
• Endpoint is user-defined (COMFYUI_URL env var > comfyui_url in config.json; bundled default http://127.0.0.1:8188). A non-local host prints a data-flow warning on every run; if no server answers, the script exits with configuration instructions instead of a raw timeout.
• Written to disk: generated audio is saved under output_dir (config.json or COMFYUI_OUTPUT_DIR). Don't submit sensitive content if the endpoint or disk is a concern.
• Secrets: never put credentials in the URL or logs.

Workflow Files

The workflow is an adapted version of a real ComfyUI graph. Original source + exactly what was changed → references/attribution.md.

CLI Usage

⚠️ The music description is MANDATORY and must be passed via --tags (or --prompt) — never as a positional word.

python3 music_with_comfyui.py music "lo-fi track" → music: error: the following arguments are required: --tags

Always write the description through --tags "..." on the first attempt. A positional word wastes a whole turn before the command ever reaches ComfyUI.

Full commands and the timeout table → references/cli-reference.md

What each knob controls

• --tags (required) — the music description: genre, tempo, instrumentation, mood, vocal style. This is the primary prompt. ComfyUI-escaped (line breaks as \t/\n); keep it clean.
• --lyrics (optional) — structured Sonic-lyrics text. Omit (or pass empty) for an instrumental; AceStep 1.5 generates instrumental audio.
• --duration (optional) — seconds (default: config default_duration, 120). The audio latent's seconds is set to match, so keep them consistent.
• --seed (optional) — not sent by default. If omitted, the script uses no seed (random per run). Pass --seed <n> only to lock/reproduce a specific output.
• Style/tuning flags (--bpm, --keyscale, --language, --timesignature, --cfg-scale, --temperature, --top-p, --top-k, --min-p) — optional overrides of the defaults in config.json music.default_params; each omits cleanly if unset.
• --tags-image (removed) — AceStep 1.5 takes no image. If the user gives a reference image, the agent reads it and derives tags/lyrics itself (see Image → Music).
• --format / --output-dir — output format (jpg/png for the image skill; here audio format is whatever the server's SaveAudioAdvanced writes).

Prompt / Tags Formatting

Guidance for crafting AceStep prompts lives in references/prompts.md. Key points:

• Write --tags as a comma/space-separated description of genre, tempo, instrumentation, mood, and vocal style. Keep it clean and natural.
• ComfyUI-escape any line breaks in the description: \\t (tab) / \\n (newline). If you write the tags through --tags "..." the script passes the string through, so keep line breaks to a minimum or escape them.
• Lyrics, if any, should be structured (sections like [Verse 1], [Chorus], ...). Keep to the language set via --language.
• If the agent picks the style itself, it should still produce a concise, coherent tag line — not a wall of contradictory keywords.

Error Handling

The script auto-detects the same categories as the image skill and reports them with install/source guidance (see references/error-handling.md):

1. Missing nodes — reports which node, its package, and a pinned GitHub install hint (the script never runs git clone itself).
2. Missing models — reports the missing model file(s); the AceStep workflow has no blind model substitution (there is only one model), so it aborts with the install hint instead of degrading.
3. Non-critical utility nodes (e.g. UnloadAllModels) — bypassed automatically without interrupting generation.

Output Delivery

Send the generated file — do not just describe it. Deliver in the user's original session, never raw paths/URLs. Per-channel prefixes and the staging/cleanup routine → references/delivery.md

• Generated audio is saved under output_dir/music/ (prefix amusic_<timestamp>, e.g. amusic_20261006_153000.mp3).
• Attach the file — send it as an audio attachment in the user's session (e.g. a MEDIA: line for WhatsApp, a native attachment elsewhere). Do not send the raw file path to the group unless asked.
• WhatsApp → transcode to 44.1kHz. Always. AceStep 1.5 writes 48 kHz MP3, which does not play on WhatsApp. Before delivering on WhatsApp, transcode to 44.1 kHz with ffmpeg first: ffmpeg -y -i <src> -c:a aac -b:a 128k -ar 44100 -ac 1 -movflags +faststart <out>.m4a then send the .m4a via MEDIA: as a plain line. (mp3 also works if 44.1 kHz; 44.1 kHz is the key.)
• Explicit original-request: if the user asks for the original/unmodified file, do NOT transcode. Instead package the original audio as a .zip (e.g. zip -j out/<name>.zip <src>.mp3) and send the zip as a document attachment (MEDIA:<path>.zip, or message(action="send", channel=..., asDocument=true, media=<path>.zip)).
• Never send raw local file paths, ComfyUI URLs, or host info unless asked.

Delivering the Result — Title + Lyrics

A generated song is not just an audio clip: hand it over as a complete package, and do it all in one message.

• Song title. If the user did not request a specific title, give the song one. Derive it from the track's mood, the --tags, and the lyrics — keep it short and evocative. State it clearly as the song's name.
• Lyrics. If the song has lyrics, return them to the user, cleanly formatted (each section on its own line, section headers such as [Verse] / [Chorus] intact, no stray control characters or unescaped line breaks).
• The audio file, attached in the user's session (channel prefixes in references/delivery.md).

Deliver all three together — title, lyrics, and audio — not in separate hand-offs:

🎵  **<Song Title>**
<style note — one line>

[Verse 1]
...

[Chorus]
...

MEDIA:/…/amusic_<timestamp>.mp3

Never omit the title or the lyrics: a finished song should arrive with its name and, when it has words, its lyrics, in the same message as the audio. If the user supplied a title, use theirs; if the song is instrumental, skip the lyrics but still give it a title.

Execution Discipline — Fast Path

The measured GPU sampling time is model-dependent (short, looped clips are fast; long durations cost seconds-to-tens-of-seconds of sampling). When a run stretches well beyond the configured timeout_seconds, the delay is almost always agent-side overhead before the run, not inference. Kill the waste:

1. Skip re-reads on repeat. For a known config / repeat / same-job, the defaults are already baked in — no need to re-open SKILL.md, config.json, or the script.
2. At most one pre-check merge. Gather any needed info (requested style/lyrics) in a single batched command, or skip it for a trivial repeat.
3. Run immediately after one check. Don't inspect, second-guess, or add tool calls between knowing the tags and running.
4. Verify only the minimum. After the run, check only what the user asked (duration / format / file exists).
5. Golden rule: sampling is bounded by timeout_seconds in config; anything past that is overhead. Never pile on tool calls before the audio exists.

References (index)

• references/image-first.md — image-then-music detection rules
• references/attribution.md — original source + modifications per workflow
• references/cli-reference.md — full commands + timeout reference
• references/prompts.md — per-model prompt / tags guidance
• references/error-handling.md — auto-handled error categories
• references/delivery.md — per-channel delivery + staging