Batch generation

In Library, open Batch generation. You import a UTF-8 JSON file, review a preview (song count, takes, requested vs admitted parallelism), then start. One project is created per song. Pause, cancel, resume, retry failures, and export of ready WAVs are available.

Takes use the same exclusive GPU queue as Create. Admitted capacity is 1: this is not two overlapping inferences. If the file asks for 2, allowReduction (default) starts at 1; requireRequested rejects the batch.

An invalid file creates no projects. After a crash, a running batch becomes “interrupted”; nothing auto-resumes.

Minimal file

UTF-8 encoding. schemaVersion must be 1. Each song needs a unique id, a title, a style, and lyrics.

{
  "schemaVersion": 1,
  "songs": [
    {
      "id": "night",
      "title": "The night moves on",
      "style": "French indie pop, warm bass, soft drums, intimate vocals",
      "lyrics": "[Verse]\nThe night moves across the roofs\nI keep a light for you\n\n[Chorus]\nCome back when day is breaking\nWe still have our dreams"
    }
  ]
}

With no other fields: 1 take, cot = full, 180 s target duration, lyrics sent as written.

Root fields

Field	Required	Default	Rule
`schemaVersion`	yes	—	Integer, only `1`
`songs`	yes	—	Array, 1 to 1000 songs
`name`	no	—	1 to 120 characters
`maxParallelGenerations`	no	`2`	1 to 32. Admitted capacity is still 1
`parallelismPolicy`	no	`allowReduction`	`allowReduction` (start at 1) or `requireRequested` (block if demand > 1)
`onError`	no	`continue`	`continue` (next song) or `pause`
`retry.maxAttempts`	no	`1`	1 to 5
`defaults`	no	—	Shared song options (below)

Any unknown property is rejected. Duplicate JSON keys too. Max file size: 10 MB. At most 10,000 takes in total.

Each song (`songs[]`)

Field	Required	Limits
`id`	yes	1 to 64 characters, unique in the file
`title`	yes	1 to 120. No trailing period. Forbidden: `/ \ : * ? " < > \|`
`style`	yes	1 to 4000 characters
`lyrics`	yes	1 to 4000, unless `instrumentalMode` is `true` (then `""` is allowed)
`seed`	no	Integer 0 to 4,294,967,295. If omitted, a seed is written
`generations`	no	1 to 1000 takes for this song
`cot`	no	`full`, `melody`, or `off`
`targetDurationSec`	no	30 to 360 (a target, not a guaranteed musical length)
`preferFullLyrics`	no	boolean
`instrumentalMode`	no	`true`: lyrics optional; this is not the YuE2 instrumental LoRA
`singingLanguage`	no	string or `null` (1 to 40 characters when set)
`tempoBpm`	no	integer 40 to 220, or `null`
`key`	no	`{ "tonic": "C", "mode": "minor" }` or `null`
`meter`	no	`{ "numerator": 4, "denominator": 4 }` or `null`

tonic: C, C#, D, Eb, E, F, F#, G, Ab, A, Bb, B.
mode: major or minor.
Meter: 4/4, 3/4, 6/8, or 2/4 only.

`defaults` and precedence

defaults uses the optional song fields (not id, title, style, lyrics, seed). For each option: song value, else defaults, else the product default:

Option	Product default
`generations`	`1`
`cot`	`full`
`targetDurationSec`	`180`
`preferFullLyrics`	`true`
`instrumentalMode`	`false`
`singingLanguage`, `tempoBpm`, `key`, `meter`	unset

Example with options

Two songs, 3 takes then 2, policy that accepts reduction to 1:

{
  "schemaVersion": 1,
  "name": "Demo — two songs",
  "maxParallelGenerations": 2,
  "parallelismPolicy": "allowReduction",
  "onError": "continue",
  "retry": { "maxAttempts": 1 },
  "defaults": {
    "generations": 3,
    "cot": "full",
    "singingLanguage": "en",
    "targetDurationSec": 180,
    "preferFullLyrics": true,
    "instrumentalMode": false
  },
  "songs": [
    {
      "id": "night",
      "title": "The night moves on",
      "style": "French indie pop, warm bass, soft drums, intimate vocals",
      "lyrics": "[Verse]\nThe night moves across the roofs\nI keep a light for you\n\n[Chorus]\nCome back when day is breaking",
      "tempoBpm": 110,
      "seed": 42
    },
    {
      "id": "road",
      "title": "On the road",
      "style": "French rock, electric guitars, energetic drums, clear vocals",
      "lyrics": "[Verse]\nThe road unfolds its turns\n\n[Chorus]\nOn the road we keep going",
      "generations": 2,
      "targetDurationSec": 120,
      "tempoBpm": 140
    }
  ]
}

In the app, Batch generation can also download a sample file. The engine is the profile’s engine (YuE2 by default), not a second XOR.

What batch does not do

No overlapping GPU inferences.
No YuE2 audio_input and no Lego mix/stems fields in the JSON (Lego runs from Production on an open song).
Not a replacement for sequential candidates on the Create tab (N calls for one song).