LTX 2.3 Trainer — Backward Audio Extension (/audio-extend-suffix)
Overview
The /audio-extend-suffix endpoint trains a LoRA for the LTX 2.3 model that generates the lead-in to an audio clip — it extends audio backward in time. During training, the last few seconds of each clip are kept as a clean "suffix" and the model learns to generate the audio leading up to them. At inference you supply a closing audio clip and the model produces the preceding section. This is audio-only.
Key features:
- Learns to extend audio backward, generating a plausible lead-in to a clean closing (suffix) window.
- Audio-only — no video is processed or generated.
- The suffix is carved from each clip's own audio automatically.
- Fixed audio length bucket via
audio_duration_seconds.
- Validation previews extend a supplied clip backward; the output preview is audio.
Provide a single .zip archive (linked via training_data_url) of plain audio clips:
- Audio:
.wav, .mp3, .flac, .ogg, .aac, .m4a
- Captions: a
.txt with the same base name as each audio clip (optional but recommended).
File names must be unique across the archive. Aim for at least 10 clips. Clips shorter than audio_duration_seconds are skipped, so make each clip at least that long — and since conditioning_seconds must be smaller than audio_duration_seconds, every kept clip leaves a lead-in to learn.
Dataset
training_data_url (required)
Type: string
URL to the .zip archive of audio clips.
trigger_phrase
Type: string
Default: ""
Phrase prepended to captions during training; include it at inference.
Training Parameters
rank
Type: integer (8, 16, 32, 64, 128)
Default: 32
LoRA capacity.
number_of_steps
Type: integer
Default: 2000 (range 100–20000)
Number of optimization steps.
learning_rate
Type: number
Default: 0.0002
Optimization step size.
conditioning_seconds
Type: number
Default: 1.0 (range 0.1–30.0)
Seconds of trailing audio kept as the clean suffix the model leads up to. Must be less than audio_duration_seconds so there is a lead-in to generate (and not so close to it that nothing is left after rounding).
Audio Configuration
audio_duration_seconds
Type: number
Default: 5.0 (range 0.5–60.0)
Target audio clip length in seconds (the audio duration bucket). Clips shorter than this are skipped; longer clips are trimmed.
audio_normalize
Type: boolean
Default: true
Peak-normalize audio for consistent loudness across the dataset.
audio_preserve_pitch
Type: boolean
Default: true
Preserve pitch when fitting audio to the target duration.
Validation
validation
Type: array
Default: [] (max 2 entries)
Validation samples, each an object with:
prompt (string) — the text prompt.audio_url (string, required) — an audio clip to extend backward. Its closing conditioning_seconds are used as the suffix.
The validation clip must be at least conditioning_seconds long. When validation prompts are provided, conditioning_seconds is also bounded by the preview length: the validation preview generates an audio span sized from audio_duration_seconds (capped at 121 frames divided by validation_frame_rate), and a conditioning_seconds that would fill that span is rejected at request time (HTTP 422). A very low validation_frame_rate shrinks this preview window, so keep conditioning_seconds comfortably below audio_duration_seconds.
validation_negative_prompt
Type: string
Default: a built-in quality negative prompt.
validation_frame_rate
Type: integer
Default: 24 (range 8–60)
Used together with the audio bucket to size the preview audio length.
stg_scale
Type: number
Default: 1.0 (range 0.0–3.0)
debug_dataset
Type: boolean
Default: false
Return an archive of the preprocessed data for inspection.
Note: this audio-only mode also accepts the shared video/validation-video fields, but they have no effect — no video is processed.
Outputs
lora_file — the trained LoRA weights (.safetensors).config_file — JSON describing the trigger phrase and training type.audio — a combined preview of the generated validation audio.debug_dataset — preprocessed-data archive, only when debug_dataset is enabled.
Billing
A successful run is billed max(100, number_of_steps) billable units. Requests that fail before training completes (input-validation errors / HTTP 422, or dataset-download failures) are billed 0 units.
How the Training Works
Pipeline Overview
- Preprocessing — the archive is extracted, audio matched to captions, and each clip fit to the
audio_duration_seconds bucket.
- Training — for each clip, the closing
conditioning_seconds of audio are held clean as the suffix and the model learns to generate the preceding lead-in. Validation previews run at intervals.
- Output — the LoRA, config, and combined audio preview are uploaded.
What Happens to Your Data
- Archive extraction: the
.zip is unpacked; macOS metadata and hidden files are ignored.
- File matching: each audio clip is paired with the
.txt of the same base name.
- Audio fitting: each clip is fit to the
audio_duration_seconds bucket (shorter clips skipped, longer ones trimmed; normalized and pitch-fit as configured).
- Suffix carving: the closing
conditioning_seconds of each clip are kept clean as conditioning; everything before is the generation target. This happens automatically.
- Captions: the trigger phrase (if set) is prepended.
Tips for Getting Good Results
Dataset Quality
- Use at least 10 clean clips containing the kind of backward continuation (lead-in) you want learned.
- Clips should be meaningfully longer than
conditioning_seconds.
- Pick an
audio_duration_seconds that fits most clips so few are skipped.
Caption Best Practices
- Describe the sound and lead-in plainly, optionally with a trigger phrase.
- Keep captions consistent across clips.
Good caption: a drum fill building up to a cymbal crash
Weak caption: drums
Trigger Phrases
- Use a distinctive trigger phrase to invoke a particular lead-in style; include it in every caption and at inference.
At inference, supply a closing audio clip at least conditioning_seconds long, use the same caption style and trigger phrase, and keep durations in line with audio_duration_seconds.
Recommended Starting Configuration
{
"training_data_url": "https://example.com/audio_clips.zip",
"trigger_phrase": "",
"rank": 32,
"number_of_steps": 2000,
"learning_rate": 0.0002,
"conditioning_seconds": 1.0,
"audio_duration_seconds": 5.0,
"validation": [
{ "prompt": "a drum fill building up", "audio_url": "https://example.com/closing.wav" }
]
}
Diagnosing Issues
- Lead-in does not flow into the suffix: add more representative clips; try a slightly longer
conditioning_seconds for more closing context.
- Validation rejected for short clip: provide a clip at least
conditioning_seconds long, or lower conditioning_seconds.
- Window covers the whole clip: reduce
conditioning_seconds relative to audio_duration_seconds.
- Overfitting: fewer steps, lower
rank, more clips.
Validation Prompt Tips
- Use a fresh closing clip to gauge generalization.
- Keep the prompt aligned with the lead-in you expect.
Common Pitfalls
conditioning_seconds too close to (or above) audio_duration_seconds — nothing left to generate.audio_duration_seconds so long that most clips are skipped.- Background noise polluting the learned lead-in.
fal-ai/ltx23-trainer-v2/audio-extend-suffix