LTX 2.3 Trainer — Audio+Video Reference IC-LoRA (/ic-lora/av2av)
Overview
The /ic-lora/av2av endpoint trains an IC-LoRA (In-Context LoRA) for the LTX 2.3 model that performs a combined audio+video → audio+video transformation. An IC-LoRA is a small adapter that does not generate from text alone — instead it conditions on a reference clip supplied at inference (both its video and its audio track) and learns to produce a corresponding target clip with both generated video and generated audio. You teach it that joint mapping by providing pairs of clips — a reference and a target, each carrying audio — and the LoRA learns to go from one to the other across both modalities at once.
Use this endpoint when you want a control-driven transformation across both modalities together, for example:
- Restyling a video and its sound at the same time.
- Applying a recurring audiovisual edit that touches both picture and audio.
- Any reference-to-(audio+video) mapping you can demonstrate with paired examples.
Key features:
- Trains an IC-LoRA from paired reference→target clips, each with audio.
- Both video and audio are generated, each conditioned on its own reference.
- Optional reference video downscaling / temporal downsampling so the LoRA can be driven by a coarse / low-FPS video control proxy.
- Validation previews transform a supplied reference clip (video + audio) into a combined video preview that carries the generated audio.
Provide a single .zip archive (linked via training_data_url) of paired clips, each with an audio track:
<name>_start.<ext> — the reference clip (video + audio).<name>_end.<ext> — the target clip (video + audio).<name>.txt — optional caption.
Video formats: .mp4, .mov, .webm, .mkv, .avi. Both the reference and the target of every pair must contain an audio track — silent clips are rejected with a clear error. Each _start must have a matching _end. File names must be unique across the archive. Aim for at least 10 pairs.
Minimum clip length: with auto_scale_input off (the default), both clips in each pair (_start and _end) must already have at least number_of_frames frames (default 89, ≈ 3.7 s at 24 fps). A pair with either side too short is skipped, and if no pair qualifies the request fails (422). Turn on auto_scale_input to resample shorter clips instead.
Example layout:
sample01_start.mp4 sample01_end.mp4 sample01.txt
sample02_start.mp4 sample02_end.mp4 sample02.txt
Dataset
training_data_url (required)
Type: string
URL to the .zip archive of _start/_end clip pairs (with audio).
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
IC-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.
reference_downscale_factor
Type: integer
Default: 1 (range 1–8)
Spatially downscale the reference video by this factor before encoding (coarse / low-resolution control proxy → full-resolution output). 1 means no downscaling. Both width and height must be divisible by the factor, and width ÷ factor and height ÷ factor must each be divisible by 32 (checked against both the training and validation resolutions); an incompatible value fails the request with a 422. (Applies to the video reference only; audio has no spatial knob.)
reference_temporal_scale_factor
Type: integer
Default: 1 (range 1–8)
Temporally downsample the reference video (lower FPS) by this factor before encoding. 1 means no change. (number_of_frames − 1) must be divisible by the factor, and after subsampling (frames − 1) must remain a multiple of 8 — checked against both number_of_frames and validation_number_of_frames. An incompatible factor/frame-count combination fails the request with a 422. (Example: with the default 89 frames, a factor of 2 is invalid because (89 − 1) ÷ 2 = 44, which is not a multiple of 8.)
Video Configuration
number_of_frames
Type: integer
Default: 89 (range 9–121)
Frames per training clip. Must satisfy frames % 8 == 1; other values are snapped down to the nearest valid count.
frame_rate
Type: integer
Default: 24 (range 8–60)
Target frames per second.
resolution
Type: string (low, medium, high)
Default: medium
| Resolution | 16:9 | 1:1 | 9:16 |
|---|
| low | 512×288 | 512×512 | 288×512 |
| medium | 768×448 | 768×768 | 448×768 |
| high | 960×544 | 960×960 | 544×960 |
aspect_ratio
Type: string (16:9, 1:1, 9:16)
Default: 1:1
Type: boolean
Default: false
Fit videos to the target frame count and frame rate.
Type: boolean
Default: false
Off for this mode: scene splitting would desync the reference/target pair. Provide pre-split clips instead.
Type: number
Default: 30.0 (range 1.0–60.0)
Duration threshold for scene splitting (only relevant if splitting were enabled).
Audio Configuration
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 video duration.
Validation
validation
Type: array
Default: [] (max 2 entries)
Validation samples, each an object with:
prompt (string) — the text prompt.reference_video_url (string, required) — the reference video (its audio track is also used as the audio reference) that conditions the generated audio+video.reference_audio_url (string, optional) — a separate reference audio; if omitted, the audio is taken from the reference video's own track — in which case that reference video must contain an audio track, or the validation sample is rejected with a 422.
validation_negative_prompt
Type: string
Default: a built-in quality negative prompt.
validation_number_of_frames
Type: integer
Default: 89 (range 9–121)
validation_frame_rate
Type: integer
Default: 24 (range 8–60)
validation_resolution
Type: string
Default: high
validation_aspect_ratio
Type: string
Default: 1:1
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.
Outputs
lora_file — the trained IC-LoRA weights (.safetensors).config_file — JSON describing the trigger phrase and training type.video — combined validation reel; the preview video carries the generated audio track.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,
_start/_end pairs and captions matched, both clips verified to have audio, clips fit to the resolution bucket, and the video reference optionally downscaled / temporally downsampled. The target's audio is extracted from its own track; the reference's audio is extracted from the reference clip.
- Training — the IC-LoRA trains for
number_of_steps. Both modalities are generated: the video conditioned on the reference video, the audio conditioned on the reference audio. Validation previews run at intervals.
- Output — the IC-LoRA, config, and validation reel (with generated audio) are uploaded.
What Happens to Your Data
- Archive extraction: the
.zip is unpacked; macOS metadata and hidden files are ignored.
- Pair matching: each
<name>_start is paired with its <name>_end and the optional <name>.txt. Targets missing a reference cause a clear error.
- Audio requirement: both clips of every pair must carry audio; silent pairs are rejected.
- Video fitting: both clips are resized to fill the resolution bucket and center-cropped, staying aligned.
- Reference scaling: when set above 1, the reference video is downscaled / temporally downsampled before encoding so the LoRA learns to drive full-resolution output from a coarse reference.
- Captions: the trigger phrase (if set) is prepended.
What an IC-LoRA Is
An IC-LoRA performs an in-context transformation: it conditions on a reference clip supplied at inference and produces a transformed result — here, both video and audio together. The trained file specializes the base model at your reference→target audiovisual mapping. At inference you supply a reference video (and optionally a separate reference audio) plus a prompt, and the LoRA produces the corresponding target clip with both generated video and audio.
Tips for Getting Good Results
Dataset Quality
- Use at least 10 well-aligned reference→target pairs, each with clean, in-sync audio.
- Reference and target should describe the same scene/sound differing only by the transformation you want learned.
Caption Best Practices
- Describe the target's content and sound plainly, optionally with a trigger phrase.
- Keep captions consistent in style across pairs.
Good caption: av_st9le a neon-lit street with a synthwave soundtrack
Weak caption: street
Trigger Phrases
- A distinctive trigger phrase helps invoke the transformation cleanly; include it in every caption and at inference.
Reference Scaling
- To drive generation from a coarse video control map, set
reference_downscale_factor above 1 and supply matching coarse references at inference; the validation localizer mirrors training-time scaling for you.
At inference, supply a reference video (and optionally a separate reference audio), use the same trigger phrase, and apply the same reference scaling you trained with.
Recommended Starting Configuration
{
"training_data_url": "https://example.com/av2av_pairs.zip",
"trigger_phrase": "av_st9le",
"rank": 32,
"number_of_steps": 2000,
"learning_rate": 0.0002,
"reference_downscale_factor": 1,
"reference_temporal_scale_factor": 1,
"number_of_frames": 89,
"frame_rate": 24,
"resolution": "medium",
"aspect_ratio": "1:1",
"validation": [
{ "prompt": "av_st9le a neon-lit street", "reference_video_url": "https://example.com/ref.mp4" }
]
}
Diagnosing Issues
- Transformation too weak: more steps, higher
rank, or more consistent pairs.
- Audio or video ignores the reference: ensure pairs are well aligned and differ only by the intended transformation; check reference scaling matches at inference.
- Dataset rejected for silence: both clips of every pair must contain audio.
- Overfitting: fewer steps, lower
rank, more pairs.
Validation Prompt Tips
- Use a fresh reference clip to gauge generalization.
- Match the caption style and trigger phrase you trained with.
Common Pitfalls
- Silent reference or target clips (rejected).
- Missing or mismatched
_start/_end pairs.
- Using different reference scaling at inference than at training.
fal-ai/ltx23-trainer-v2/ic-lora/av2av