lightricks/ltx23-video-trainer
Ltx23 Video Trainer by Lightricks - advanced AI model for training. Delivers high-quality results with fast inference, suitable for both creative and production workflows.
Input
The aspect ratio of the generated image. Allowed values: 21:9, 16:9, 3:2, 4:3, 5:4, 1:1, 4:5, 3:4, 2:3, 9:16.
Resolution to use for training. Higher resolutions require more memory. Allowed values: low, medium, high.
0.0000011
Learning rate for optimization. Higher values can lead to faster training but may cause overfitting. Range: 0.000001 to 1.
860
Target frames per second for the video. Range: 8 to 60.
A list of validation prompts to use during training. When providing an image, _all_ validation inputs must have an image.
The rank of the LoRA adaptation. Higher values increase capacity but use more memory. Allowed values: 8, 16, 32, 64, 128.
When audio duration doesn't match video duration, stretch/compress audio without changing pitch. If disabled, audio is trimmed or padded with silence.
9121
Number of frames per training sample. Must satisfy frames % 8 == 1 (e.g., 1, 9, 17, 25, 33, 41, 49, 57, 65, 73, 81, 89, 97). Range: 9 to 121.
10020000
The number of training steps. Range: 100 to 20000.
The aspect ratio to use for validation. Allowed values: 16:9, 1:1, 9:16.
Enable joint audio-video training. If None (default), automatically detects whether input videos have audio. Set to True to force audio training, or False to disable.
03
STG (Spatio-Temporal Guidance) scale. 0.0 disables STG. Recommended value is 1.0. Range: 0 to 3.
160
The duration threshold in seconds. If a video is longer than this, it will be split into scenes. Range: 1 to 60.
If true, videos above a certain duration threshold will be split into scenes.
9121
The number of frames in validation videos. Range: 9 to 121.
A negative prompt to use for validation.
860
Target frames per second for validation videos. Range: 8 to 60.
Whether to generate audio in validation samples.
PNG, JPEG, WebP, or GIF · 20 MiB maximum
URL to zip archive with videos or images. Try to use at least 10 files, although more is better.
**Supported video formats:** .mp4, .mov, .avi, .mkv
**Supported image formats:** .png, .jpg, .jpeg
Note: The dataset must contain ONLY videos OR ONLY images - mixed datasets are not supported.
The archive can also contain text files with captions. Each text file should have the same name as the media file it corresponds to.
01
Probability of conditioning on the first frame during training. Higher values improve image-to-video performance. Range: 0 to 1.
Normalize audio peak amplitude to a consistent level. Recommended for consistent audio levels across the dataset.
A phrase that will trigger the LoRA style. Will be prepended to captions during training.
When enabled, the trainer returns a downloadable archive of your preprocessed training data for manual inspection. Use this to verify that your videos, images, and captions were processed correctly before committing to a full training run.
If true, videos will be automatically scaled to the target frame count and fps. This option has no effect on image datasets.
The resolution to use for validation. Allowed values: low, medium, high.
Idle
Example output — click Run to generate your own
Production endpoint
Send your first request
OpenAI-compatible endpoint with unified authentication and usage tracking.
POST
https://api.sandbase.ai/v1/runModel ID
lightricks/ltx23-video-trainer01
Input Schema
25 parameters · 0 required · 25 optional
| Parameter | Type | Required | Description |
|---|---|---|---|
rank | integer | Optional | The rank of the LoRA adaptation. Higher values increase capacity but use more memory. · Options: 8, 16, 32, 64, 128 · Default: 32 8163264128 |
stg_scale | number | Optional | STG (Spatio-Temporal Guidance) scale. 0.0 disables STG. Recommended value is 1.0. · Min: 0 · Max: 3 · Default: 1 |
frame_rate | integer | Optional | Target frames per second for the video. · Min: 8 · Max: 60 · Default: 25 |
resolution | string | Optional | Resolution to use for training. Higher resolutions require more memory. · Options: low, medium, high · Default: "medium" lowmediumhigh |
validation | object[] | Optional | A list of validation prompts to use during training. When providing an image, _all_ validation inputs must have an image. · Default: [] |
with_audio | boolean | Optional | Enable joint audio-video training. If None (default), automatically detects whether input videos have audio. Set to True to force audio training, or False to disable. |
aspect_ratio | string | Optional | The aspect ratio of the generated image. · Options: 21:9, 16:9, 3:2, 4:3, 5:4, 1:1, 4:5, 3:4, 2:3, 9:16 21:916:93:24:35:41:14:53:42:39:16 |
debug_dataset | boolean | Optional | When enabled, the trainer returns a downloadable archive of your preprocessed training data for manual inspection. Use this to verify that your videos, images, and captions were processed correctly before committing to a full training run. · Default: false |
learning_rate | number | Optional | Learning rate for optimization. Higher values can lead to faster training but may cause overfitting. · Min: 0.000001 · Max: 1 · Default: 0.0002 |
trigger_phrase | string | Optional | A phrase that will trigger the LoRA style. Will be prepended to captions during training. · Default: "" |
audio_normalize | boolean | Optional | Normalize audio peak amplitude to a consistent level. Recommended for consistent audio levels across the dataset. · Default: true |
number_of_steps | integer | Optional | The number of training steps. · Min: 100 · Max: 20000 · Default: 2000 |
auto_scale_input | boolean | Optional | If true, videos will be automatically scaled to the target frame count and fps. This option has no effect on image datasets. · Default: false |
number_of_frames | integer | Optional | Number of frames per training sample. Must satisfy frames % 8 == 1 (e.g., 1, 9, 17, 25, 33, 41, 49, 57, 65, 73, 81, 89, 97). · Min: 9 · Max: 121 · Default: 89 |
training_data_url | string | Optional | URL to zip archive with videos or images. Try to use at least 10 files, although more is better. **Supported video formats:** .mp4, .mov, .avi, .mkv **Supported image formats:** .png, .jpg, .jpeg Note: The dataset must contain ONLY videos OR ONLY images - mixed datasets are not supported. The archive can also contain text files with captions. Each text file should have the same name as the media file it corresponds to. |
audio_preserve_pitch | boolean | Optional | When audio duration doesn't match video duration, stretch/compress audio without changing pitch. If disabled, audio is trimmed or padded with silence. · Default: true |
validation_frame_rate | integer | Optional | Target frames per second for validation videos. · Min: 8 · Max: 60 · Default: 25 |
validation_resolution | string | Optional | The resolution to use for validation. · Options: low, medium, high · Default: "high" lowmediumhigh |
split_input_into_scenes | boolean | Optional | If true, videos above a certain duration threshold will be split into scenes. · Default: true |
validation_aspect_ratio | string | Optional | The aspect ratio to use for validation. · Options: 16:9, 1:1, 9:16 · Default: "1:1" 16:91:19:16 |
first_frame_conditioning_p | number | Optional | Probability of conditioning on the first frame during training. Higher values improve image-to-video performance. · Min: 0 · Max: 1 · Default: 0.5 |
validation_negative_prompt | string | Optional | A negative prompt to use for validation. · Default: "worst quality, inconsistent motion, blurry, jittery, distorted" |
validation_number_of_frames | integer | Optional | The number of frames in validation videos. · Min: 9 · Max: 121 · Default: 89 |
generate_audio_in_validation | boolean | Optional | Whether to generate audio in validation samples. · Default: true |
split_input_duration_threshold | number | Optional | The duration threshold in seconds. If a video is longer than this, it will be split into scenes. · Min: 1 · Max: 60 · Default: 30 |
02
Output Schema
| Field | Type | Description |
|---|---|---|
id | string | Unique identifier for the generation task |
status | string | Task status: pending, running, completed, failed, timeout |
model | string | Model used for the generation |
outputs | array | Array of output items |
outputs[].url | string | URL of the generated artifact |
outputs[].content_type | string | MIME type (e.g. image/png, video/mp4) |
error | object | null | Error details if failed, null on success |
error.type | string | Machine-readable error type code |
error.message | string | Human-readable error description |
Async Workflow
This model uses asynchronous execution. Submit a request and poll for the result.
- Submit — POST to /v1/run, receive an
id - Poll — GET /v1/run/{id} until status is
completed,failed, ortimeout - Retrieve — Read
outputsfrom the completed response
03
Code Examples
Ready-to-run snippets
# Step 1: Submit
curl -X POST https://api.sandbase.ai/v1/run \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "lightricks/ltx23-video-trainer",
"rank": 32,
"stg_scale": 1,
"frame_rate": 25,
"resolution": "medium",
"validation": [],
"debug_dataset": false,
"learning_rate": 0.0002,
"trigger_phrase": "",
"audio_normalize": true,
"number_of_steps": 2000,
"auto_scale_input": false,
"number_of_frames": 89,
"audio_preserve_pitch": true,
"validation_frame_rate": 25,
"validation_resolution": "high",
"split_input_into_scenes": true,
"validation_aspect_ratio": "1:1",
"first_frame_conditioning_p": 0.5,
"validation_negative_prompt": "worst quality, inconsistent motion, blurry, jittery, distorted",
"validation_number_of_frames": 89,
"generate_audio_in_validation": true,
"split_input_duration_threshold": 30,
"prompt": "a beautiful sunset over mountains"
}'
# Step 2: Poll result (replace <id>)
curl https://api.sandbase.ai/v1/run/<id> \
-H "Authorization: Bearer YOUR_API_KEY"API README
LTX-2.3 22B Video Trainer
LTX-2.3 22B Video Trainer is a LTX route built for custom video-model training. It turns a curated training set and training controls into a reusable personalized video model, giving creators a focused endpoint instead of forcing one generic workflow across materially different production tasks. The route belongs to a family known for audio-visual timing, cinematic motion, and shot-level controllability, so it is best evaluated as a creative system for intentional shots and assets rather than as a one-click novelty generator.
Use this endpoint when the input contract and deliverable match that job exactly. Its documented controls include rank choices (8, 16, 32, 64, 128); stg_scale (STG (Spatio-Temporal Guidance) scale. 0.0 disables STG. Recommended value is 1.0); frame_rate (Target frames per second for the video); resolution choices (low, medium, high). Together, these controls help teams plan predictable iterations, compare outputs under stable settings, and connect generation to iterative film, advertising, and social-video workflows without hiding the operational choices that shape the result.
Highlights
- LTX-2.3 22B Video Trainer: dedicated custom model training. This exact route centers on custom model training; it uses the schema-defined source fields to separate its job from neighboring endpoints in the same family.
rankcontrol. options8,16,32,64,128; this setting provides a concrete lever for route-specific output planning.stg_scalecontrol. range0to3; use it to build controlled comparisons without changing the rest of the request.frame_ratecontrol. range8to60; together with the other documented fields, it supports repeatable review of this model path.
Pricing
| Configuration | Price |
|---|---|
| Standard request | $9.600000 |
When to Use
| Scenario | Why this model fits |
|---|---|
| Create the exact route output | Choose it when you need custom video-model training and already have a curated training set and training controls. |
| Develop controlled variations | Keep the main brief fixed while changing one documented setting at a time to compare motion, framing, quality, or asset behavior. |
| Build repeatable batches | Use a consistent request shape for catalog, campaign, storyboard, game-asset, or social-content production. |
| Preserve source intent | Prefer this route when the supplied reference material must remain the foundation of a reusable personalized video model. |
| Connect a media pipeline | Use asynchronous results in an automated review, approval, post-production, or asset-management workflow. |
Prompt Guide
Start with the desired result, then describe the source relationship, subject action, composition or camera behavior, lighting, style, and timing. For custom video-model training, state what must remain stable as clearly as what should change. Use only fields exposed by the schema; the example below is structurally valid for this route.
{
"rank": 32,
"stg_scale": 1,
"frame_rate": 25,
"resolution": "medium",
"validation": []
}
Technical Specs
| Specification | Value |
|---|---|
| Model ID | lightricks/ltx23-video-trainer |
| Workflow | Custom video-model training |
| Required inputs | No fields marked required by schema |
rank | integer; options: 8, 16, 32, 64, 128 |
stg_scale | number; minimum: 0; maximum: 3 |
frame_rate | integer; minimum: 8; maximum: 60 |
resolution | string; options: low, medium, high |
validation | array |
with_audio | boolean |
aspect_ratio | string; options: 21:9, 16:9, 3:2, 4:3, 5:4, 1:1, 4:5, 3:4, 2:3, 9:16 |
debug_dataset | boolean |
learning_rate | number; minimum: 1e-06; maximum: 1 |
trigger_phrase | string |