training
LTX-2 Audio-Video LoRA Trainer lets you train custom LoRA models with synchronized audio-video generation support. Train action, motion, and video effect models by uploading a ZIP file containing videos with optional audio. Ready-to-use REST inference API, best performance, no coldstarts, affordable pricing.
Idle
您的請求將花費 $0.35 每次運行。
使用 $10 您可以運行此模型大約 28 次。
README
LTX-2 19B Video-LoRA Trainer
LTX-2 19B Video-LoRA Trainer is a high-performance custom model training service for the LTX-2 19B video generation model. Train lightweight LoRA (Low-Rank Adaptation) adapters directly from video clips — capturing motion patterns, visual styles, and character appearances for personalized video generation with synchronized audio.
Why Choose This?
-
Train from videos Upload video clips directly to capture motion dynamics, temporal patterns, and visual styles that static images cannot convey.
-
Motion learning The trainer learns from video sequences, enabling LoRAs that understand movement, transitions, and temporal consistency.
-
Character consistency Create LoRAs that maintain character identity and motion style across generated video clips.
-
Seamless integration Trained LoRAs work directly with LTX-2 Text-to-Video LoRA and Image-to-Video LoRA models.
Training Process
-
Data Upload Prepare and upload a ZIP file containing your training videos. Include diverse clips that represent the style, character, or motion you want to capture.
-
Configure Trigger Word Set a unique trigger word (e.g., "p3r5on") that will activate your trained style or character in prompts.
-
Adjust Training Parameters
- steps — Total training iterations (default: 500)
- learning_rate — Training speed (default: 0.0002)
- lora_rank — Adapter capacity (default: 32)
-
LoRA Training The system runs a tailored LoRA optimization loop:
- Freezes the base model weights
- Trains only the low-rank adapter layers
- Learns from video sequences for motion-aware adaptation
-
Model Export After training completes, you receive a LoRA adapter file (.safetensors) compatible with:
Parameters
| Parameter | Default | Description |
|---|---|---|
| data | — | ZIP file containing training videos (required) |
| trigger_word | — | Unique word to activate your trained concept |
| steps | 500 | Total training iterations |
| learning_rate | 0.0002 | Training speed (lower = more stable, higher = faster) |
| lora_rank | 32 | Adapter capacity (higher = more detail, larger file) |
Pricing
| Training Steps | Price (USD) |
|---|---|
| 100 | $0.35 |
| 500 | $1.75 |
| 1,000 | $3.50 |
| 2,000 | $7.00 |
Billing Rules
- Base price: $0.35 per 100 steps
- Total cost = $0.35 × (steps / 100)
- Billed proportionally to the total number of steps in your job
Best Use Cases
- Motion Styles — Train on dance videos, action sequences, or specific movement patterns.
- Character Animation — Capture how a character moves and behaves across multiple clips.
- Brand Videos — Create consistent motion and visual style for marketing content.
- Art Styles — Learn animated art styles from reference video clips.
Pro Tips
- Use 5-10 diverse video clips that clearly show the style or character you want to capture.
- Shorter clips (5-15 seconds) with consistent quality work better than long mixed footage.
- Choose a unique trigger word that won't conflict with common words.
- Higher lora_rank (32-64) captures more detail but increases training time and file size.
- Start with default settings, then adjust if needed.
Try More Trainers
- LTX-2 19B IC-LoRA Trainer — Train LoRAs from images for LTX-2 video generation.
- Qwen Image 2512 LoRA Trainer — Train LoRAs for Qwen Image text-to-image model.
- Z-Image LoRA Trainer — Train LoRAs for Z-Image models.
Guidance
Notes
- Higher parameter values (steps, lora_rank) will increase training time.
- Training time scales with the number and length of videos configured.
- For faster iterations, start with lower settings and increase gradually.
- Video-based training captures motion patterns that image-based training cannot.