Usage Guides

LTX-2 LoRAs

LoRAs (Low-Rank Adaptation) are lightweight model modifications that customize LTX-2’s output for specific styles, effects, or visual characteristics. Unlike full model fine-tuning, LoRAs require only 1-128MB of additional weights, making them efficient and easy to share.

Capabilities

Implementing a LoRA allows you to:

  • Enforce Style Consistency: Lock video generation to specific aesthetic guidelines (e.g., specific cinematic looks, line-art styles, or brand-aligned color palettes).
  • Enhance Subject Fidelity: Improve the model’s ability to retain specific character or object details during motion.
  • Fine-Tune Motion Dynamics: Adjust how the model interprets movement for specific use cases.
  • Add Structural Control: Enable precise control through depth maps, pose skeletons, or edge detection.

Available LoRAs

LTX-2 provides several official LoRAs for common use cases:

Using LoRAs

ComfyUI

ComfyUI provides the most streamlined workflow for using LoRAs with LTX-2. Two methods are available:

Method 1: Official ComfyUI-LTXVideo Nodes (Recommended)

  1. Install the ComfyUI-LTXVideo custom nodes
  2. Download LoRA files from Hugging Face
  3. Place LoRA files in ComfyUI/models/loras/
  4. Load workflows from the example workflows

For IC-LoRA control workflows, see the IC-LoRA guide.

Method 2: Community LoRA Loader

For advanced users requiring specific LoRA loading behavior, install ComfyUI-LTXVideoLoRA:

$cd ComfyUI/custom_nodes
>git clone https://github.com/dorpxam/ComfyUI-LTXVideoLoRA.git

This provides additional nodes for handling different LoRA key formats and GGUF support.

Adjusting LoRA Strength

LoRA strength controls how much influence the LoRA has on generation. The value ranges from 0.0 (no effect) to 1.0 (full effect):

  • 0.9-1.1: Subtle effect, preserves base model characteristics
  • 1.2-1.4: Balanced, recommended for most use cases
  • 1.5-1.6: Strong effect, maximum style transfer

Finding the Right Strength:

  • Start at 1.0 for effect LoRAs
  • Use 1.0 for IC-LoRA control models (they’re designed for full strength)
  • Lower strength if the effect overwhelms the scene
  • Increase if the effect is too subtle

Combining Multiple LoRAs

You can stack multiple LoRAs, though this requires careful strength balancing:

1pipe.load_lora_weights("path/to/lora1", adapter_name="style")
2pipe.load_lora_weights("path/to/lora2", adapter_name="effect")
3
4# Adjust individual strengths
5pipe.set_adapters(["style", "effect"], [0.6, 0.8])

Best Practices:

  • Keep total combined strength under 2.0
  • Test combinations incrementally
  • Test different resolutions and aspect ratios
  • Effect LoRAs generally combine better than control LoRAs
  • Avoid mixing multiple IC-LoRA control types

For Effect LoRAs

  • Use clear, descriptive prompts that align with the LoRA’s training data
  • Combine with appropriate negative prompts to avoid artifacts
  • Test at different resolutions - some effects scale differently

Performance Optimization

  • FP8 quantized models work with LoRAs and reduce VRAM usage
  • LoRAs add minimal compute overhead (less than 5% typically)
  • Distilled models work with LoRAs trained on dev models

Training Custom LoRAs

Create your own LoRAs to capture specific visual styles and effects using the LTX-2 Trainer.

For complete training instructions, see the Trainer documentation.