PyTorch API

For developers who want direct Python integration or custom workflows beyond ComfyUI, LTX-2 offers two paths: the native ltx-pipelines package (full control, all features) and the HuggingFace Diffusers integration (familiar API, quick start).

Repository Structure

The LTX-2 codebase is a monorepo with three packages:

Requirements

• Python >= 3.10
• CUDA > 12.7
• PyTorch ~= 2.7

See System Requirements for full hardware specifications.

Installation

$
# Clone the repository
$
git clone https://github.com/Lightricks/LTX-2.git
$
cd LTX-2
$
$
# Set up the environment
$
uv sync --frozen
$
source .venv/bin/activate

Download Models

$
# Download the distilled model (recommended for speed)
$
huggingface-cli download Lightricks/LTX-2.3 \
>
  --include "ltx-2.3-22b-distilled-1.1.safetensors" \
>
  --local-dir models/
$
$
# Download the full model (higher quality, slower generation)
$
huggingface-cli download Lightricks/LTX-2.3 \
>
  --include "ltx-2.3-22b-dev.safetensors" \
>
  --local-dir models/
$
$
# FP8 variant (smaller download, requires less VRAM)
$
huggingface-cli download Lightricks/LTX-2.3-fp8 \
>
  --include "ltx-2.3-22b-dev-fp8.safetensors" \
>
  --local-dir models/

A full list of available checkpoints (including spatial/temporal upscalers, distilled LoRA, IC-LoRA variants, and camera control LoRAs) is on the LTX-2 GitHub repo.

Option 1: Native Pipelines (ltx-pipelines)

The ltx-pipelines package provides the most complete feature set, including two-stage generation, IC-LoRA, and fine-grained guidance control.

Available Pipelines

Text-to-Video Example

This example generates a video using the distilled pipeline with a two-stage approach — base generation followed by upscale refinement.

1
import torch
2
from ltx_pipelines.distilled_pipeline import DistilledPipeline
3
4
# Initialize
5
pipeline = DistilledPipeline.from_config("path/to/config.yaml")
6
7
# Generate
8
output = pipeline(
9
    prompt="A golden retriever running through a sunlit meadow, "
10
           "wildflowers swaying in a gentle breeze. "
11
           "Camera follows at ground level, tracking the dog. "
12
           "Warm afternoon light with soft bokeh in the background. "
13
           "Sound of panting, rustling grass, distant birdsong.",
14
    width=768,
15
    height=512,
16
    num_frames=97,
17
    fps=24.0,
18
    seed=42,
19
)

Guidance Parameters

The native pipelines expose MultiModalGuiderParams for fine-grained control over generation:

Memory Optimization

For consumer GPUs, enable FP8 quantization to reduce VRAM usage by ~40% with minimal quality loss:

1
from ltx_core.quantization.policy import QuantizationPolicy
2
3
# FP8 Cast — broad GPU compatibility
4
pipeline = DistilledPipeline.from_config(
5
    "path/to/config.yaml",
6
    quantization=QuantizationPolicy.fp8_cast(),
7
)
8
9
# FP8 Scaled MM — optimized for Hopper GPUs (H100)
10
# Requires: uv sync --frozen --extra fp8-trtllm
11
pipeline = DistilledPipeline.from_config(
12
    "path/to/config.yaml",
13
    quantization=QuantizationPolicy.fp8_scaled_mm(),
14
)

From the command line:

$
# FP8 Cast (works on most GPUs)
$
python run_pipeline.py --quantization fp8-cast
$
$
# FP8 Scaled MM (Hopper GPUs only, uses TensorRT-LLM)
$
python run_pipeline.py --quantization fp8-scaled-mm

Additional tip: Set the environment variable PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to improve memory allocation.

Option 2: HuggingFace Diffusers

If you’re already working within the Diffusers ecosystem, LTX-2 is available as a native pipeline:

1
import torch
2
from diffusers import LTX2Pipeline
3
4
pipeline = LTX2Pipeline.from_pretrained(
5
    "Lightricks/LTX-2",
6
    torch_dtype=torch.bfloat16,
7
)
8
pipeline.to("cuda")
9
10
# Generate
11
result = pipeline(
12
    prompt="A golden retriever running through a sunlit meadow, "
13
           "wildflowers swaying in a gentle breeze.",
14
    width=768,
15
    height=512,
16
    num_frames=97,
17
)

For more on the Diffusers integration, see the HuggingFace documentation.

Generation Parameters Reference

Resolution

Standard aspect ratios:

Higher resolutions are supported (up to 4K) but require significantly more VRAM. Start with lower resolutions for testing.

Frame Count & Duration

Sampling

What’s Next

• Prompting Guide — write prompts that produce coherent motion and audio
• LoRA Guide — customize style, motion, and character appearance
• IC-LoRA Guide — video-to-video with depth, pose, and edge control
• Trainer Documentation — train your own LoRAs