This is an FP8 / INT8 quantized version of LTX-2.

• Optimized for efficient inference with reduced memory footprint. Same-seed first-frame LPIPS vs the bf16 model (lower is better): 0.23 INT8, 0.26 FP8 (same-seed 12-step first frame; suite means 0.24 / 0.26; per-case hard checks pass).

Samples

Prompt: "cute sloth typing on a computer" (first frames)

LTX-2 Model Card

This model card focuses on the LTX-2 model, as presented in the paper LTX-2: Efficient Joint Audio-Visual Foundation Model. The codebase is available here.

LTX-2 is a DiT-based audio-video foundation model designed to generate synchronized video and audio within a single model. It brings together the core building blocks of modern video generation, with open weights and a focus on practical, local execution.

Model Checkpoints

Model Details

• Developed by: Lightricks
• Model type: Diffusion-based audio-video foundation model
• Language(s): English

Online demo

LTX-2 is accessible right away via the following links:

• LTX-Studio text-to-video
• LTX-Studio image-to-video

Run locally

Direct use license

You can use the models - full, distilled, upscalers and any derivatives of the models - for purposes under the license.

ComfyUI

We recommend you use the built-in LTXVideo nodes that can be found in the ComfyUI Manager. For manual installation information, please refer to our documentation site.

PyTorch codebase

The LTX-2 codebase is a monorepo with several packages. From model definition in 'ltx-core' to pipelines in 'ltx-pipelines' and training capabilities in 'ltx-trainer'. The codebase was tested with Python >=3.12, CUDA version >12.7, and supports PyTorch ~= 2.7.

Installation

git clone https://github.com/Lightricks/LTX-2.git
cd LTX-2

# From the repository root
uv sync
source .venv/bin/activate

Inference

To use our model, please follow the instructions in our ltx-pipelines package.

Diffusers 🧨

LTX-2 is supported in the Diffusers Python library for text & image-to-video generation. Read more on LTX-2 with diffusers here.

Use with diffusers

To achieve production quality generation, it's recommended to use the two-stage generation pipeline. Example for 2-stage inference of text-to-video:

import torch
from diffusers import FlowMatchEulerDiscreteScheduler
from diffusers.pipelines.ltx2 import LTX2Pipeline, LTX2LatentUpsamplePipeline
from diffusers.pipelines.ltx2.latent_upsampler import LTX2LatentUpsamplerModel
from diffusers.pipelines.ltx2.utils import STAGE_2_DISTILLED_SIGMA_VALUES
from diffusers.pipelines.ltx2.export_utils import encode_video

device = "cuda:0"
width = 768
height = 512

pipe = LTX2Pipeline.from_pretrained(
    "Lightricks/LTX-2", torch_dtype=torch.bfloat16
)
pipe.enable_sequential_cpu_offload(device=device)

prompt = "A beautiful sunset over the ocean"
negative_prompt = "shaky, glitchy, low quality, worst quality, deformed, distorted, disfigured, motion smear, motion artifacts, fused fingers, bad anatomy, weird hand, ugly, transition, static."

# Stage 1 default (non-distilled) inference
frame_rate = 24.0
video_latent, audio_latent = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    width=width,
    height=height,
    num_frames=121,
    frame_rate=frame_rate,
    num_inference_steps=40,
    sigmas=None,
    guidance_scale=4.0,
    output_type="latent",
    return_dict=False,
)

latent_upsampler = LTX2LatentUpsamplerModel.from_pretrained(
    "Lightricks/LTX-2",
    subfolder="latent_upsampler",
    torch_dtype=torch.bfloat16,
)
upsample_pipe = LTX2LatentUpsamplePipeline(vae=pipe.vae, latent_upsampler=latent_upsampler)
upsample_pipe.enable_model_cpu_offload(device=device)
upscaled_video_latent = upsample_pipe(
    latents=video_latent,
    output_type="latent",
    return_dict=False,
)[0]

# Load Stage 2 distilled LoRA
pipe.load_lora_weights(
    "Lightricks/LTX-2", adapter_name="stage_2_distilled", weight_name="ltx-2-19b-distilled-lora-384.safetensors"
)
pipe.set_adapters("stage_2_distilled", 1.0)
# VAE tiling is usually necessary to avoid OOM error when VAE decoding
pipe.vae.enable_tiling()
# Change scheduler to use Stage 2 distilled sigmas as is
new_scheduler = FlowMatchEulerDiscreteScheduler.from_config(
    pipe.scheduler.config, use_dynamic_shifting=False, shift_terminal=None
)
pipe.scheduler = new_scheduler
# Stage 2 inference with distilled LoRA and sigmas
video, audio = pipe(
    latents=upscaled_video_latent,
    audio_latents=audio_latent,
    prompt=prompt,
    negative_prompt=negative_prompt,
    num_inference_steps=3,
    noise_scale=STAGE_2_DISTILLED_SIGMA_VALUES[0], # renoise with first sigma value https://github.com/Lightricks/LTX-2/blob/main/packages/ltx-pipelines/src/ltx_pipelines/ti2vid_two_stages.py#L218
    sigmas=STAGE_2_DISTILLED_SIGMA_VALUES,
    guidance_scale=1.0,
    output_type="np",
    return_dict=False,
)

encode_video(
    video[0],
    fps=frame_rate,
    audio=audio[0].float().cpu(),
    audio_sample_rate=pipe.vocoder.config.output_sampling_rate,
    output_path="ltx2_lora_distilled_sample.mp4",
)

For more inference examples, including generation with the distilled checkpoint, visit here.

General tips:

• Width & height settings must be divisible by 32. Frame count must be divisible by 8 + 1.
• In case the resolution or number of frames are not divisible by 32 or 8 + 1, the input should be padded with -1 and then cropped to the desired resolution and number of frames.
• For tips on writing effective prompts, please visit our Prompting guide

Limitations

• This model is not intended or able to provide factual information.
• As a statistical model this checkpoint might amplify existing societal biases.
• The model may fail to generate videos that matches the prompts perfectly.
• Prompt following is heavily influenced by the prompting-style.
• The model may generate content that is inappropriate or offensive.
• When generating audio without speech, the audio may be of lower quality.

Train the model

The base (dev) model is fully trainable.

It's extremely easy to reproduce the LoRAs and IC-LoRAs we publish with the model by following the instructions on the LTX-2 Trainer Readme.

Training for motion, style or likeness (sound+appearance) can take less than an hour in many settings.

Citation

@article{hacohen2025ltx2,
  title={LTX-2: Efficient Joint Audio-Visual Foundation Model},
  author={HaCohen, Yoav and Brazowski, Benny and Chiprut, Nisan and Bitterman, Yaki and Kvochko, Andrew and Berkowitz, Avishai and Shalem, Daniel and Lifschitz, Daphna and Moshe, Dudu and Porat, Eitan and Richardson, Eitan and Guy Shiran and Itay Chachy and Jonathan Chetboun and Michael Finkelson and Michael Kupchick and Nir Zabari and Nitzan Guetta and Noa Kotler and Ofir Bibi and Ori Gordon and Poriya Panet and Roi Benita and Shahar Armon and Victor Kulikov and Yaron Inger and Yonatan Shiftan and Zeev Melumian and Zeev Farbman},
  journal={arXiv preprint arXiv:2601.03233},
  year={2025}
}

Quantized transformer checkpoints (this repo)

This repo adds pre-quantized diffusion transformer checkpoints for Lightricks/LTX-2, built with torchao dynamic activation quantization from the dense bf16 transformer of the diffusers-format release (Lightricks/LTX-2). The official model card above is unchanged from the source repo.

Files:

• LTX-2-INT8.pt (19.1 GB)
• LTX-2-FP8.pt (18.9 GB)

Details:

• int8: Int8DynamicActivationInt8WeightConfig (per-token activation, per-channel weight, torch._int_mm).
• fp8: Float8DynamicActivationFloat8WeightConfig with PerRow granularity (e4m3, torch._scaled_mm). The loader must floor the dynamic activation scale (activation_value_lb=1e-12 on torchao 0.13+) so all-zero activation token rows cannot produce a zero scale.
• Loading a checkpoint is bit-identical to quantizing the dense bf16 transformer on the fly; the checkpoint skips the dense load and quantize step.
• Dual-expert pipelines (Wan2.2 A14B) carry one file per expert and scheme; the -2 suffix is the transformer_2 expert. A partial pair never loads (all-or-none).
• LTX-2 is a joint audio-video transformer; the checkpoints cover the full module. Prequant load 21s (int8) / 25s (fp8) vs 34-50s dense+quantize and 82-92s bf16 load; ~37 GB resident vs 55 GB bf16.
• Validated against same-seed dense bf16 clips (LPIPS-vgg, PSNR, non-finite and black-frame checks) on torch 2.10 and torchao 0.17.

Samples

Prompt: "cute sloth typing on a computer" (first frames, family default settings, seeds 0-2).

int8

fp8

Pre-cast fp8 text encoder (this repo)

LTX-2-text_encoder-FP8.pt (13.2 GB) is the pipeline's text encoder (Gemma3ForConditionalGeneration, the text_encoder subfolder of Lightricks/LTX-2) with the layerwise fp8 storage cast Unsloth Studio applies at load time, saved pre-cast:

• Loading it is bit-identical to downloading the dense encoder and casting on load (verified tensor for tensor: 1066 tensors, 660 cast to fp8 storage).
• Cuts the text-encoder download from 48.7 (fp32 hub store) GB to 13.2 GB.
• Plain-tensor state dict: loads with torch.load(weights_only=True). Metadata records scheme fp8, component text_encoder, base Lightricks/LTX-2.