This is an FP8 / INT8 quantized version of the LTX-2.5 distilled video DiT.

Each file replaces the 38.0 GB bf16 transformer of Lightricks/LTX-2.5-Diffusers with a 19.1 to 25.1 GB one. The text encoder, connectors, VAEs, vocoder and latent upsampler are unchanged and come from the Diffusers repo.
Same-seed LPIPS vs the bf16 DiT, averaged over all 121 frames of 4 clips (lower is better): 0.265 INT8-ConvRot, 0.366 FP8, 0.385 INT8. INT8-ConvRot is the recommended file.
FP8 and INT8 are bit-identical to quantizing the bf16 DiT at load with the same recipe, so they add no error over on-the-fly quantization; they save the download and the bf16 staging.

file	size	scheme	LPIPS vs bf16 (mean / max frame)	PSNR vs bf16 (mean / min frame)
`LTX-2.5-INT8-ConvRot.safetensors`	25.11 GB	INT8 W8A8, group-256 Hadamard rotation	0.265 / 0.449	20.2 / 16.8 dB
`LTX-2.5-FP8.safetensors`	19.06 GB	FP8 e4m3, per-row scales	0.366 / 0.518	17.6 / 11.6 dB
`LTX-2.5-INT8.safetensors`	25.11 GB	INT8 W8A8, no rotation	0.385 / 0.632	17.4 / 13.9 dB

Built from the transformer/ of Lightricks/LTX-2.5-Diffusers at revision da0b690960026eab794b8ceaf841233cf7064478 (tensor-identical to ltx-2.5-22b-distilled-transformer-bf16.safetensors in Lightricks/LTX-2.5). FP8 quantizes every linear with at least 512 features. INT8 keeps the audio stream, the audio-video cross attention, the adaLN and timestep projections and the norms in bf16. INT8-ConvRot rotates the inputs of all 480 quantized video linears with a block Hadamard (group 256) before quantizing; the rotation is listed in the file's metadata and applied to the activations at run time.

Scores are against the bf16 DiT at the reference distilled settings (1280x704, 121 frames, 24 fps, 8 + 3 two-stage steps, CFG 1), text-to-video and image-to-video, seeds 42 and 7, with the same bf16 text encoder in every run; a repeated bf16 render is bit-identical (LPIPS 0.000). The distilled sampler injects noise every step, so small weight errors move the motion more than the look of a frame: every file keeps the subject, scene and prompt, as the samples show.

Every file is safetensors in torchao's layout (tensor_names plus a per-tensor entry in the header, with the three root-level parameters stored as unsloth_root::<name>), so torchao.prototype.safetensors.safetensors_support.unflatten_tensor_state_dict rebuilds it and the stock diffusers LTX2VideoTransformer3DModel loads it strictly. Loading runs no pickled code. Checked with torch 2.13 and torchao 0.18.

Samples

Frame 60 of each clip.

bf16	INT8-ConvRot	FP8

LTX-2.5 — Video, Audio & World Simulation

Full control and customization — self-host on your infrastructure.

Homepage Docs GitHub Research API Playground Discord

LTX License

Under $10M annual revenue

Commercial and production use at no cost under the LTX-2.x Community License. Transfer of fine-tunes may require a paid license, in accordance with the LTX-2.x Community License.

Read the Documentation

Over $10M annual revenue

Paid Commercial Use Agreement for LTX-2.x with full weights, engineering support, LoRAs, and flexible deployment options. To learn about all licensing options, talk to an expert.

Talk to a Commercial Licensing Expert

LTX-2.5 is an open world model with open weights, built for local execution and fine-tuning. Its established use is generating synchronized, high-fidelity video and audio from text, image, and video inputs; applicability to emerging domains such as robotics and physical AI is developing.

Full control and customization — self-host on your own infrastructure. No per-generation billing, no per-seat lock-in, no forced API dependency. Revenue is measured across the whole entity, including subsidiaries and affiliates under common control. The full, binding terms live in LICENSE.

What's new in LTX-2.5

Native multishot generation — generate connected scenes in a single pass: multiple shots that hold character identity, environment, lighting, voice, and visual style across cuts (previous versions produced a single continuous shot).
Diffusion fidelity rendering — Instead of locking every scene to one compression rate, our model dynamically allocates compute by scene complexity and budget, rendering flawless detail where it matters, efficient everywhere else.
New diffusion video decoder — replaces the VAE reconstruction stage; sharper faces, textures, and on-screen text, better motion, and fewer artifacts in demanding scenes.
Custom Gemma 4 12B text encoder — holds complex prompts together (multiple characters, camera moves, lighting, actions) instead of dropping details across a longer sequence.
Prompt enhancer — expands a short prompt into richer cinematic instructions at minimal extra compute.
Duration predictor (optional) — an opt-in node predicts a clip's length from the prompt and sets the frame count for you, instead of relying on a fixed-duration parameter.
Substantially improved distilled model — retains much more of the full model's visual quality, prompt adherence, and motion consistency in a smaller, faster checkpoint.

Model family & checkpoints

LTX-2.5 ships as a split, Comfy-aligned pack (one .safetensors per component) rather than a single monolith. Point each CLI flag / loader at the file below.

Transformers (DiT)

File	Notes
`diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors`	Distilled DiT (bf16). Fixed 8-step schedule, CFG=1.
`diffusion_models/ltx-2.5-22b-dev-transformer-bf16.safetensors`	Full / trainable DiT (bf16).
`diffusion_models/ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors`	Distilled DiT (Comfy int8 + convrot). ComfyUI only — not for `ltx-pipelines` / PyTorch.
`diffusion_models/ltx-2.5-22b-dev-transformer-comfy-int8-convrot.safetensors`	Full DiT (Comfy int8 + convrot). ComfyUI only — not for `ltx-pipelines` / PyTorch.
`diffusion_models/ltx-2.5-22b-distilled-transformer-nvfp4.safetensors`	Distilled DiT (NVFP4). ComfyUI, or `ltx-pipelines` with `--quantization nvfp4-prequant` (Blackwell / `ltx-kernels`).

Other components

File	Notes
`text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors`	Gemma4 TE + projections (bf16)
`text_encoders/gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors`	Same TE, Comfy int8 — ComfyUI only
`vae/ltx-2.5-video-vae-bf16.safetensors`	DiffVAE — higher quality, heavier
`vae/ltx-2.5-video-vae-conv-bf16.safetensors`	Conv VAE — faster, lighter
`vae/ltx-2.5-audio-vae-bf16.safetensors`	Audio VAE + vocoder
`loras/ltx-2.5-22b-distilled-lora-450-bf16.safetensors`	Distilled LoRA (dev-transformer workflows)
`model_patches/ltx-2.5-duration-head-bf16.safetensors`	Auto duration when `--num-frames` omitted
`latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors`	x2 spatial upscaler required for multi-stage pipeline
`latent_upscale_models/ltx-2.5-latent-temporal-upscaler-x2-bf16-1.0.safetensors`	x2 temporal upscaler

Usage

Online demo

Try LTX-2.5 in the API Playground without installing anything locally.

Option A — Python (`ltx-pipelines`)

Weights on this repo are split (Comfy-aligned): one safetensors file per component. The LTX-2 ltx-pipelines package loads them via --transformer-path, --text-encoder-path, etc.

Install

git clone https://github.com/Lightricks/LTX-2.git
cd LTX-2
uv sync
source .venv/bin/activate

Python >= 3.12, CUDA >= 12.7, PyTorch ~= 2.7 recommended. See the repo README for attention backends and optional extras.

Download weights

hf auth login

# LTX-2.5 distilled split pack
hf download Lightricks/LTX-2.5 \
  diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
  text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
  vae/ltx-2.5-video-vae-bf16.safetensors \
  vae/ltx-2.5-audio-vae-bf16.safetensors \
  model_patches/ltx-2.5-duration-head-bf16.safetensors \
  latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  --local-dir models/ltx-2.5

Distilled text-to-video

uv run python -m ltx_pipelines.distilled \
  --transformer-path     models/ltx-2.5/diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
  --text-encoder-path    models/ltx-2.5/text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
  --video-vae-path       models/ltx-2.5/vae/ltx-2.5-video-vae-bf16.safetensors \
  --audio-vae-path       models/ltx-2.5/vae/ltx-2.5-audio-vae-bf16.safetensors \
  --duration-head-path   models/ltx-2.5/model_patches/ltx-2.5-duration-head-bf16.safetensors \
  --spatial-upsampler-path models/ltx-2.5/latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  --prompt "A golden retriever running through a sunny meadow, cinematic lighting" \
  --seed 42 \
  --output-path output_distilled.mp4

Omit --num-frames to let the duration head pick a length from the prompt (LTX-2.5+). Or set e.g. --num-frames 121 (must satisfy frames % 8 == 1). Width/height must be divisible by 32.

Image-to-video

Add one or more --image PATH FRAME_IDX STRENGTH flags (frame 0 = first frame conditioning):

uv run python -m ltx_pipelines.distilled \
  --transformer-path     models/ltx-2.5/diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
  --text-encoder-path    models/ltx-2.5/text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
  --video-vae-path       models/ltx-2.5/vae/ltx-2.5-video-vae-bf16.safetensors \
  --audio-vae-path       models/ltx-2.5/vae/ltx-2.5-audio-vae-bf16.safetensors \
  --duration-head-path   models/ltx-2.5/model_patches/ltx-2.5-duration-head-bf16.safetensors \
  --spatial-upsampler-path models/ltx-2.5/latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  --image path/to/first_frame.jpg 0 1.0 \
  --prompt "The camera slowly dollies out as wind moves through the grass" \
  --seed 42 \
  --output-path output_i2v.mp4

Low-VRAM tips

# Downcast bf16 transformer on the fly + CPU offload
  ...existing flags... \
  --quantization fp8-cast \
  --offload cpu

Use the bf16 checkpoints with ltx-pipelines. The *-comfy-int8-convrot.safetensors files are ComfyUI-only and are not loaded by this PyTorch path.

Python API (same split paths)

from ltx_pipelines.distilled import DistilledPipeline
from ltx_pipelines.utils.model_paths import ModelPaths

model_paths = ModelPaths.from_split(
    transformer_path="models/ltx-2.5/diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors",
    text_encoder_path="models/ltx-2.5/text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors",
    video_vae_path="models/ltx-2.5/vae/ltx-2.5-video-vae-bf16.safetensors",
    audio_vae_path="models/ltx-2.5/vae/ltx-2.5-audio-vae-bf16.safetensors",
    duration_head_path="models/ltx-2.5/model_patches/ltx-2.5-duration-head-bf16.safetensors",
)

pipe = DistilledPipeline(
    model_paths=model_paths,
    spatial_upsampler_path="models/ltx-2.5/latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors",
)
# See packages/ltx-pipelines for __call__ args (prompt, seed, num_frames, images, ...).

uv run python -m ltx_pipelines.distilled --help

Full docs: ltx-pipelines installation.

Option B — ComfyUI

Official LTX-2.5 workflow templates ship in ComfyUI. Full instructions: ComfyUI integration.

Option C — Diffusers

A Diffusers-compatible pack lives at Lightricks/LTX-2.5-Diffusers — same model, Diffusers-friendly packaging.

Install

LTX-2.5 support is not in a diffusers release yet, so install from main:

pip install git+https://github.com/huggingface/diffusers

Image-to-video, two stages

import torch
from diffusers import LTX2ImageToVideoPipeline, LTX2LatentUpsamplePipeline
from diffusers.pipelines.ltx2.latent_upsampler import LTX2LatentUpsamplerModel
from diffusers.pipelines.ltx2.utils import (
    DEFAULT_NEGATIVE_PROMPT,
    DISTILLED_SIGMA_VALUES,
    STAGE_2_DISTILLED_SIGMA_VALUES,
)
from diffusers.utils import encode_video, load_image

MODEL_ID = "Lightricks/LTX-2.5-Diffusers"
# Stage 1 resolution; stage 2 runs at 2x this.
HEIGHT, WIDTH, NUM_FRAMES, FRAME_RATE = 544, 960, 121, 24.0

pipe = LTX2ImageToVideoPipeline.from_pretrained(MODEL_ID, dtype=torch.bfloat16)
pipe.enable_model_cpu_offload()
pipe.vae.enable_tiling()  # stage 2 decodes at 2x

latent_upsampler = LTX2LatentUpsamplerModel.from_pretrained(
    MODEL_ID, subfolder="latent_upsampler", dtype=torch.bfloat16
).to("cuda")
upsample_pipe = LTX2LatentUpsamplePipeline(vae=pipe.vae, latent_upsampler=latent_upsampler)

generator = torch.Generator("cuda").manual_seed(42)
shared = dict(
    image=load_image("path/to/first_frame.jpg"),
    prompt="The camera slowly dollies out as wind moves through the grass",
    negative_prompt=DEFAULT_NEGATIVE_PROMPT,
    frame_rate=FRAME_RATE,
    guidance_scale=1.0,
    audio_guidance_scale=1.0,
    stg_scale=0.0,
    audio_stg_scale=0.0,
    modality_scale=1.0,
    audio_modality_scale=1.0,
    generator=generator,
    return_dict=False,
)

stage_1_latents, audio_latents = pipe(
    height=HEIGHT, width=WIDTH, num_frames=NUM_FRAMES,
    sigmas=DISTILLED_SIGMA_VALUES, output_type="latent", **shared,
)

upsampled_latents = upsample_pipe(
    latents=stage_1_latents, output_type="latent", return_dict=False
)[0]

# Stage 2 takes its size from the upsampled latents, so pass no height/width.
video, audio = pipe(
    num_frames=NUM_FRAMES,
    sigmas=STAGE_2_DISTILLED_SIGMA_VALUES,
    latents=upsampled_latents,
    audio_latents=audio_latents,
    noise_scale=STAGE_2_DISTILLED_SIGMA_VALUES[0],
    output_type="np",
    **shared,
)

encode_video(
    video[0],
    fps=int(FRAME_RATE),
    output_path="output_i2v_two_stage.mp4",
    audio=audio[0].float().cpu(),
    audio_sample_rate=pipe.vocoder.config.output_sampling_rate,
)

Constraints

Frame count: num_frames % 8 == 1 (1, 9, 17, …, 121, …)
Width and height divisible by 32

Prompting

Well-structured, detailed prompts materially improve results. For multishot prompting and a full guide, see How to prompt LTX-2.

Training & fine-tuning

The dev transformer is fully trainable. Reproduce published LoRAs and IC-LoRAs with the LTX-2 Trainer.

Based on our testing, the large majority of LoRAs and IC-LoRAs trained on LTX-2.3 run on LTX-2.5 without changes. A small number of exceptions exist — validate your adapters before production use.

Limitations

This model is not intended or able to provide factual information.
As a statistical model, this checkpoint may amplify existing societal biases.
Prompt following is heavily influenced by prompting style.
The model may fail to generate videos that match the prompt perfectly.
The model may generate content that is inappropriate or offensive.

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={2026}
}

Downloads last month: -

Model tree for unsloth/LTX-2.5-FP8

Base model

Lightricks/LTX-2.5-Diffusers

Finetuned

(1)

this model

Paper for unsloth/LTX-2.5-FP8

LTX-2: Efficient Joint Audio-Visual Foundation Model

Paper • 2601.03233 • Published Jan 6 • 199

Samples

LTX-2.5 — Video, Audio & World Simulation

LTX-2.5 — Video, Audio & World Simulation

What's new in LTX-2.5

Model family & checkpoints

Transformers (DiT)

Other components

Usage

Online demo

Option A — Python (ltx-pipelines)

Install

Download weights

Distilled text-to-video

Image-to-video

Low-VRAM tips

Python API (same split paths)

Option B — ComfyUI

Option C — Diffusers

Install

Image-to-video, two stages

Constraints

Prompting

Training & fine-tuning

Limitations

Citation

Model tree for unsloth/LTX-2.5-FP8

Paper for unsloth/LTX-2.5-FP8

Option A — Python (`ltx-pipelines`)