LTX-2.5 Review: A 22B Model That Generates Video and Audio at Once, From Download to First Clip

The hottest model in Hugging Face's video generation category right now is LTX-2.5 from Lightricks. As of November 5, 2026, it has 10,653 likes and 2,561,845 downloads, far ahead of other video models. It was released on August 11, 2026, and replaces both LTX-Video (v0.9.x) and LTX-2.

This article is an analysis based on the official GitHub repository (Lightricks/LTX-2) and the Hugging Face model card, not a hands-on generation test. The server used for this article has no GPU, so nothing was run locally. Therefore this piece makes no claims about generation quality; it only records what to download and what is required, factually.

What Changed

The core of LTX-2.5 is that it generates audio and video in a single model. Its tags cover not only text-to-video, image-to-video, and video-to-video, but also text-to-audio, audio-to-video, text-to-audio-video, and image-text-to-audio-video. That is a different approach from the common pattern of attaching a separate speech model as post-processing.

The architecture is a 22B transformer (DiT), and the official selling points are:

• Synchronized audio and video generation
• Native multi-shot generation
• Stronger prompt adherence
• Better local performance

The text encoder is Gemma 4 12B (fine-tuned for LTX with the projection bundled in). Stock Gemma 4 from Google is not a substitute: the documentation states that loading checks the encoder version the checkpoint was trained with (gemma4-12b-ltx-v1).

Files and Sizes You Need

The repository splits each component into separate files, so you only download what your pipeline needs. The actual file sizes as of November 5, 2026 are:

The quick-start combination (distilled bf16, text encoder, video and audio VAE, spatial upscaler) is about 66 GiB. Keeping at least 100 GB of free disk space is safer. The int8 and nvfp4 files are quantized versions of the same model, so pick only one.

Choosing a Pipeline

According to the official documentation, pipelines are split by purpose.

The defaults are 1024x1536 at 24 fps, and 4K is --width 3840 --height 2176. The documentation stresses 2176, not 2160. The example command's --num-frames 121 is about five seconds at 24 fps.

When Memory Runs Short

The official documentation does not publish absolute VRAM figures; instead it recommends techniques.

• Quantize and offload weights with --quantization fp8-cast --offload {cpu,disk}
• Use --tile for 4K (running 4K untiled may OOM)
• Install the natten extra for the fastest DiffVAE decoding (Linux and CUDA only; falls back to Triton without it)
• Disabling automatic memory cleanup between stages speeds things up when you have enough VRAM
• Use the single-stage pipeline when you do not need high resolution

In other words, running a 22B model on 24 GB of VRAM requires a combination of quantization, offloading, and tiling, and speed drops accordingly.

Using ComfyUI

The official ComfyUI nodes live in the Lightricks/ComfyUI-LTXVideo repository. The model repository also ships ComfyUI-specific int8 and nvfp4 files, so users who prefer a GUI over the Python pipeline can take that path.

The Other Riser: Pika-H3-Flex

Pika-Labs/Pika-H3-Flex, released on October 29, 2026, is another notable entry with 439 likes and 863 downloads at the time of writing. It is a LoRA adapter built on MiniMax-H3 that adds audio generation to image-to-video. It provides two examples, first-last frame to video with audio (fl2va) and reference to video with audio (ref2va), and includes ComfyUI LoRA files. Its license is the MiniMax-H3 community license.

A giant single model (LTX-2.5) and an adapter ecosystem (Pika-H3-Flex) rising at the same time is the defining shape of this week's video generation landscape.

How to Download

1. Accept the License (Required)

LTX-2.5 is a gated model. Downloading without logging in returns 401 or 403. You must sign in with a Hugging Face account and accept the terms on the model page.

1. Go to <https://huggingface.co/Lightricks/LTX-2.5> and sign in
2. Click the license agreement button on the page (LTX-2.x Community License)
3. Create a Read token under Settings and Access Tokens
4. For fine-grained tokens, enable the read gated repos scope

2. CLI Login

pip install -U huggingface_hub
hf auth login
# Paste the token (input is hidden)

With older CLI versions, huggingface-cli login also works.

3. Download Only the Quick-Start Files (About 66 GiB)

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 \
  latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  --local-dir models/ltx-2.5

With --local-dir, the repository folder layout (diffusion_models/, vae/, and so on) is preserved, which is why the run command below includes those folder names.

If VRAM is tight, download the int8 or nvfp4 files instead of bf16.

# Example: a light combination for 24 GB-class VRAM
hf download Lightricks/LTX-2.5 \
  diffusion_models/ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors \
  text_encoders/gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors \
  vae/ltx-2.5-video-vae-bf16.safetensors \
  vae/ltx-2.5-audio-vae-bf16.safetensors \
  latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  --local-dir models/ltx-2.5

4. Install the Code and Generate a First Clip

git clone https://github.com/Lightricks/LTX-2.git
cd LTX-2
uv sync --extra natten   # Skipped automatically on Windows and macOS

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 \
  --spatial-upsampler-path models/ltx-2.5/latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  --num-frames 121 \
  --seed 42 \
  --output-path output.mp4 \
  --prompt "A medium close-up shot of a man looking at the camera, saying 'This is the best tool I have ever seen.'"

5. Verify the Download

ls -lh models/ltx-2.5/diffusion_models/
du -sh models/ltx-2.5

If a file size differs from the table above, the download was interrupted. Running hf download again resumes from where it stopped.

6. ComfyUI Users

git clone https://github.com/Lightricks/ComfyUI-LTXVideo.git

For ComfyUI, distribute the files into models/diffusion_models, models/text_encoders, models/vae, models/upscale_models, and models/loras. Accepting the license and running hf auth login are still required.

Summary

LTX-2.5 is a model trying to end the era of generating video and audio separately. Before performance, the first walls are the 66 GiB download, the gated license, and the memory to run 22B. Check disk, VRAM, and token access before downloading.

• Gate acceptance and a Read token with read gated repos
• About 66 GiB for quick start, or int8/nvfp4 light combinations
• 24 fps by default, 4K at 3840x2176
• When short on memory: fp8-cast quantization, cpu/disk offload, and tiling
• Code is Apache-2.0; weights are under the LTX-2.x Community License