AI • Technology • Digital InnovationIdeas, tools and guides for the future
AI GAME CHANGER PRO
Discover AI. Explore Technology. Create More.

LTX-2.5: Free Open AI Video That Makes the Sound Too — Full Setup Guide

November 17, 2026 • Ai Game Changer
A film set in Moscow with the camera crew at work

Most free AI video tools give you a silent clip — then you go and find the sound somewhere else. LTX-2.5, the newest open model from Lightricks, does both in one pass: it generates synchronised video and audio from a text prompt, an image, or another video. The weights are open, the code is on GitHub, and commercial use is free for individuals and small companies. Here is what it does, what it needs, and how to try it free — including the honest answer about the hardware.

What LTX-2.5 is

  • Video and audio together. LTX-2.5 is a 22-billion-parameter DiT-based audio-video foundation model. One prompt produces a clip with matching sound — dialogue, footsteps, ambient noise — instead of a silent video you have to dub later.
  • Open weights. Lightricks released the open weights in November 2026. The code lives on GitHub at github.com/Lightricks/LTX-2 (10,280 stars, checked live 17 November 2026, Apache 2.0 licence for the code), and the weights are on Hugging Face at huggingface.co/Lightricks/LTX-2.5.
  • Free for most people. The model is released under the LTX-2.x Community Licence: free for individuals and companies under $10 million in annual revenue. Bigger companies need a paid commercial licence. The weights are gated on Hugging Face — you need a free account and must accept the model terms before downloading.
  • Multi-shot scenes. New in LTX-2.5: native multishot generation — connected scenes in a single pass that hold character identity, environment, lighting, voice and visual style across cuts, instead of one continuous shot.
  • Production quality mode. DFR (Diffusion Fidelity Rendering) is the slower, higher-quality path: generated keyframes plus a spatial detailing pass, with 4K output available. The fast path is the distilled pipeline, which runs in only 8 steps.
  • More than text-to-video. Image-to-video, audio-to-video (generate video from an audio file), keyframe interpolation, video-to-video, SDR-to-HDR conversion, and DubIt — re-dubbing a clip while matching the speaker's identity and lip movements. There is also a LoRA trainer, so you can train the model on your own character or style.

What you need — the honest version

  • Strong GPU for local use. The full distilled split pack is roughly 66 GB of downloads (22B transformer, Gemma 4 12B text encoder, video and audio VAEs, spatial upscaler). You need a strong NVIDIA GPU, Python 3.12 or newer, CUDA 12.7 and PyTorch ~2.7. The README's own low-memory options are --quantization fp8-cast and --offload cpu.
  • Free Colab will not run this one. Be straight about it: a free Colab T4 GPU (about 15 GB of memory) is not enough for a 22B model. If you want to run it yourself, you need a serious GPU or a rented cloud GPU — which costs per hour, not per video. The genuinely free way to use LTX-2.5 today is the demo in Step 1.

Step 1: Try it free — no install (the free demo)

Open the LTX API Playground at console.ltx.video — the official demo linked from the Hugging Face model page. Sign in with a free account, type a prompt, and generate a clip with sound, right in the browser. Nothing to install, nothing to download. LTX Studio also has a free tier. Use the demo first: if the look and the sound are what you want, then decide whether local installation is worth it.

Step 2: Install it on your own GPU

  1. Clone the official repository (the only official source for the code):
    git clone https://github.com/Lightricks/LTX-2.git
    cd LTX-2
  2. Install the dependencies:
    uv sync --extra natten
  3. Log in to Hugging Face and accept the model terms on the LTX-2.5 model page, then download the distilled split pack (about 66 GB):
    hf auth login
    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
    If you get a 401/403 error, you have not accepted the terms on Hugging Face yet, or your token lacks access to gated repos.

Step 3: Generate your first clip with sound

Run the distilled pipeline — the fast starting point:

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 golden retriever running through a sunny meadow, cinematic lighting"

Defaults are 1024×1536 at 24 fps. Add --quantization fp8-cast --offload cpu if your GPU memory is tight. For production quality, run the DFR pipeline instead (same files, plus a detailing IC-LoRA from a separate Hugging Face repo) — slower and heavier, but sharper.

Step 4: Write prompts the way LTX-2 likes them

The official prompting guide says: write detailed, chronological descriptions of actions and scenes in a single flowing paragraph, under 200 words. Start directly with the action, then add movements, appearances, camera angles, lighting and colours — like a cinematographer describing a shot list. Keep it literal and precise. There is also an automatic prompt enhancer (enhance_prompt) that expands a short prompt into richer cinematic instructions, and an optional duration head that picks the clip length from your prompt.

Step 5: The other routes — ComfyUI, Diffusers, and training your own LoRA

  • ComfyUI: official LTX-2.5 workflow templates ship in ComfyUI — follow the instructions at github.com/Lightricks/ComfyUI-LTXVideo.
  • Diffusers: a Diffusers-compatible pack lives at Lightricks/LTX-2.5-Diffusers (install diffusers from main, as the model card explains).
  • Your own style: the repository includes ltx-trainer for LoRA, full fine-tuning and IC-LoRA — train it on your character, setting or visual style. Most LoRAs trained on LTX-2.3 run on LTX-2.5 without changes, according to the model card.

What it cannot do

The model card is clear: LTX-2.5 is not a source of facts, it may amplify existing biases, prompt following depends heavily on prompting style, and it may generate inappropriate content. The weights are gated and the licence is a community licence, not a standard permissive open-source licence — read the LTX-2.x Community Licence before using it commercially. And the honest hardware line again: this is a 22B model. The free demo is genuinely free; running it at home is for people with a strong GPU.

Bottom line

FramePack gave you free video on a 6 GB card, and Kokoro, ZONOS3 and Haru gave you free voice. LTX-2.5 is the next piece: free, open-weights video with synchronised sound, multishot scenes, and a LoRA trainer for your own style — with a free browser demo so you can hear it before you install anything. If you make videos, try the playground first; it costs nothing and takes two minutes.

Sources: the LTX-2 GitHub README and LTX-2.5 Hugging Face model page, both checked live 17 November 2026 (10,280 GitHub stars, ~66 GB download, LTX-2.x Community Licence, free under $10M revenue). Hero photo: a film set in Moscow, by Lesless, CC BY-SA 4.0, via Wikimedia Commons.