Instructions to use Lightricks/LTX-2.5 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusion Single File
How to use Lightricks/LTX-2.5 with Diffusion Single File:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- LTX-2
How to use Lightricks/LTX-2.5 with LTX-2:
# Install the LTX-2 pipelines git clone https://github.com/Lightricks/LTX-2.git cd LTX-2 uv sync --extra natten
# Download weights from this repo # Substitute filenames from this repo's "Files and versions" if they differ hf download Lightricks/LTX-2.5 \ diffusion_models/<distilled-transformer>.safetensors \ text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \ vae/<video-vae>.safetensors \ vae/<audio-vae>.safetensors \ latent_upscale_models/<spatial-upsampler>.safetensors \ latent_upscale_models/<temporal-upsampler>.safetensors \ --local-dir models/LTX-2.5 # DFR requires the detailing IC-LoRA (separate repo; strength is fixed at 0.5) hf download Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler --local-dir models/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler# Distilled LTX-2.5 pipeline (fast) uv run python -m ltx_pipelines.distilled \ --transformer-path models/LTX-2.5/diffusion_models/<distilled-transformer>.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/<video-vae>.safetensors \ --audio-vae-path models/LTX-2.5/vae/<audio-vae>.safetensors \ --spatial-upsampler-path models/LTX-2.5/latent_upscale_models/<spatial-upsampler>.safetensors \ --num-frames 121 \ --prompt "A beautiful sunset over the ocean" \ --output-path output.mp4 # For image-to-video, add: --image path/to/image.jpg 0 0.8# DFR pipeline (higher detail fidelity; optional temporal 2x/4x) uv run python -m ltx_pipelines.dfr_pipeline \ --transformer-path models/LTX-2.5/diffusion_models/<distilled-transformer>.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/<video-vae>.safetensors \ --audio-vae-path models/LTX-2.5/vae/<audio-vae>.safetensors \ --spatial-upsampler-path models/LTX-2.5/latent_upscale_models/<spatial-upsampler>.safetensors \ --temporal-upsampler-path models/LTX-2.5/latent_upscale_models/<temporal-upsampler>.safetensors \ --detailing-lora models/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler/ltx-2.5-22b-ic-lora-pixel-spatial-upscaler-x2-1.0.safetensors \ --spatial-upscalings 1 \ --temporal-upscalings 1 \ --height 1088 \ --width 1920 \ --num-frames 121 \ --prompt "A beautiful sunset over the ocean" \ --output-path output.mp4 # For 4K: --spatial-upscalings 2 --width 3840 --height 2176 # For image-to-video, add: --image path/to/image.jpg 0 0.8 - Notebooks
- Google Colab
- Kaggle
I can't control it
[Using LTX 2.5 12V. Considering what I've seen from AI, what I want to do what must be simplest thing there is. I want to see a man sitting behind a desk and speaking. So I made the set, created the character sheet and a voice for him - all that stuff. Now no matter what I do the camera pushes in. Claude and I tried all day yesterday to simply get LTX to make the clip without pushing in. We failed.
To keep interest during 2-4 minutes of speech I would like to establish several views of the desk. Now if I tell it to pull back it does, then pushes in again. ](Title: LTX 2.5 (MSR + Prompt Relay): camera slowly pushes in on a seated talking-head shot, can't stop it
Setup (ComfyUI, RTX 5060 Ti 16GB, RAM sits around 93% during renders):
- ltx-2.5-22b-distilled-transformer-comfy-int8-convrot
- gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot text encoder
- ltx-2.5-video-vae-bf16 and ltx-2.5-audio-vae-bf16
- Licon MSR V2 IC-LoRA (ComfyUILTX25MSRICLoRALoader, strength 1) plus the MSR Multi Reference Guide, with an image reference, a scene image as background, and an audio reference for the voice
- Kijai's PromptRelayEncode with a global prompt and two local prompts
- SamplerCustomAdvanced, exp_heun_2_x0_sde, ManualSigmas 1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0 (8 steps), fixed seed
- 10 second clips, 241 frames at 24 fps
What I'm making: a character seated at a desk, speaking to the camera. Identity and voice come out great. The problem is the camera: it slowly pushes in over the clip even when nothing in the prompt asks for it.
What I've tried, all on the same seed:
- CFG 1, 1.5 and 2: no change in the push-in
- "camera zoom" in the negative prompt: no effect. I later learned "zoom" isn't a recognized camera term, only push in and pull back
- "static frame" in the prompt (twice): still drifts. It seems to leave the timing of pushes and pulls to the model rather than locking the camera
- "pull back" in a local prompt: the camera did pull back, then pushed in again on its own
- "wide establishing shot" added to the local prompts: same result
- "no push in": no help
I know the official camera guide lists: tracks, pans across, circles around, tilts upward, pushes in and pulls back, overhead view, handheld movement, over-the-shoulder, wide establishing shot, static frame. It would be a big help. Thanks!doesn't say how any of them behave in a long clip.
Questions:
- Is there a keyword, prompt structure, or setting that actually holds a fixed camera for a 10 second talking shot?
- Could the MSR LoRA or Prompt Relay be adding the drift?
- Is there any way to control how much of the clip each local prompt gets?
I'd rather not fix it by keyframing in post. Any working prompts or workflows for a static seated speaker)
Any help REALLY appreciated!
here is my help, I've tried this model some time ago when i was less experimented with ai video models in general. I'm now using minimaxh3 because it has the reference to video mode which is very powerful even if the human realism quality is less than what this ltx model can do out of the box. BUT here is what I've figured out related with your question. both minimaxh3 and this model are missing the part that controls and interact with the model to make it do what you want. we may call this part "THE DIRECTOR" the module that can write the prompt like the model wants so that the camera will move at specific speeds and keep the angle or make certain things happen in the scene exactly as you want. this very important module is not open source and is using by these companies to make money with their website. now here is the good thing IF you figure out the secret sauce to talk with the model then you won the lottery as the models are very powerful, here is the bad news, no one actually knows or at least will not reveal the correct prompt layout. Also I've tried for example google argon at highest settings and it doesn't know the correct prompt structure, it just gives you the bad information that is already on internet. So no help from AIs either, they are wrong, don't trust them.