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
Ideas
How are the captions being made for the dataset? Is it Qwen? or human captioning? I think human captioning is far more powerful than using ai to auto-caption. If Qwen captions a video it usually leaves out the actions and describes one frame of the video in my experience using fizgig to create Loras. If a man is sitting at a bench with shrubs or bushes behind him, stands up and walks to the left of the frame, Qwen will say a blond man sitting on a bench, describe his clothing, with a blurry background. It will leave out the action. Also, it seems that it was trained on up-close videos of people only, as the camera zooms out and away, the faces lose integrity like it never saw that before. It should be trained on individuals (and maybe groups) approaching the camera, and up-close videos of people where the camera zooms out and away, or people walking backwards. I feel like this careful engineering could improve the model greatly. It could also help to train it on individuals who leave the frame and return, so that it can remember faces better (only a guess, I'm not really sure that would work)