ltx2

Run the ltx2.py script with a --prompt describing the scene and an --output path, for example a cinematic sunset description saved as an MP4. The script calls the deployed Modal endpoint and returns a roughly 5-second H.264 clip.

Pass the image via the --input flag along with a motion-focused prompt such as gentle camera drift or subtle head movement. The model performs image-to-video generation and outputs an animated MP4 at the chosen resolution.

Common resolutions include 768x512, 512x512, 1024x576, and 576x1024, with dimensions divisible by 64. Frame counts must satisfy (n-1) % 8 == 0, ranging from 25 frames up to a practical maximum of 193 frames, about 8 seconds.

No, the model cannot reliably render readable text in video, so use Remotion overlays for titles. Clips are limited to about 8 seconds each, and longer content requires stitching multiple generated clips together.

Roughly 30 percent of generations may include logos or text artifacts inherited from training data. Re-run the generation with a different --seed value to obtain a clean result.

Create a Modal secret with a HuggingFace token, deploy the Modal app which downloads about 55GB of weights, and save the endpoint URL as MODAL_LTX2_ENDPOINT_URL in the .env file. The HuggingFace token needs read access and the Gemma 3 license must be accepted.