ltx2

Run the ltx2.py tool with a --prompt describing the scene and an --output path, for example a cinematic description of a sunset over the ocean. The script calls the Modal endpoint and saves an H.264 MP4 clip of about 5 seconds by default.

Use image-to-video mode by passing --input with your image file alongside a motion-focused prompt like "gentle camera drift, soft ambient motion". The model animates the still image into a short MP4 clip at the specified resolution and frame count.

Frame counts must satisfy (n-1) % 8 == 0, giving valid options like 25, 49, 73, 97, 121 (default, ~5s), 161, and 193 frames (~8s max). Common resolutions include 768x512, 512x512, 1024x576 widescreen, and 576x1024 portrait, with dimensions divisible by 64.

No, the model cannot reliably render readable text in video, so use Remotion overlays for titles instead. Clips are limited to about 8 seconds (193 frames), so longer content requires stitching multiple generated clips together.

Roughly 30% of generations may include logos or text artifacts inherited from training data. Re-run the generation with a different --seed value to get a clean result, and use the negative prompt defaults to reduce quality issues.

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 your .env file. The HuggingFace token needs read access and acceptance of the Gemma 3 license.