MiniMax / h3-i2v

## Overview H3 I2V is MiniMax’s image-to-video model from the MiniMax H3 system. You provide a starting frame plus a written scene description. The model generates a short MP4 video clip with synchronized stereo audio. It’s useful when you need realistic motion, camera intent, and sound without manual animation. ## What you can build - Animate a product photo into a short ad shot with subtle camera movement - Turn character art into a short cinematic clip with controlled actions - Create first-to-last-frame transitions for scene changes and match cuts - Generate short storyboard shots with a consistent starting composition - Produce short talking scenes with scripted on-screen dialogue in supported languages ## Inputs - A first frame image as a JPG or PNG file to lock the opening composition and subject identity - An optional last frame image as a JPG or PNG file to steer the ending pose or scene state - A required written description of the scene, including actions, camera framing, and audio intent - A clip length in whole seconds, between 4 and 15 seconds - An output resolution choice, with a 768p-style base output and an optional 2K mode - An optional aspect ratio selection for text-only generation when you don’t provide an image ## Outputs - An MP4 video clip - 24 FPS video - 32 kHz stereo audio embedded with the video - 4–15 seconds duration - A resolution where the shorter side is 768 pixels by default, with an optional 2K workflow ## Limitations - Clip length is limited to 4–15 seconds. - Frame rate is fixed at 24 FPS. - The standard generation path produces a 768p output where the shorter side is 768 pixels. - 2K output requires an additional regeneration stage in the full H3 workflow. - Prompt quality matters. Vague motion descriptions often produce weak or random movement. - Low-quality, heavily compressed, or blurry input images can cause identity drift and artifacts. ## Safety & compliance - MiniMax H3 is released under the MiniMax H3 Community License Agreement, which includes territory limits for open-weight use. - The license defines an Applicable Territory that excludes the European Union, the United Kingdom, the Republic of Korea, and the United States of America. - You must follow the Acceptable Use Policy. It forbids harmful use, bypassing safeguards, exploitation of minors, malware generation, and doxxing. - The policy also forbids generating or spreading election-influencing misinformation and fake online engagement. - If you publish outputs in public, you must clearly disclose that the content is machine-generated. - The license forbids using the model, its outputs, or results to improve other AI models, except MiniMax H3 derivatives.