MiniMax / h3-r2v

## Overview MiniMax H3 R2V is MiniMax’s reference-to-video generation model. It reads your prompt plus optional reference images, video clips, and audio clips in one context. It generates a short MP4 video with native stereo audio, not a separate audio pass. This helps when you need stable identity, style, motion, and voice across a single shot. ## What you can build - Product hero shots that keep the product design identical while the camera moves - Character-consistent clips from multiple reference images (outfit, face, props) - Motion transfer using a short reference video as movement or pacing guidance - Style transfer using a style board image plus a separate subject image - Video edits driven by a source clip, like replacing a subject while keeping motion - Dialogue and sound effects timed to the action inside one generated clip ## Inputs - Reference images you want the model to follow, such as a subject, outfit, or style board (JPG, PNG, WEBP, HEIC, or HEIF). You can upload up to 9 images. - Reference video clips you want to use for motion, pacing, or as a source for editing (MP4 or MOV). You can upload up to 3 clips. Each clip must be 2–15 seconds, with 15 seconds total. - Reference audio clips you want to use for voice or sound guidance (WAV or MP3). You can upload up to 3 clips. Each clip must be 2–15 seconds, with 15 seconds total. Audio can’t be the only reference. - A text prompt that describes the target video. Cite your uploads by their order, like “Image 1”, “Video 1”, and “Audio 1”. - The target clip length in whole seconds. It must be between 4 and 15 seconds. - An aspect ratio choice. You can let the model adapt to your references or force a fixed ratio like 16:9 or 9:16. - An output resolution choice. You can generate at 768P or 2K. ## Outputs The model returns a single MP4 video. The file includes the video stream plus a synced stereo audio track. Video is generated at 24 fps. Audio is 32 kHz stereo. The duration matches your requested 4–15 seconds. ## Limitations - Maximum clip length is 15 seconds. - Reference inputs have hard caps: 9 images, 3 videos, and 3 audio clips, with 12 files total. - Audio references must be paired with at least one image or video reference. - Reference-to-video prompting is sensitive to how you name and assign each reference. Vague prompts can ignore parts of your intent. - Conflicting references can cause identity drift, style blending, or sudden changes between shots. - Low-quality inputs can degrade results. This includes compressed videos, noisy audio, low-resolution images, and scanned or heavily filtered assets. ## Safety & compliance MiniMax applies automated moderation to user inputs and generated prompts. Content suspected of being unlawful, pornographic, or infringing third-party rights may be blocked. Filtering can still produce false positives and false negatives. You are responsible for lawful use, rights clearance, and following the MiniMax H3 Community License terms.