MiniMax H3: Image to Video
Turn a starting image into a short, high-quality video clip with MiniMax H3, a unified multimodal model that generates visuals and synchronized audio in a single pass. The workflow takes 1 input image and produces 1 video output, with optional text or reference media to steer motion, style, or sound. Ideal for rapid concept visualization, character-driven animation tests, and creating shareable social media clips with built-in sound design.
试试一个示例
关于此工作流
This workflow turns a single still image into a short, coherent video clip with synchronized stereo audio using MiniMax H3. At its core is the MiniMax H3 generation node (id: 4c314f31-ecda-4b08-ae98-faaba1bf613f), which performs a single forward pass to synthesize both video frames and audio that match the scene. The image you provide through LoadImage sets the subject and overall style; you can optionally add short text guidance inside the MiniMax H3 node to suggest motion, pacing, or mood.
To help the model run efficiently and preserve composition, the workflow routes the input through GetImageSize and a Use Image Size group that keeps aspect ratio intact. ResolutionSelector and ImageScaleToTotalPixels work together to pick a target size that fits MiniMax H3’s recommended megapixel budgets and the multiple-of-32 constraint required by the model. The generated clip is then written out by SaveVideo as a single MP4 containing the video stream (and the model’s generated audio, if enabled), ready for quick review or sharing.
Model assets used by the generator include the video VAE (minimax_h3_video_vae_fp16.safetensors) and the audio VAE (minimax_h3_audio_vae_fp32.safetensors).
常见问答
Download the MiniMax video VAE (minimax_h3_video_vae_fp16.safetensors) and audio VAE (minimax_h3_audio_vae_fp32.safetensors) from the links in the workflow notes. In the MiniMax H3 node (id: 4c314f31-ecda-4b08-ae98-faaba1bf613f), set the paths if the node exposes them, then reload models if required.
Use ResolutionSelector with a megapixel budget that matches your GPU VRAM. The ImageScaleToTotalPixels node will output dimensions that are multiples of 32 and preserve your input image’s aspect via GetImageSize. If you encounter OOM errors, lower the MP target or shorten the duration.
Yes. In the MiniMax H3 node, disable audio generation (if exposed) or ignore the audio output when wiring to SaveVideo. The SaveVideo node will still write the video stream as MP4.
Your input image anchors style and identity. For motion, add concise guidance in the MiniMax H3 node (e.g., camera moves, tempo, ambience). Keep prompts short and concrete. For stronger effects, try a higher duration or different seed; for more subtle motion, reduce duration or simplify the guidance.