MiniMax H3 与 fal 联合直播：开放权重多模态视频模型实战

MiniMax has introduced H3, its latest multimodal AI model capable of understanding text, images, video, and audio within a unified context to generate high-quality videos with native stereo sound. The model supports up to 15-second videos in 2K resolution, offers advanced editing capabilities such as motion transfer, image-to-video, and video-to-video generation, and is designed for commercial applications including advertising, e-commerce, product design, gaming, and UI/UX. MiniMax also plans to release H3 as an open-weight model, giving developers greater flexibility to customize and deploy it. The company claims H3 delivers industry-leading price-performance, producing premium-quality videos at a significantly lower cost than competing models. H3 is built on a unified multimodal [...] than competing models. H3 is built on a unified multimodal architecture that improves consistency across visual and audio generation, reducing artifacts and enhancing realism. It also supports faster i

## Common questions about MiniMax H3 MiniMax H3 is an open-weights, general-purpose multimodal video model. Instead of a separate model for each task, MiniMax H3 reads text, images, video, and audio in one unified context and generates coherent audiovisual results from any mix of them. It supports text-to-video, first-and-last-frame, reference-to-video, and precise video editing. MiniMax H3 is released with open weights, so it is an open foundation you can explore, customize, and build on rather than a closed endpoint. fal is a Day 0 ecosystem partner, which means you can call the hosted MiniMax H3 API on fal.ai from launch without provisioning GPUs, and still have the option to work with the weights directly for your own research and fine-tuning. [...] ### Sound Composed to Picture Every generation returns native stereo audio: original score, dialogue, foley, and room tone timed to the cut. Give MiniMax H3 a reference recording and it will transfer or clone that voice onto your cha

AIH3MultimodalVideo Generation Today, we're launching MiniMax H3, a general-purpose multimodal generation model. H3 understands unified context across text, images, video, and audio, generating video with native stereo sound, up to 15 seconds at 2K resolution. Early testing shows H3 is ready for commercial content creation across a wide range of use cases, excelling at instruction following, accurate text and brand rendering, and V2V motion transfer. With precise, controllable multimodal generation and editing, H3 is built for advertising, branding, e-commerce, product design, UI/UX, gaming, and more. [...] To get there, we built dedicated models and a full-modality understanding pipeline. Most source material requires around 100K tokens of inference, distilled down to an average of roughly 4K tokens. H3-VAE: A Major Boost in Architectural Efficiency The H series has continuously pushed forward on tokenizer technology. With H3, we completely overhauled our previous tokenizer, achie

H3-Context-IR: As inputs become increasingly complex, we build a dedicated system to deeply understand and refine the input multimodal instructions, then convert them into a form that H3 can readily understand—the Context Intermediate Representation—for generation. H3-Context-IR is critical to the quality of the final output, so we strongly recommend incorporating it into your generation pipeline or following the “Prompting Guidance” to build your own context-processing system. H3-Base: Generates audio and video based on the H3-Context-IR output, producing results at 768p resolution. [...] MiniMax H3Video GenerationOpen Source Today, we are officially open-sourcing MiniMax H3, our next-generation general-purpose video model. ## System Overview MiniMax H3 is a general-purpose, omni-modal generative system. It supports unified understanding of multimodal contexts composed of text, images, video, and audio, and can generate video with native stereo audio at resolutions up to 2K and du

The model accepts any combination of text, images, video clips, and audio as input simultaneously. Specifically, it accepts up to nine reference images, three video clips, and three audio tracks in a single generation request — each contributing a different layer of creative control. A production team can supply character reference images, a video sample showing desired camera movement, and an audio reference defining the soundscape, and H3 produces a clip incorporating all of them in a single pass. MiniMax calls this the "omni-reference" system. The model also supports instruction-based editing: a creator can specify a change to part of an existing clip — swap a product, rewrite visible signage, change a background — and H3 applies the edit while leaving the rest of the frame intact. [...] MiniMax H3 is a multimodal video generation model that accepts text, images, video clips, and audio as simultaneous inputs and produces native 2K resolution video at 24fps with synchronized stereo

Fill (@MachineDelusions) shows us some crazy ComfyUI workflows where he is using it for beat synced music/video production, and shows how you can really treat this model like a Swiss army knife, and that we are only scratching the surface. Links + Notes 📝 Join Fine-Tune Fridays 🔧 Discord 🗿 Use Oxen AI 🐂 Oxen.ai gives developers, creators, and teams easy access to the latest AI models, plus the data infrastructure to organize, version, collaborate on, and customize the data behind them. Use 200+ image, video, audio, and language models through one API. Save prompts, reference assets, generations, metadata, and training data into collaborative repositories. Track every change, branch experiments, and fine-tune custom models on your proprietary data.