生成式影片
MiniMax H3 Opens Video Generation Weights, ComfyUI Simultaneously Provides a Runnable Workflow
MiniMax has released the weights for its H3 video model, enabling developers to run image-to-video generation with synchronized audio locally. ComfyUI has added an official workflow, but consumer GPU tests largely rely on pruned and INT8 derivative files and should not be treated as full-quality reproductions of the original weights.
After launching the H3 audiovisual generation model, MiniMax uploaded its model weights to Hugging Face on August 3, turning its earlier promise of a release “within days” into a downloadable artifact. H3 can generate video from text, a starting image, or reference material, incorporating visuals and stereo audio into a single generation process. Unlike traditional pipelines that first generate silent video and then connect speech or sound-effect models, H3 handles temporal synchronization directly within the same model.
The practical value of the open weights is that usable workflows are already emerging. ComfyUI’s official template library has added `video_minimax_h3_i2v.json`, connecting model loading, text encoding, image conditioning, sampling, and audiovisual output nodes into an inspectable graph. Developers can therefore swap samplers, adjust resolution and frame length, or integrate H3 into existing asset-management and post-production pipelines without relying on MiniMax’s hosted API.
However, “downloadable” does not mean the original model can run unconditionally on typical graphics cards. Many community tests conducted on the first day used pruned, INT8, or rotation-quantized derivative files. In one reported case, an RTX 4090 Laptop GPU took about three minutes to generate a five-second video at 960×540. These results demonstrate that local inference is feasible, but they also reflect the effects of quantizers, disk offloading, node implementations, and unofficial checkpoints. Their memory usage and speed cannot be directly applied to the original BF16 weights, nor can they serve as a basis for comparing quality with the cloud API.
Research and product teams should next verify whether the license covers commercial deployment and review restrictions involving training data and human generation. They should also establish side-by-side comparisons of the API, original weights, and quantized versions using identical prompts. Technical evaluations should focus particularly on character consistency across long sequences, lip-sync and audio synchronization, object persistence after camera cuts, and whether the audio produces hallucinatory dialogue. If community inference frameworks can provide stable support, H3 could become one of the few fully self-hostable foundation video models capable of generating both visuals and audio simultaneously.