ComfyUI 提供可编辑的 MiniMax H3 图，用于本地视频工作流
Yannik Marek 的基于节点的引擎新增了对 MiniMax H3 原生工作流程的支持，为创作者提供了用于该开放权重视频模型的可检视图表。
By Ryan Merket · Published
Primary source: ComfyUI Newsletter
Why it matters
ComfyUI is positioning its node-based engine as the execution layer between open-weight media models and creators, where memory optimization, editable templates and hardware support determine whether downloadable weights become usable production tools.
ComfyUI, the node-based AI creation engine created by Yannik Marek, added native MiniMax H3 workflows on Aug. 2, turning the video model into editable graphs with local-runtime support. ComfyUI says the integration can run locally on Nvidia RTX 3060-class hardware.
The release adds the workflow layer around MiniMax's H3 weights, including editable templates and local-runtime support. RuntimeWire covered MiniMax's checkpoint release, 2K service path and community license separately. Here, the news is how ComfyUI has made the model inspectable and configurable inside its node-based engine.
Marek began ComfyUI in early 2023 as a solo Stable Diffusion experiment. ComfyUI says the project started when no available tool could chain two AI models into a repeatable workflow, prompting Marek to build and open-source his own node-based system. The resulting interface exposes models, parameters and processing stages on a graph rather than hiding them behind a single prompt box.
That structure makes rapid model integration central to ComfyUI's product. ComfyUI's announcement says H3 accepts text, images, video and audio. MiniMax's H3 model card says the model generates clips at 24 frames per second and up to about 15 seconds, with native stereo audio. ComfyUI packaged those capabilities into workflows for text-to-video, image-to-video, first-and-last-frame generation and reference-driven video. Reference media can carry a subject, motion or voice into a new clip, according to the ComfyUI announcement.
ComfyUI's H3 documentation requires version 0.30.0 or later and lists text-to-video, image-to-video and reference-to-video templates. Developers can inspect and modify each graph, replacing a model or processing stage without reconstructing the entire workflow.
为什么 ComfyUI 的图很重要
ComfyUI 表示 它已针对本地使用对 H3 进行了优化，并将该模型配对了现成的工作流，而不是让开发者自行组装图表。该公告支持 RTX 3060 兼容性的说法，但并未证明不同 GPU 之间的生成速度或输出质量具有可比性。
开发者可以使用 ComfyUI 的 H3 文档 从文本到视频、图像到视频和参考到视频的图表开始。由于 ComfyUI 将管线以节点形式暴露，创作者可以更改单个模型、参数或处理步骤，重新运行图表并保留其余过程。
本地工作流有其限制。MiniMax 的 官方模型卡 表示 H3-Base 在 768p 生成，而文档化的完整 2K 流程则将本地部署的基础模型与托管的 H3-Context-IR 和 H3-Regenerate-2K 服务结合。ComfyUI 的 教程 描述了一个原生画布，短边为 768 像素，上限为 768×1,344 像素。
MiniMax 将 H3 描述为开放权重，但其条款比不受限制的开源许可更为限制。 MiniMax H3 社区许可协议 对地理区域施加了限制，并且当使用 H3 的商业产品或服务产生超过 2000 万美元的年收入时，需要单独的书面授权。
ComfyUI 的工作流押注
ComfyUI 正在为视觉模型、预处理步骤和后期制作工具构建一个通用执行层。 ComfyUI 在 4 月表示 它筹集了 3000 万美元，使总融资达到 4700 万美元。 TechCrunch 报道 此轮估值为 5 亿美元。Craft 主导了这笔融资，Pace Capital、Chemistry 和 TruArrow 参与。ComfyUI 报告当时拥有 400 万用户、超过 60,000 个社区构建的节点和超过 150,000 次的日下载量。
H3 的整合展示了 Marek 的图模型如何在同一天吸收新发布的检查点并将其转化为可修改的生产流程。MiniMax 提供了模型。ComfyUI 构建了可编辑的模板和本地执行路径，使其能在现有的创作引擎中运行。