MiniMax H3 Day-0 Support in ComfyUI: Open Weights, Native Audio, and 2K Video

4 hours ago

MiniMax H3 是一款开源的全模态视频模型，能够生成高达 2K 分辨率和 15 秒时长的视频，并自带原生立体声音频。
它支持文本转视频、图像转视频、首尾帧控制，以及使用图像、视频或音频作为参考的视频生成。
该模型集成了多模态上下文理解能力，通过同时处理图像、音频和视频，将多个任务合并为一个。
原生立体声音频与视频在同一流程中生成，而非后期处理。
编辑与动作迁移功能允许参考视频的运动轨迹，同时应用不同的主体和风格。
该模型针对 ComfyUI 中的本地推理进行了优化，显存占用减少 66%（从 123.6 GB 降至 42.5 GB），使其能在消费级 GPU（如 RTX 3060）上运行。
优化措施包括剪枝调制权重、int8 卷积量化，以及结合动态 VRAM 内存卸载的自定义内核。
示例输出展示了多种创意应用场景：漫画风格、产品广告、高定时尚大片，以及活力产品推广视频。

Seedance 2.5 is a next-generation audio-video joint generation model designed for 30-second storytelling with precise reference control and editing capabilities.
It allows creation of videos up to 30 seconds in a single generation, extendable twice for richer storytelling, with smoother motion and more realistic visuals.
The model understands reference videos more precisely, capturing intention, framing, and cinematic language for creative interpretation beyond motion transfer.

Seedance 2.5 generates up to 30 seconds of high-quality audio-video in a single pass, supports multi-round extensions, and improves shot transitions and scene changes for coherent long-form storytelling.
Multimodal referencing is fully upgraded, allowing up to 30 images, 10 videos, and 10 audio clips as inputs, with enhanced capabilities for clay render, motion, and creative references.
Editing is more precise with timestamp-level control, green screen, camera perspective, and reference-based editing, supporting professional film and advertising workflows.

Inkling-Small is a general-purpose multimodal model that accepts text, image, and audio inputs and generates text outputs.
It can be used with libraries like Transformers, vLLM, SGLang, and Docker, as well as through APIs and third-party providers.
The model uses a 42-layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) architecture, totaling 276B parameters but activating only 12B per token.

Seedance 2.5 is a next-generation audio-video joint generation model designed for 30-second storytelling with precise reference control and editing capabilities.
It allows creation of videos up to 30 seconds in a single generation, extendable twice for richer storytelling, with smoother motion and more realistic visuals.
The model understands reference videos more precisely, capturing intention, framing, and cinematic language for creative interpretation beyond motion transfer.

Seedance 2.5 generates up to 30 seconds of high-quality audio-video in a single pass, supports multi-round extensions, and improves shot transitions and scene changes for coherent long-form storytelling.
Multimodal referencing is fully upgraded, allowing up to 30 images, 10 videos, and 10 audio clips as inputs, with enhanced capabilities for clay render, motion, and creative references.
Editing is more precise with timestamp-level control, green screen, camera perspective, and reference-based editing, supporting professional film and advertising workflows.

Inkling-Small is a general-purpose multimodal model that accepts text, image, and audio inputs and generates text outputs.
It can be used with libraries like Transformers, vLLM, SGLang, and Docker, as well as through APIs and third-party providers.
The model uses a 42-layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) architecture, totaling 276B parameters but activating only 12B per token.

Hasty Briefsbeta