Kandinsky 6.0 Video

FreeVideo is an open-source local inference engine and ComfyUI plugin for generating videos with the MiniMax H3 model on consumer hardware. Built on OpenVDN's eight-step VDN-H3 model and Video DeltaNet's hybrid attention, it adapts weight placement, compute precision, attention kernels and acceleration to available hardware. Weight streaming, asynchronous prefetching and chunked computation reduce peak memory use, with the project reporting operation on systems with as little as 8 GB of VRAM and 16 GB of RAM. It accepts text prompts, first and last frames, and image, video and audio references, and supports MiniMax H3 LoRAs, two-pass sampling, batch generation, four quality levels and preview-first generation. Windows and Apple silicon launchers configure ComfyUI, the runtime and models, while Linux supports setup through a shell script and command-line generation. The code is licensed under Apache License 2.0; the model weights use the MiniMax H3 Community License.

MiniMax H3 is MiniMax's general-purpose omnimodal generative system for understanding text, images, video, and audio and generating video with native 32 kHz stereo audio. It supports 4–15-second outputs at 24 FPS across a wide range of aspect ratios; the shorter side is 768 pixels by default, with 2K generation available through H3-Regenerate-2K. Its H3-Base-FL2VA variant supports text-to-video, first-frame-to-video, last-frame-to-video, and first-and-last-frame-to-video modes, while H3-Base-Ref2VA accepts mixed image, video, and audio references. The repository also provides a prompt-writing skill and style-specific video-generation skills for the MiniMax Hub canvas workflow. H3 is available through MiniMax's global and China APIs, web apps, and desktop apps; the video describes its weights as open and runnable at 768p.

ComfyUI is an open-source AI content-creation engine and diffusion-model GUI, API, and backend developed by Comfy-Org. Its visual node-graph interface lets users build and reuse no-code workflows for image, video, audio, 3D, and text generation, editing, and processing, including workflows using diffusion models, VAEs, text encoders, LoRAs, ControlNets, adapters, and upscalers. Workflows can be saved and loaded as JSON, exposed through App Mode, or integrated into applications through local API endpoints. The engine supports asynchronous queueing, partial graph re-execution, VRAM and RAM management, model offloading, quantized models, custom nodes, and fully offline core operation unless optional resources are requested. It can run locally on Windows, Linux, and macOS through desktop, portable, or manual installations, with support for NVIDIA, AMD, Intel, Apple Silicon, and Ascend hardware. Comfy-Org also provides an official paid cloud version.

An AI video-generation model from ByteDance's Seed team that creates cinematic, multi-shot videos from text prompts, images, video clips, and audio. It supports native audio-visual synchronization and can generate complete 2D animation story videos from storyboard images and detailed video prompts.

MAGI-2 Preview is the inference implementation for SandAI’s unified audio-video generation model. The 114-billion-parameter architecture, built on MagiMoE, activates about 6 billion parameters per token and generates 10-second clips from text prompts (T2V) or a prompt plus a still image (I2V), with sound generated alongside the video and muxed into the output file. Generation runs in two stages: `magi2_preview` denoises the clip at low resolution, after which `magi2_refiner` increases it to 1080p. The current base release is not step-distilled, so denoising steps account for most of the generation time; a distilled release is described as forthcoming. The repository contains inference code, while the model weights are downloaded separately from the `sand-ai/MAGI-2-preview` Hugging Face repository and occupy hundreds of gigabytes in total. An optional prompt-enhancement stage sends the input to an OpenAI-compatible instruction-following LLM, which rewrites it as a structured JSON caption for the 10-second clip, renders that result as Markdown, and passes it to the generation pipeline. It can be disabled to use the raw prompt. The documented runtime requires eight NVIDIA Hopper GPUs, Python 3.12, a recent CUDA toolkit, and FFmpeg for audio-video muxing; Docker images and a source installation are provided.