At the time of writing, the Comfy-Org repository showed more than 4.2 million downloads on Hugging Face. The repository collects Wan 2.1 checkpoints, reduced-precision weights, adapters, and workflow components under ComfyUI’s standard directory structure. Developers gain one source for text-to-video, image-to-video, controlled editing, transparent video, and audio-driven animation pipelines.

Wan 2.1 under the hood

Alibaba Cloud released Wan 2.1 in February 2025 as an open-source family of four foundation-model checkpoints. Its diffusion transformer generates video by iteratively removing noise from a compressed representation of the clip. A 3D causal variational autoencoder called Wan-VAE handles that compression across space and time, reducing the memory required to process full-resolution frames.

According to Alibaba’s launch post, Wan 2.1 scored 86.22% overall on VBench and was the only open-source model among the benchmark’s top five at publication. VBench evaluates qualities such as subject consistency, motion, spatial relationships, color, and prompt alignment. Alibaba also presented Wan 2.1 as the first video model to support text effects in both Chinese and English.

One tree, eight workflow families

Comfy-Org consolidates files from Wan-AI, Alibaba PAI, MAGREF-Video, htdong, MeiGen-AI, and related projects. The repository includes fp16, bf16, and scaled fp8 builds where available, allowing developers to balance compatibility, memory use, and checkpoint size.