The VAE used by Qwen Image 2.1, hosted separately so any Qwen Image model - ours or a community finetune - can point at it instead of shipping its own copy. One file, BF16, about 644 MB.

Originally released by the Qwen team on Hugging Face, and repackaged as a single file for ComfyUI by Comfy-Org. All credit for the weights goes to the Qwen team.

What you need it for

Running Qwen Image 2.1 locally, or running a finetune whose author did not bundle a VAE. It is unchanged from the upstream release, so it is the same file already included on the Qwen Image 2.1 model page - this page exists so it can be referenced on its own.

Where it goes

In ComfyUI, models/vae/. This is a component rather than a runnable model, so there is no Generate button and no on-site generation - if you want to generate with Qwen Image, that is the Qwen Image 2.1 page, where the VAE is already handled for you.

The matching text encoder

Qwen Image 2.1 runs on the Qwen3-VL 8B encoder, which has its own page: Qwen3-VL, in BF16, FP8, NVFP4, int8 and w4a8. Pick one precision there to go with this VAE.

Links

• Source: Qwen on Hugging Face
• Repackaged file: Comfy-Org/Qwen-Image-2.1
• Text encoder: Qwen3-VL
• Checkpoint: Qwen Image 2.1