Qwen Image 2.1 VAE

The original Qwen Image 2.1 VAE, unchanged, on Civitai.

What it is

• The original Qwen Image 2.1 VAE: 64-channel latents, 338M parameters, bf16, 644 MB

• Same file name as in ComfyUI's official templates, so they load it as is

• Works with :

  • Qwen Image 2.1, base and turbo, for text-to-image and editing (it encodes as well as decodes)

  • Noct Q, the uncensored Qwen Image 2.1 in Turbo and Base versions: https://civarchive.com/models/2958896

  • Not with Qwen Image 1 or Krea 2: they use the 16-channel qwen_image_vae

Install

Put it in ComfyUI/models/vae/ and select it in the Load VAE node (ComfyUI 0.37.0 or newer)

On 8 GB cards, decode large images with VAE Decode (Tiled): tile 1024, overlap 128

About the images

Rendered with Noct Q V2 Turbo (5 steps, 832x1248) and decoded with this file. Raw decodes, no upscaling.

中文：Qwen Image 2.1 原版 VAE，未经修改（bf16，644 MB），与 Comfy-Org 发布的文件逐字节一致。适用于 Qwen Image 2.1（文生图与编辑）和 Noct Q，不适用于 Qwen Image 1 和 Krea 2。放入 ComfyUI/models/vae/。仅限非商业用途。

License

Built with Qwen. Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved. Free, non-commercial use only. LICENSE and NOTICE are in the zip.