The vae model and the Qwen-image vae are very different
#11
by taoliu123 - opened
The qwen_image_vae.safetensors here are very different from the Qwen-image vae, as is the structure of the wan_2.1_vae.safetensors model, is this a bug or is it intentional?
It is indeed wan2.1 vae renamed. Why is that? I am unable to get the official VAE from QWEN to work.
The original seems to be in diffusers format, and the ComfyUI version has keys named to support how the VAEs are loaded in ComfyUI.
It is indeed identical in structure to Wan 2.1 VAE, but the weights are slightly different. You can use the QwenImage VAE to decode single frame from WanVideo as well.