https://huggingface.co/spacepxl/Wan2.1-VAE-upscale2x/tree/main
I saw someone complaining that the VAE of Qwen Image is very poor. So I conducted a test.
Finally, the aforementioned trained VAE was chosen. The 'decoder.head.2' was averaged along the RGB channels (to be able to use it on Krea 2 without the need for additional custom nodes). Then, the decoder part was merged with the original VAE. Compared to the original VAE, there are only differences, which are suitable for those who aim to achieve realistic style of character skin details.
Description
FAQ
Comments (14)
my eyes are not that keen to differentiate b/w example images
Thanks, 100% better than the original. The difference is even more apparent after a massive upscale (like using ultimate sd ulscale).
Do you know if you need to install anything else aside downloading this VAE? for some reason my comfyUI doesn't like it and throws an error
@yamstapia316 no clue i just load it with the default node as usual
@yamstapia316 Did you use that huggingface link to download it? There will be additional nodes required there, and the image pixels will be enlarged by a factor of two. And what you download on the right side of the page will work properly in the comfyui.
works very well, especially if followed by a SeedVR2 upscaling step.
PS: you might want to convert this to BF16 to save on file size, since that's what ComfyUI will convert it to anyway. I tried this locally and there is no difference between FP32 and BF16
Comfy can be set to FP32 VAE.
This is a significant improvement!
How is it compared to using wan2.1 vae?
this works so good with SeedVR2
Add too much details, make the results not natural. Didn't tried seedvr2 on it, but when you generate in 2048, the results are already sharp and don't need upscale, so original vae feel more natural in that case. Do you recommend seedvr2?
Black color is crashing :( Best is qwen_image_vae. But if you want just a pinch softer image use Wan2_1_VAE_fp32.safetensors for KREA2