Quality of the image is poor.

[sunzj]

I try sd1.5 and SDXL, the result seems dont as i expected.

i tried the following prompt on sdxl, https://anakin.ai/blog/sdxl-photorealistic-prompts/

expected result are:

but the result are blur and low quality.

I am not sure it's my config problem or it's a common issue.

[happyme531]

I think on SDXL the issue is it is using vae config of sd1.5

https://github.com/happyme531/RK3588-stable-diffusion-GPU/blob/37957e25e3af6d959165829a1483e3c5416fcfc8/web_stable_diffusion/trace/model_trace.py#L93

And on SD1.5 the issue is the vae scale is wrong (see the commit in the above link)

[sunzj]

@happyme531

As i verified, the vae scale should always 1 / 0.18215.

• sd1.5 latents = 1 / 0.18215 * latents vae = utils.get_vae(pipe, "1.5")

• SDXL latents = 1 / 0.18215 * latents vae = utils.get_vae(pipe, "XL")

I think on SDXL the issue is it is using vae config of sd1.5

https://github.com/happyme531/RK3588-stable-diffusion-GPU/blob/37957e25e3af6d959165829a1483e3c5416fcfc8/web_stable_diffusion/trace/model_trace.py#L93

And on SD1.5 the issue is the vae scale is wrong (see the commit in the above link)

[sunzj]

After test, some image quality issue should cause by the model，both 1.5 and xl have these issues. Such as, eyes, nose abnormal for full-body image, and portrait is normal.