There is a visual result issue in V 2.1.66

[kubilaykilinc]

I run Fooocus with standard run.bat. When I use an SD1.5 model as a refiner, these types of textures occur. I'm using standard settings. When I use an SD1.5 refiner as a refiner, textures begin to appear in the image and deformation occurs.

I run Fooocus with standard run.bat. When I use an SD1.5 model as a refiner, these types of textures occur.

I used the SDXL base 1.0 model as the base model. I used RealisticVision5.1 as Refiner. As for Style, only Fooocusv2 was active. No other styles were active.

This problem did not exist before 2.1.66.

[lllyasviel]

give prompt seed and model name

[kubilaykilinc]

My Hardware : Windows 11 - 32 GB Ram - AMD Ryzen 5 5600G with Radeon Graphics - NVIDA RTX 3060 12GB (Latest Driver 537.58) - 1 TB SSD

Fooocus Version 2.1.66 Running - run_realistic.bat

Base SDXL model: realisticStockPhoto_v10.safetensors Refiner model : (SD 1.5 model) realisticVisionV51_v51VAE.safetensors

LORA 1 : SDXL_FILM_PHOTOGRAPHY_STYLE_BetaV0.4.safetensors LORA 1 weight : 0.25

Positive Prompt a vintage style Olive oil product in a vintage style room

Negative Prompt (Standart Fooocus prompt) unrealistic, saturated, high contrast, big nose, painting, drawing, sketch, cartoon, anime, manga, render, CG, 3d, watermark, signature, label

Seed : 0

Other settings are standard fooocus settings (style, cfg level, developer debug mode etc.) FreeU mode : disable

and this is result

I tried this setup with another SD.1.5 model as refiner. I am getting same issue.

This is same settings but no any refiner.

[kubilaykilinc]

All the setup I mentioned above is the same. seeds 0 prompt an apple

A. Base SDXL model: realisticStockPhoto_v10.safetensors Refiner model: (SD 1.5 model) realisticVisionV51_v51VAE.safetensors

B. Base SDXL model: realisticStockPhoto_v10.safetensors Refiner model: (SDXL model) sd_xl_refiner_1.0_0.9vae.safetensors

I use SD 1.5 refiners for better photorealistic results. When I use any SDXL model as a refiner. No problem. However, if the refiner is SD1.5 model. There is this problem.

[lllyasviel]

thanks will take a look soon

[lllyasviel]

This is the default result If you use another sd1.5 as refiner, result will be a bit worse

But I cannot reproduce your examples.

The solution is to just do not use 1.5 refiner for realistic XL. The 1.5 refiner only make things better for anime images in most cases.

Previous 2.1.60 has same problem but less obvious. Since this 1.5 refiner is mostly for anime, the implementation will mainly target improving anime images. This problem does not exist in anime images.

[lllyasviel]

but i will take a look again since this is very strange

Edit - I think I find the cause, but i am still considering what should be a correct formulation - I will probably fix this in several days

[lllyasviel]

see if fixed in 2.1.67