V2.0

Main model used: flux2 klein base 9b

Steps: 8 20 (If it doesn't work, first increase the lora strength. If it still doesn't work, add steps. Default is 8 steps.)

CFG: 1 (Do not add accelerated lora or other loras, otherwise it will affect the model weights.)

Model weight: 0.3 1 (Default recommendation is 0.75)

The lora weight and steps for each image are different because the texture depth, lighting, and angle of each face in the images vary. Therefore, there are no overly uniform parameter recommendations, only approximate values. Adjust the specific values according to your needs.

This lora is mainly used to add facial texture details, but there is a prerequisite: the original image must be AI-generated (some real photos show effects, while others have severe redness issues; in any case, real photos are not very stable). Ideally, the image should be around 1k resolution with a clearly visible face (side profile, tilted, or looking up/down are all fine). The facial texture details in the photos I generated with GPT Image 2 were insufficient, so I trained this lora to enhance facial texture. This lora provides native enhancement for images generated by GPT Image 2. It is only effective for images below 2k resolution!! Input images above 2k will inherently show less effect, which is a limitation of open-source models. Surprisingly, when I input a 4k image, the facial texture was predictably reduced, but it also refined the facial texture, giving a more natural feel, which can be considered a workaround for fixing some messy texture details. Nevertheless, I still do not recommend images above 2k resolution.

Model address: https://www.runninghub.cn/model/public/2072311076204662785