Guide : SDXL + inpainting

[glucauze]

I have been testing SDXL with post-inpainting. The idea is to use the refiner in post-inpainting. Non cherry picked example.

Here is the prompt :

raw photography, portrait of Alan Turing
Negative prompt: blurry, bad art
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 2929728811, Size: 1024x1024, Model hash: 31e35c80fc, Model: sd_xl_base_1.0, Version: v1.5.1

Here is the configuration :

Here is the original :

Here is without inpainting :

Here is the result

Note that he prompt is important. Here is another try :

with the following inpainting prompt + codeformer instead of gfpgan:

pos : cinematic photo of a [gender],  35mm photograph, film, bokeh, professional, 4k, highly detailed
neg : drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly

And here another one :

pos : cinematic photo of a [gender],  close portrait, 35mm photograph, film grain,  professional, 64k, highly detailed, ((natural skin)), ((skin pores)), ((realistic skin))
neg : drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly

0.3 seems to be the limit for denoising strenght

The usefulness of inpainting really depends on what you want to do. For reference it is running on GPU 12Gb of vram, but it reaches the limit.

[glucauze]

Here is another example :

cinematic photo of Alan Turing, close portrait, 35mm photograph, film grain, professional, 64k, highly detailed, ((natural skin)), ((skin pores)), ((realistic skin))
Negative prompt: blurry, bad art
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 2929728811, Size: 1024x1024, Model hash: 31e35c80fc, Model: sd_xl_base_1.0, Version: v1.5.1

Original :

Result :