TUTORIAL #1 - Perfect feet with AI(Stable Diffusion webui 1111) [Update: 2023/12/05]

TUTORIAL #1 - Perfect feet with AI(Stable Diffusion webui 1111)

(The same considerations apply to hands/fingers)

Here, on internet, I have seen feet with 6 toes and toes of wrong lengths.

How can I make perfect feet with AI(Stable Diffusion webui 1111)?

1-Use a very good model

2-Have patience

3-Use inpaint with at least a batch of 4+ (if you have 8gb vram or more)

4-In inpaint you ONLY cover the wrong parts of the foot, if all the toes have strange shapes/there are 7 or more toes, you better cover all the toes with inpaint

5-if you use the only masked option put 768x768 as resolution otherwise leave the original resolution(with inpaint area=whole picture), as prompt you can use "feet" or "toes"

6-use an optimal negative prompt, I use:

blurry, blurry_image, Lowres, Low_resolution, Low_picture_quality, Low_picture_anime, extra_anatomy, extra_body, extra_navel, extra_face, extra_eyes, extra_chest, extra_nipples, extra_hips, extra_arms, extra_hands, extra_fingers, extra_legs, extra_feet, extra_toe, missing_anatomy, missing_body, missing_navel, missing_face, missing_eyes, missing_chest, missing_nipples, missing_hips, missing_arms, missing_hands, missing_fingers, missing_legs, missing_feet, missing_toe, sketch, sketching, (worst quality: 2), (low quality:2), ((monochrome)), ((grayscale)), inpainting, HDR, bad_prompt_version2, jpeg artifacts, signature, watermark, text, error, missing fingers

But that's just what I use, maybe you are better off with other negative prompts

7-Controlnet can help you, but it is a solution I don't use often, also because it is NOT necessary for the feet

8-Use 'Denoising Strength' at 0.75 and if you see that the feet are deforming lower it gradually, between 0.6 and 0.75 you get the best results

9-Seed at -1

(to always have 4 completely different results, if you have chosen 4 as batch size)

10-Sampling Method DPM++ SDE KARRAS, it's the slowest but gives the best results, I've seen people recommend Euler A... certainly faster... but the results aren't as good...

11-Sampling Steps=28(some people put 20 here, but personally I prefer 28, sometimes 40 steps), CFG Scale=7, Mask blur=4, Only masked padding, pixels=32

Example(with inpaint you can add the missing toe/toes):

Anime: