Improving image generations on Auto1111 Stable Difusion (Patreon)

Do you have a problem achieving the perfect image generation, even if your prompt is right? In this tutorial I will quickly show you how you can find the desired result from images that are already close to the final intent.

1. Do your prompt as normal

Do your prompt as normal, when you find a result almost like the one you want, but still imperfect, go to the next step. If the result is perfect go to step 3.

2. Extra seed

Use the extra option available from auto1111 or similar. This tool allows you to transition over small variations within a seed that is already close to the generation you want. Don't forget to recycle the seed of your main generation!

Use the extra seed option. I recommend using variation strengths between 0.01 and 0.1, where higher values means more variation. Find the generation that suits you best.

By increasing the batch size and batch count, the generated images will already be around the main seed when you click on "generate".

Don't forget that for the next step you need to recycle the extra seed as well.

3. Use High Resolution fix (or Hires)

Use the hires fix to increase the resolution and enlarge the details, I recommend leaving the denoise between 0.2 and 0.6 (higher values equals more differences from the original image).

When you set denoise to .2, you need drop the Hires steps. When I use 35 steps at normal generation, I drop to 20 at hires fix with .15-.25 denoise.

List of upscalers I recommend:

• Latent: Make more differences from original image. General Use.
• Latent (bicubic antialiased): Works well when you want to increase details of a pattern or a intricated thing, but you need increase denoise a bit as well.
• 4x-UltraSharp, 4x_NMKD-Siax, 4x_NMKD-Superscale or 8x_NMKD-Superscale: I use this when I just want to increase the size of the image without losing fidelity. I transit between them to find the result I desire.

To install put the model(on attachments) on \models\ESRGAN folder of your auto1111. All the credits of the models to the creators Kim2091 and Nmkd.

4. Inpainting

To make it even better, you can still use inpaint on details. Use between 0.3 - 0.5 denoise and DDIM sampler, with the same CFG you used for the original prompt. Also describe better the detail that you want to change, and put 1.2~1.3 weight on it.

To speed up the process use "only masked" in inpaint area option and set a square resolution, such as 512x512 or 768x768.