Stuffy Stable Diffusion Guide (backup)

Setup

Install webui
I use this webui on an NVIDIA GPU.

  1. Update NVIDIA drivers.
  2. Download sd.webui.zip from here and extract in a folder.
  3. Double click update.bat, once it finishes, close it and then double click run.bat (the first time it's going to take a while).
  4. Go to the "webui" folder, find a file called webui-user.bat, right click to edit it and add arguments like below
    set COMMANDLINE_ARGS=--autolaunch --update-check --no-half-vae --opt-sdp-attention
  5. Any time you want to open the program, double click run.bat. When you want to shut it down, close the command terminal.
  6. In the settings tab, go to User interface and in Quicksettings list, add "CLIP_stop_at_last_layers", apply and reload.

Models

  • Download my model or any other you like and place it in your stable diffusion models folder.
  • Download this vae file, place it in your VAE models folder, then go in the Settings tab > VAE, select and apply settings.
  • I also use two negative embedding, EasyNegative and Negative_Hand. Download them and place them in your embeddings folder.
  • If you are going to use controlnet, you need to install the extension, then download the models you plan to use and place them in your controlnet models folder.
  • To make characters, I make and use my own Loras, here's my free Patreon guide if you're interested in that.

Workflow

The workflow below is what I'm currently using, I often tweak and modify things but I try to keep this guide up to date.

Parameters
Tip: You can set your preferred values once, then go in settings tab > Defaults, view and save the new defaults so you don't have to change them every time.

  • Base prompt: (masterpiece, best quality:1.2),
  • Negative prompt: (worst quality, low quality:1.2), (easynegative:0.8), negative_hand-neg, (low contrast), (monochrome)
  • Sampling method: DPM++ 2M Karras
  • Sampling steps: 16
  • CFG: 5.5
  • Clip skip: 2

Txt2img

  • My goal at this step is to get a decent base image I know I can work with.
  • Most resolutions are fine but play around with it.
  • I usually prompt for whatever I need, but I add or change many things later.
  • Sometimes I use controlnet depth, canny or softedge if I have a specific composition in mind or if I can't directly get what I want, for the input image I use other generations I've made.

Img2img

  • Once I have an image that I believe is good enough, I run it through img2img while manually making changes and fixing issues in Photoshop in between each pass.
  • I find it more convenient to make big changes in the first pass, so this is where I try to redraw anything major and make a lot of attempts, the subsequent passes are more for refining.
  • I roughly go up to around 800x1200 or equivalent in pixels for the resolution.
  • I enable controlnet here too if I think it might help. Controlnet Tile can be very useful when increasing the resolution here.
  • For the resize mode I mostly use Just Resize (latent upscale). I use around 0.6 for the denoising strength.
  • But it's best to play around with all these settings to better understand what to do.

Photoshop
Here's some of the tools I use and some examples.

  • Brush, eraser, eyedropper tool: I use them all the time to roughly redraw parts manually.
  • The new remove tool: it works really well to remove undesired parts or artifacts from an image, it can also fix lines and more...
  • Selection tools: I often use them to mix parts of different generations together, also to select parts so I can modify them individually.
  • Liquify filter: it's very useful to change the shape of the body, but it can be used for other things too.
  • Adjustment layers: I use them to change colors, brightness, etc...
  • After upscaling the image, I do some simple post processing and remove any remaining artifact.
  • Free Patreon post with some more tips.

Inpaint

  • I use inpainting to make medium and small adjustments, and most importantly for the face. I go back and forth between inpainting and Photoshop until I've made all of the changes I wanted.
  • I use "Whole picture" for most things, and "Only masked" for the face.
  • For the denoising strength I use anywhere from 0.3 to 0.6. But for the face I stay at around 0.4. But play around with all the settings.

Upscale
Finally I upscale the image with these settings:

  • Resize: 2
  • Upscaler: 4x-AnimeSharp (download and place the .pth file in your ESRGAN models folder)

https://x.com/StuffyAi
https://www.patreon.com/StuffyAi (NSFW)

Edit
Pub: 23 Mar 2024 13:39 UTC
Views: 389