Amusingly, they do not have a logo.

Stable Diffusion >

• See also ComfyUI usage notes

---

TODO - img2img ∞

I do this through a ComfyUI refiner workflow.

• The steps value is critical. It remixes the image to give something mostly-similar. The larger this number, the more alike the result will be.

  • Should be extremely low: 3-7

• denoise is how much of that img2img image will be gotten rid of. This is a good number to play with.

  • Try 0.01 to 0.99
  • 0.85 seems nice.
• The initialiser phase of the refiner should have maybe 1-5 steps to just sketch things out.
• The base phase of the refiner should have a fair number of steps, perhaps 20. This will fill in the denoised amount.

• The final refiner phase cleans things up.

  • I have to do some manual math to set start_at_step and steps. start_at_step is the addition of the previous two phases, and the steps is that plus the number I want to contribute. This might be as few as 3 steps.

cfg ∞

• Stable Diffusion Deep Dive - CFG - Don't Accidentally Fry Your Images

  • by SiliconThaumaturgy
  • https://docs.google.com/spreadsheets/d/1YkX1pzJvYKrj_w6fhrpyaTyZvR687THp7QkxOVI4qEg/edit

ascore ∞

These are the weights for your Prompts - see also Prompt weights

Good starting points are:

• Positive prompt 6

• Negative prompt 2.50

Prompts ∞

The text which describes what you want for your art. This comes in two forms:

• Positive Prompt - The elements you want to appear.

• Negative Prompt - The elements you do not want to appear.

-

• ascore defines how much, in general, these are to be respected.

• Prompt weights let you take certain items (words or phrases) and give them different priority compared to other items in that prompt.

-

• See Photography AI art terms

• Stable Diffusion Basics - How to execute many prompts at once AND specify settings for each one

  • by SiliconThaumaturgy

• https://promptbase.com/stable-diffusion

  • Buy/sell

• https://huggingface.co/datasets/Gustavosta/Stable-Diffusion-Prompts

Prompt weights ∞

Prompt[not existing] weights are for items in your #prompts, and using them lets you take certain items (words or phrases) and give them different priority compared to other items in that prompt.

• ascore is a priority system for the whole of the prompt. It is a separate setting for the negative prompt and the positive prompt.

Where to get prompts ∞

First and foremost, many places which host AI generated art will give you access to the prompts used in posts. Also, some images will have prompts and settings embedded into their images, so check inside. TODO - instructions.

• https://prompthero.com/
• https://civitai.com/images
• https://www.mage.space/explore
• https://lexica.art/

• https://www.spellpt.com/

Image sizes: Height and width ∞

• Make sure the value is divisible by 16 for best results.
• Changing these numbers can dramatically change your art.

• Determine what size images a your #checkpoint was trained on, and keep your workflow to around that size or larger; there might be quality issues if you make art much smaller.

Training ∞

Training is the process where images are analyzed in order to create a #checkpoint.

TODO - how to train your own checkpoint ∞

Stable Diffusion 1.5 versus SDXL ∞

• SDXL is the new hotness, but 1.5 and 1.6 have a hell of a lot of usage, models, and peripheral software. It's worth leaning heavily into 1.5 and ignoring SDXL just to learn about Stable Diffusion for now.

  • Switching to SDXL is trivial, at least with ComfyUI. There are some additional nodes to use with it, but they're straightforward to use when necessary.

-

• SDXL 1.0 blows away Stable Diffusion 1.5. And here is the testing to prove it.

  • by SiliconThaumaturgy
  • https://docs.google.com/spreadsheets/d/1wLJ27ySec_pkSiMW1IJlAKlB_Gw3kTngpjpjAnGVwBU/edit#gid=116461310

SDXL ∞

• See Stable Diffusion 1.5 versus SDXL

• See Stable Diffusion § Where to get SDXL models

SDXL the new hotness, but 1.5 is better-supported.

Seeds ∞

When starting out

1. Specify the value (in ComfyUI it is noise_seed)
2. Make it fixed
3. Refine your process until you are confident

4. Randomize it as your last step

Checkpoints ∞

A file which contains the magic which Stable Diffusion uses to generate new art. Generally this is a checkpoint.

Some people use checkpoint and #model interchangeably.

Models ∞

• See Diffusion model

• See Stable Diffusion § Where to get models

ComfyUI uses the term "#checkpoint". To it, all sorts of files are considered models.

Most models are trained with 512x512 images, and so you should try to use that size or larger. If your images seem cropped, try 1024x1024. Changing these numbers will change your results dramatically; I don't know why.

Upscale models ∞

• See Stable Diffusion § Where to get upscale models

LORAs ∞

• See Stable Diffusion § Where to get LORAs

Checkpoint ∞

This is the primary #model from which all other art is created.

Samplers ∞

In ComfyUI the sampler node is called KSampler.

This stuff matters.

These are good with high cfg settings.

• DDIM
• DPM++ 2M
• DPM++ 2M Karras

• Heun

-

• https://stable-diffusion-art.com/samplers/

• Stable Diffusion Deep Dive - Samplers - which are worth using and which are redundant?

  • by SiliconThaumaturgy

Issues ∞

• Multiple limbs = Cohesion / fractionalization

• Higher resolution issues

  • Most #checkpoints are made for 512x512, and a side can sometimes be pushed to 768 without issue. For these checkpoints, higher resolutions such as 1024x1024 can lead to "twinning", which is the duplication of characters. Some custom checkpoints are trained for higher resolutions.
  • "Stretching" can occur

Resources / tutorials ∞

• ComfyUI usage notes
• AUTOMATIC1111 might have stuff specific to it.

• was sdcompendium.com/doku.php?id=frequently_asked_questions_0005

Last updated 2025-02-04 at 01:28:26