A merged model created by AIBooru's own @redjoe23. The model was designed to combine dreamlike_diffusion's hyper responsiveness to artist prompts with the NSFW of anything, or some alternative model (see below).
- Clip Skip: 1
- Steps: 30 - 50
- CFG Scale: 8 - 14
HiRex.Fix to 2x with denoising at 0.5-0.7. Set HiRes steps to half of normal steps to save time.
AnyDream works best when artists are added to the prompts. Begin by writing the first prompt as a short statement describing the subject as well as some combination of medium, setting, and action. For example:
a painting of a woman walking through a forest, or
a portrait of a cyborg ninja. After that, add any number of prompts to modify your image, the same as you would for any other model.
Finally, add your artist prompts at the end. Typically, some combination two or three artists works best, such as the popular "Artgerm mix". Put the word by before each name, so the Artgerm mix would become
by artgerm, by greg rutkowski, by alphonse mucha.
- Experiment, experiment, experiment. Keep notes. When you find an artist mix you like, save it as a style in Webui so you can add it easily.
- Contrasting artists work much better than you would think. The Artgerm mix, for example, is a combination of comic art, fantasy paintings, and art nouveau.
- For artist lists with AI-generated samples, look here or here.
- Add the tag
dreamlikeartto enhance the dreamlike effect on the image.
- Adding more prompts tends to water down the effect of your artist prompts. Enhance them 20% to 50% in WebUI by highlighting them and hitting hitting ctrl + up a few times, so
by Ralph Horsleybecomes
(by Ralph Horsley:1.2).
- Removing the word by from the artist prompt can actually change the results a bit. Experiment.
- If you aren't getting results you expect from your artist mix, you may need to cut a few. Build a simple prompt with your artists and use a prompt matrix to get a simple matrix for comparing all the possible combinations.