Spaces:

adamelliotfields

/

diffusion

like 9

Running on Zero

In A1111 and Fooocus, styles are templates with {prompt} placeholders. Fooocus includes the styles from twri/sdxl_prompt_styler. You can view all (100+) of them here.

The prompt styler templates are meant for SDXL, which has 2 text encoders: ViT-G and ViT-L. The templates have a period (.) delimiter to split the template into G and L parts. The G part is the subject and the L part is for refinement and emphasis.

Using the example from the docs, ("spider man", "robot mech").blend(1, 1), we get a cyborg spiderman. Reducing the robot mech weight to 0.8, the cyborg becomes much more humanoid. At 0.5, it's just a cool looking spiderman (barely any attention on the "robot mech" part).

Create a dropdown for the styles and a slider for the weight of the style. User's prompt can be weighted at 1 since they can use Compel's inline weighting. Also do the same for the negative prompt (negative weight will be same as positive).

Don't prefix prompts with "a breathtaking photo of" like the original implementation; just include the word "breathtaking" in the style instead.

Users will be able to see the final prompt when the metadata display is implemented.