Oct 14, 2022 21:00:00

``MagicPrompt'' Stable Diffusion version that allows you to easily and automatically create many prompts and spells in one shot just by telling the content you want to draw in the image generation AI ``Stable Diffusion''

In the image generation AI 'Stable Diffusion', if you want to generate an image, you need a character string called 'prompt' in advance, but this prompt does not mean that you can simply write it in sentences, but you can enter it in your own format and order. You have to use a service like

KREA or Lexica to examine and combine or rewrite existing prompts. A web application that automatically creates a prompt for generating an image when you enter the image you want to generate in sentences. is possible.

MagicPrompt Stable Diffusion - a Hugging Face Space by Gustavosta
https://huggingface.co/spaces/Gustavosta/MagicPrompt-Stable-Diffusion

In the 'Initial Text' field on the left, enter the content you want to generate in English. This time, enter 'Illustration of a beautiful girl in white dress sitting by the window' and click 'Send' to create a prompt to generate 'Illustration of a beautiful girl in a white dress sitting by the window'.

Then, 4 patterns of prompts were displayed in 'Generated Prompts' on the left.

The prompt created is as follows.
1: Illustration of a beautiful girl in white dress sitting by the window, highly detailed face, beautifull face, beautiful detailed eyes, fantasy art, in the style of JAMES MCCARTHY, illustration, epic, fantasy, intricate, hyper detailed, artstation, concept art, smooth, sharp focus, ray tracing, vibrant, photorealistic
2: Illustration of a beautiful girl in white dress sitting by the window of a spaceship, dark atmosphere, beautiful detailed face, by Stanley Artgerm Lau, WLOP, Rossdraws, James Jean, Andrei Riabovitchev, Marc Simonetti, and Sakimichan, trending on artstation
3: Illustration of a beautiful girl in white dress sitting by the window wearing a black dress, beautifull face, by Alphonse Mucha, by artgerm, body tilted, 4k resolution, highly detailed, beautiful
4: Illustration of a beautiful girl in white dress sitting by the window of her father as a baby. She has light blue eyes, black hair and bright blue hair in a bob cut. By Makoto Shinkai, Stanley Artgerm Lau, WLOP, Rossdraws, James Jean, Andrei Riabovitchev, Marc Simonetti, krenz cushart, Sakimichan, trending on ArtStation, digital

I entered the four created prompts into the

AUTOMATIC 1111 version Stable Diffusion web UI , made the generation settings, and clicked 'Generate' to actually generate the image. The AI models used are Stable Diffusion v1.4 (model.ckpt), Waifu-Diffusion v1.2 (wd-v1-2-full-ema.ckpt), Waifu-Diffusion v1.3 (wd-v1-3-full .ckpt), the number of steps is 100, the sampler is Euler a, the resolution is 512 x 512 pixels, the CFG scale is 7, and the seed values are 1234, 56789, and 24681357.

The result of entering the first prompt is as follows. Although the white dress has been converted to 'white dress', the window is drawn in the background because it reflects the element 'sitting by the window'. However, in Waifu-Diffusion v1.3, for some reason, a mysterious cosmic horror-like existence is drawn, and the finish is somewhat horror-like.

In the second prompt, for some reason, 'sitting by the window of a spaceship' and spaceship elements were added without permission, so everything is an illustration like a science fiction animation. increase. The images generated by Waifu-Diffusion, which is learning from the dataset of the 2D illustration site Danbooru, give the impression that they are more complete as 2D illustrations. is drawn in fine detail.

When generated with the third prompt, the prompt 'by Alphonse

Mucha ' has a strong influence.However, the picture by Stable Diffusion has a strong Mucha touch, but Waifu-Diffusion The image by is quite a manga-like pattern.

The image generated by the fourth prompt is below. `` She has light blue eyes, black hair and bright blue hair in a bob cut '', which clearly reflects the girl's appearance. Also, for some reason, it was interpreted as a composition holding a baby, and although I got the impression that the baby was somewhat collapsed, I was able to draw the `` beautiful girl in a white dress sitting by the window '' without any problems.

In addition, this 'MagicPrompt' Stable Diffusion version of the model was trained in 150,000 steps, filtered with about 80,000 datasets, and extracted from '

Lexica.art ', which shows several examples of stable diffusion. made using data from Mr. Gustavosta , the author, has already completed the ' Dall-E 2 ' version, plans to create a Midourney version in the future, and is also developing a 'complete version' that integrates these.