B.A. Thanksgiving - How it was made. (Patreon)

Hello all you marvelous supporters!

I wanted to give a step-by-step process on how I created this image. The tools I used are NovelAI's Diffusion Anime v3 model and Photoshop. There's a lot to go through, so here we go...

1) First, I created a generic base image of four girls seated at a table with a Thanksgiving meal. The actual details of the image aren't as important, I'm just looking for general image composition. The original image was of four girls in maid outfits around the table, but unfortunately NovelAI had a server hiccup so I wasn't able to save that image.

2) My next step was to Inpaint the Bayshore Academy girls on top of the existing characters, one at a time. As I add to the prompt, I use de-emphasis brackets [[[[ ]]]] around descriptions of characters already in the image, and use emphasis braces {{{{ }}}} around the character I'm trying to add. This takes a bit of time and luck to get the AI to give me the proper character designs, especially as I add more characters. Image 2a is about midway through my inpainting, before I decided to put all the girls in dresses. Image 2b is the final result of inpainting all the characters, with additional inpainting to fix small details like eyes and hands.

3) With all the characters added to the image, I used NAI's Enhance and Upscale function to help clean up the image after inpainting. I included all descriptions of the characters in the prompt, with each grouped in its own set of emphasis braces {{ }}. This takes some time and multiple rolls to get a good clean image to work with (image 3), along with some extras for compositing (3a and 3b)

4) Using Photoshop, I composite parts of images 3a and 3b into image 3 to clean up bits of the image like hands and shadows. This gives me an intermediate image (image 4) that I could stop at. However, I want to improve the details on the image, so I kept going using this image.

5) Using the intermediate image (image 4), I used Photoshop to mask out sections of the image into other images with the same dimensions that NAI uses for it's standard images. I also had to do some painting to remove parts of the image (like clearing the table). I did this for each of the girls and the table (which I had to break into two images because the table was too large).

6) Next thing I did was feed the image sections back into NAI, and used its image2image function to remake and restore details. By using a low img2img strength and a very small amount of noise, the image sections keep their composition and structure, which allows them to be composited back together in the original intermediate image. While doing this, I edited the prompt to only include the aspects of the image sections, and used this opportunity to help emphasis and fix little details.

7) With the details restored to sections of the images, it's back into Photoshop to composite all the pieces back together into the intermediate image (image 4). This was a lot of time and finesse to mask all the pieces back together and blend them as best I can together. I had to do a bit of painting in the image as well to help fix and composite the new table on top of the girls. After a lot of effort and patience, I ended up with image 7. But, as much as I could stop here, I'm not done yet.

8) For the final touches and cleanup, I feed the detailed image (image 7) back into NAI again, and restore the prompt to include everything in the image. I use a very very low strength img2img with very low noise to help clean up seams and noticeable painting spots in the image. This didn't take too many pulls, since the image is already so close. Once I got a satisfactory image generated, I ran it through NAI's Enhance for that little bit of extra cleanup. The image looks great, but I'm still not done yet!

9) I take the now detailed and enhanced image back into Photoshop for some final touchups, mainly to fix the eyes. This isn't a lot of work, but makes a big difference in helping all the characters have the same eye style and shine. This is the actual final step, and the resulting image looks fantastic.

10) Now that I had my final image, I used Photoshop one last time to add the Bayshore Academy logo and resize the image into wallpaper sizes.

And that's all folks! That's the whole process, start to finish. I had a lot of fun making this image, but it was a lot of work. I wasn't sure if I would be able to add details back into the image once I had all four characters in place, but I was again pleasantly surprised by NAI's new v3 model. If you're not using NovelAI, this process should still work for any Stable Diffusion setup, as the only function you need is Image2Image. The compositing in Photoshop is also very simple, and only utilizes basic layers, masks, and some brush painting. I hope this is informative for anyone interested in attempting to make images with multiple specific characters. I'll also be making a post with this guide in the research channels of the Bayshore Academy Discord.

If you like this kind of breakdown and would like to see more content like this in the future, or if you have questions about the process, let me know here or in the game's Discord.

And, as always, thank you for your continued support! - K. Bro