GymDreams

Before deciding on a name for a brand / product / identity, I always do a few sketches and make sure that I can design a decent logo for it before committing on the name.

For this identity, after a few sketches, I saw that I could make a nice mirrored form for the G and D, and that’s why I went ahead and registered the username GymDreams on DeviantArt — the original place where I started posting my AI works. There’s a longer backstory of where this name came from, but I’ll write about it in the future.

Once that account got some traction, I realized that I should probably start making profiles on other social media, but it didn’t take long for me to realize that this name faces some challenges since it’s often taken. I don’t want to randomly add spaces and dashes to the name — seeing that the logo I designed can be seen as a rotated figure 8, I decided to register @GymDreams8 on Instagram, Twitter, YouTube, and also bought the domain GymDreams8.com

Regarding the design, my art education was deeply rooted in the Bauhaus philosophy, so I also prefer forms that are made out of the basic shapes — squares, circles, and triangles. As such, logos that I design tend to have structures that can fit nicely in a grid, with proportions that can be described in x, 2x, 3x, 5x, etc.

The form of each letter is based on a circle with 5x radius or 10x diameter. They are separated by a gap that’s x long. Together, the two letterforms create a mark that’s exactly 21:10.

Have you ever wondered why so many images generated with AI feature a single person? Are you disappointed to find that AI generated couples often look like twins? There’s a reason for that.

Text-to-image engines can’t natively know which part of the text prompt belongs to each person. So when you asked for different outfits for each of your characters, they often mix everything together. Midjourney is one such engine. A follower asked me why I don’t post images of couples often. That’s because I have to generate 100+ images and hope that the RNG gods will bless me with 5 usable images.

Regional Prompter was created to deal with this issue. You can divide the canvas into regions, then individually prompt for things which don’t collide. The two sets before this show the results.

2-3

I divided the canvas into two regions horizontally (1:1) and prompt individually. To show that the AI was not simply associating proper attire with the expected age group, I swapped the clothes.

4

You can define base and common regions that make interesting results. For example, here I divided into two regions, but asked for 3 men. It creates an image with 3 men with a progression of age, hair color, and clothing.

5

If you asked for two men, the split goes all the way through! Try doing this with just txt2img and you know that this is impossible, or at least very difficult.

6-8

You can get creative and split a single person to look different for different regions.

9-10

Consistency and productivity. Prayers to RNGods not required.

Text prompts in Stable Diffusion with Regional Prompter. Source image as is. No upscaling, no post-processing, no in-painting. The results are really that good!

• Model: A-Zovya Photoreal v1 Ultra

• Sampler: DPM++ 2M Karras

• Res: 512x512

• Steps: 30

• Control Net: None

• Regional Prompter: Horizontal 1,1, varying base ratio (0.2-0.5) depending on image

Note: Actual implementation uses 0-based region IDs.

Image Sets

I haven’t posted these images to DeviantArt because it feels spammy, as these are process images to demonstrate the workflow, but if you think that they’re helpful, I will consider posting some here.

• Blog post: https://www.instagram.com/p/CsqcAwyL1kH/

• Age difference (45 vs 25), set of 10: https://www.instagram.com/p/CsqD1i2L-Te/

• Size difference (from Twink to Twunk), set of 10: https://www.instagram.com/p/CsqMnecLkqy/

Resources

• Automatic1111 extension (works with Vlad also): https://github.com/hako-mikan/sd-webui-regional-prompter

• Tutorial on Stable Diffusion Art: https://stable-diffusion-art.com/regional-prompter/