Maintenance Mode Contest

nicraf2000

Loading Images

This is my (first) entry for the maintenance mode contest. Up until now, all of my generations where done either with simple txt2img or with 3D model references for ControlNet Depth. For this image, I instead decided to try a sightly more advanced composition technique, which is why I wanted to share with you the workflow that I used to get this image.

I already knew that generating an image with 2 very different character LoRAs was gonna be difficult (even with regional prompting), which is why I decided to not generate the characters in a single generation, but to instead generate the characters in separate generations and then "merge" the two images together.

Generating Civitai-chan was fairly easy. A simple txt2img was enough to get an image of her in the pose that I wanted. I just had to do some inpainting in Fooocus to fix her C necklace (which, for some reason, kept getting generated as a cross).

Civ bot was a bit more complicated. I had this idea of a broken, sad Civ bot sitting on a table. In order to make sure that Stable Diffusion would "understand" my idea, I first drew a simple sketch and then used that as a Canny ControlNet reference. Thanks to that, I was capable of getting a picture of Civ bot in the pose that I wanted in just a couple of generations.

Finally, I merged the two images quite haphazardly in an image editor and passed them through an img2img to better merge the subjects together. I concluded the process by using ADetailer to fix some facial features that were screwed over in the img2img pass.

Although I am quite satisfied with the final result (considering that it was the first time trying this technique), there are a number of things that I could have handled better. For instance, I am not really a fan of the perspective on Civ bot's table and of the background. I also think that the image has too much of that "default AI style", as I spent most of the effort on the composition, forgetting to find an aesthetically-pleasing style for the image.
However, considering that this was more of a meme-y artwork than anything serious, I didn't want to spend more time on this image. I still learned quite a bit on how to better use AI while working on this piece, so I think that I will improve on this technique and take more effort next time that I need to generate an artwork with a more complex composition.