FamilyDev
patreon
Arts generating workflow
🕑 Added 2024-01-03 12:45:43 +0000 UTCTo explain how the arts is created, I use multiple methods based on the level of complexity of each scene.
For instance, the first complex scenario involves two chars embracing, which makes it challenging for a neural network to accurately depict. Thus, the initial approach is to take a pre-existing template from the internet that contains most of the necessary elements. Afterwards, I cover up the superfluous details on the template to ensure that neuro will do what I want from her. First, the background changes, I change the tiles to the grass. Then, I refine texture of skin. In addition, I remove the tattoo on the hand as Amy don't have tattoos. I not redraw hands, because it will only get worse! Lastly, I modify the faces and hairstyles to match the desired appearance of the characters.
The second example is easier. Initially, the template was of small resolution (626 x 417). Stable Diffusion was trained at 512 x 512, so you can't use smaller pics as template. At first the resolution was increased with minimal changes. Everything has been changed here, silhouette to Amy's silhouette, the telescope and the stars.
In the third scenario set in a strip club, generating an appropriate template is challenging, what I need is simply not available on the Internet or I just did not find it) Hence, I resort to using masks obtained from various pics and editing them accordingly. I then combine these masks together to form a single mask that can be utilized for art generation. Once the basic setting is established, I experiment with different prompts and settings until I achieve the desired result. The resulting art is being finalized, faces are drawn separately in higher resolution, since the stripper had a short skirt, I also removed it in inpaint. A separate pain is the stripper's face, because she is bent over. Therefore, in order to finish drawing her face in a higher resolution, the entire image was rotated by 90 degrees for that)) To add authenticity to the environment, I include props such as poles and neon lights.
Lastly, when creating an image of a young version of a Veronika, I rely solely on neural network generation without the use of external templates. I begin by selecting suitable prompts and configuring the generator accordingly. For some reason, the "18 years old" prompt was still drawing an adult Veronika about 30 years old, so to make her look 18, the prompt was "16 years old". Yes, it's weird, but it works. Through trial and error, I manipulate the parameters until I obtain a satisfactory outcome. Subsequently, I redraw the facial features separately and incorporate them into the final art.
Overall, my workflow involves a combination of manual interventions and automation using neural networks depending on the complexity of the scenario. Thank you once again for your interest in my game.
P.S. If you are interested, let me know, I will post more often about my methods.
Comments
Shmurg!
All AI’s have a different golden age that produces the most attractive faces. Don’t feel bad about typing in 16 years old because the sampling for that AI might have more mature looking models that it gathers its samples from. Each AI is different
Shmurg!
Oh wow… this is awesome! I love seeming the artistic workflow. It’s also pretty cool to see the visual differences between your models and the AI generated models.
