Some stuff I've been generating in my free time on weekends. Everything's in 1920x1280px or 1280x1920px.
^-- Special mention must be given to the basketball-themed pair...which almost drove me insane with their VRAM crashes at the upscaling phase. No matter what settings I used in the startup parameters OR in the tiled VAE and hypertile tabs, Stable Diffusion kept faceplanting. I had to resort to img2img to upscale the base image, which...IMHO, isn't as ideal as a good old Hires.fix even if others keep telling me Hires.fix
is img2img. The black and white version is actually the original one, gotten through the well-known Meinamix diffusion model, but After Detailer inpainted a different face style. The more colorful one was rebuilt through the Re:imagination model, with the help of ControlNet.
And for some reason, SD has been having more trouble fashioning accurate hands than before, which is rather perplexing. Adding "disfigured" (with other keywords like "mutated", "distorted", "mutilated" or "malformed" thrown in for good measure) to the negative prompt helped...somewhat, but Jesus Christ...
And here below is what I called the "Beach Girl series/collection" - some personal experiments on prompt grammar and syntax variation (basically take one prompt and modify it little by little, see how wildly the output changes). The results were both interesting and perplexing (there were some unexpected deviations at times), and the process highlighted certain weaknesses and idiosyncrasies of the diffusion models I'm currently favoring. Generating multiple distinct characters
without the use of LoRAs OR regional prompting was a tad tricky, but I found that combining "group of (x)" and "multiple (x)" with weights did the trick, especially when working with canvases set to landscape orientation (768x512px).
Cheers.