Something more Artistic 

   
   
   

Catgirls
   
   

Scifi/Cyberpunk film scenes that don't exist

   
   
   
   
   
   
       

Whipped this up as a new DeviantArt cover image, it's not super polished but it's still nice. It works as a wallpaper too. 
   

I'm taking a break for a while, but these were in my queue of stuff to come back to and aren't too uncanny. I think I'll just post them here.

Contemplation I, II, III
   
   
   

   

Experimenting with depth mapping to give a 3d effect. 

Short video loops made with Depthy. Depthy can also do red/blue anaglyphs. There's probably some way to make stereoscopic pics for vr viewing.

This can ostensibly be done with any input image. In practice some will work better than others, especially plain backgrounds.

Statue

Curvy Nude

It's not the same as having a 3d mesh generated (which is something AI researchers are working on) but it's still fun to play with.

(November 4, 2022, 7:59 am)GravidaSomnia Finally have my own hardware and it's a much better workflow for creating artistic pieces. Render speed is about the same but viewing and transferring the output files was always slow and clunky for me when using the a1111 webui colab.

How did you get these results? They're really good. What model and settings did you use?

(November 30, 2022, 1:23 pm)someguy115 How did you get these results? They're really good. What model and settings did you use?

I don't have the settings on hand but I'll look them up when I have a chance. If I still have the original output images they can be retrieved. The process I followed for those images is in the comments of the Deviant Art Page for one of them. I generated a low resolution image using a blended Stable Diffusion/Waifu Diffusion model, then used that as the img2img input, changed back to the Stable Diffusion 1.5 model, and doubled the resolution. From the DA comments:

Quote:The first outputs come from a subtly blended SD/waifu model but the final result is vanilla Stable Diffusion 1.5. The second output reuses most of the prompt but is rewritten to better match the image as it already is. Sometimes I'll have to drop cues from the prompt as they become too emphasized.


I had been using the euler scheduler but DPM2 seems to add life to photographic reproductions. 

The implementation of Stable Diffusion I'm using will generate a grid of outputs, I can open that in an editor and clone the bits that work over the bits that don't, touch up other imperfections, and just mess around with things until I'm satisfied with the final product. Feel free to try it for yourself.

I'll also edit pics before applying img2img, this helps a lot with the awful hands that AI is notorious for producing. This technique works well with artistic styles as well, I just prefer to aim for photorealism. There's also an embedding called bad_prompt that seems to help when used as a negative prompt keyword. Get it Here and put it in your Embeddings folder, assuming you have one.

Anyway,