ORIGINALS
My different figure drawing styles/approaches
over the years
AI ALTERED
.gif)
The above originals altered with two passes of KREA AI,
with one pass of Stable Diffusion AI in between
OBSERVATION: No matter what drawing style I started out with, the final AI altered images all looked fairly similar
DIFFERENT STYLES
altered with AI
I altered my drawings with two AI ART programs -- KREA and Stable Diffusion -- with AI models trained on my drawings (Open Art model used in Stable Diffusion):
.png)
Mixture of the KREA and Stable Diffusion
versions, remixed in KREA
Is the blend from both AI programs
the best result?
.png)
.png)
.png)
.png)
.png)
.png)
.png)
.png)
.png)
The above AI generations are more of a test -- to see how the AI models trained on my drawings, effect my drawings in various styles. The AI results could have been a lot prettier had I used other AI models in Stable Diffusion.
The last AI alteration seems to be so powerful in fact, that it can radically improve two bad images -- specifically those mixed together in KREA:
The two top bad Stable Diffusion versions
mixed and improved in KREA
(no prompts added)
NOTE: The KREA website changed their interface on February 26, 2026. The results seem to be better now, stilling using the FLUX model, but that might be an illusion.
AI ART Models
Again, I believe that the power and future of AI ART is from TRAINING ONE'S OWN MODELS...specifically on one's own drawings and artwork.
- Book_model_final.pkl -- (364 MB) (trained on short pose drawings)
- Long_Krrrl_Drawings__final.pkl -- (364 MB) (trained on long pose drawings)
NOTE: Stable Diffusion accepts the "ckpt" model -- not the more robust "pkl" models. However there might be a way to convert the "pkl" models to "ckpt" models -- and then use them in Stable Diffusion:
- model_20221213233657.ckpt (2 GIGS)
2023 Leonardo.AI: I trained an AI model inside Leonardo.AI on 30 of my drawings. These models only work inside Leonardo.AI. I liked how that worked until Canva bought Leonardo.AI in July 2024.
2025 KREA: I trained various AI models from my drawings in KREA (mostly for FLUX). Again, these models only work inside KREA, and would train on no more than 50 of my drawings.
2025 QWEN: I found an AI program that will run locally (on a computer not connected to the Internet). DiffSynth-Studio/Qwen-Image-i2L is a program that will train LoRA AI models from one or more images. The resulting LoRAs will work in Stable Diffusion...however the model I generated in Hugging Face was 281.43 MB, really too big to post online.
*****
- On February 25, 2026, Mitchell and I tried to upload my AI model to the newish powerful AI computer at Quelab
- This we after the shooting in front of the makerspace that night
- I generated the Stable Diffusion images above with my AI model on a different computer at Quelab, with less VRAM
No comments:
Post a Comment
Note: only a member of this blog may post a comment.