Original
Generated with model "Krrrl Drawings 01"
Generated with model "Krrrl Long Drawings 07"
I created another model in Leonardo.AI -- Krrrl Long Drawings 07 -- trained on 30 of my best long poses (I made a previous model of my long poses trained on 348 drawings). The images I uploaded however were small, 640 pixels in one direction. I altered all the drawings in this blog post with that model; and altered most with the model "Krrrl Drawings 01" (based on my short pose drawings) as well.
Original
Generated with model "Krrrl Long Drawings 07"
Original
Generated with model "Krrrl Drawings 01"
Generated with model "Krrrl Long Drawings 07"
Original
Generated with model "Krrrl Long Drawings 07"
Original
Generated with model "Krrrl Long Drawings 07"
Original
Generated with model "Krrrl Drawings 01"
Generated with model "Krrrl Long Drawings 07"
Original
Generated with model "Krrrl Drawings 01"
Generated with model "Krrrl Long Drawings 07"
Original
Generated with model "Krrrl Drawings 01"
Generated with model "Krrrl Long Drawings 07"
Original
Generated with model "Krrrl Drawings 01"
Generated with model "Krrrl Long Drawings 07"
Original
Generated with model "Krrrl Drawings 01"
Generated with model "Krrrl Long Drawings 07"
Original
Generated with model "Krrrl Drawings 01"
Generated with model "Krrrl Long Drawings 07"
Original
Generated with model "Krrrl Drawings 01"
I made the most barebones AI alterations to the drawings in this post in Leonardo.AI, thinking that the most basic treatment would reveal something about how the AI works.
I was pleased how the new model - Krrrl Long Drawings 07 -- transformed my drawings. It was trained on 30 of my long pose drawings, which have more detail. However, even though I made a larger 768x768 model, I trained on low resolution images (640 pixels maximum dimension). Ultimately Leonardo.AI generated more detail, but in low resolution.
I trained the first model -- Krrrl Drawings 01 -- on high resolution images from my book which I downloaded from my GDrive, and was quite impressed with the images that Leonardo.AI generated from this model. Even though it only generates 512x512 images, I think that training on the higher resolution images made a difference.
I still have all the physical original long pose drawings. Perhaps the next step is to photograph these drawings in high resolution and train a Leonardo.AI model on higher resolution photographs. I have reached my limit, and can only train one more model before Leonardo.AI refreshes me in a month or so.
Other LONG POSE Model
Generations
I pushed the new model -- Krrrl Long Drawings 07 -- with all the richer recipes from Leonardo.AI, and generated the figures below:
.jpg)
.jpg)
I get a huge boost when I include "Bernini sculpture" in my prompt:
.jpg)
_1.jpg)
The new long pose model does give a different touch from the old first model, as one can see in the two comparisons below.
The first sculpture in both images has a more lumpy random clay-like treatment, as it was generated with "Krrrl Long Drawings 07." The second sculpture has a more linear treatment, and was generated with "Krrrl Drawings 01."
Long pose model generation
vs short pose model
Long pose model generation
vs short pose model
Scribble Diffusion
GAME OF TELPHONE
From bad contour sketch to well rendered figure --
Using Scribble Diffusion and Midjourney
Scribble Diffusion (YouTube) pushes the AI game even further. I used my mouse to make a crude drawing of the last sculpture above, and added the prompt: "a drawing of a standing woman," and the AI generated a better charcoal drawing than I could have made from my own hand:
Double Dipping with AI: I then uploaded the above drawing to Midjourney and enhanced it:
A second try with Midjourney improves the image substantially, using this YouTuber's prompt:
- Midjourney - Turn A Sketch Or Drawing Into Photorealistic (Hyper-Realistic) AI Art In 2 MINUTES!
- natural lighting::1 natural features::1 35mm lens::1 global illumination::1 uplight::1 dof --ar 16:9
Conclusion
I thought that then I could use AI during the model break, to guide me during the second 20 minutes of the pose. Scott Eaton was pioneering this by using AI and his photographs to inform his drawing. Perhaps I could use an AI model of my own drawings (like "Krrrl Long Drawings 07") to AI adjust my drawings during the model break.
Scribble Diffusion implies that will happen in the future. Probably the next generation of life drawing students will be coached with AI Assistance in this manner (even drawing from the computer screen with Posemaniacs?).
_by_Thomas_Rowlandson.jpg)
_by_Thomas_Rowlandson.jpg)
Life Drawing
WITHOUT AI
_by_Thomas_Rowlandson.jpg){kind=link}
NOTE: In 2021 I enjoyed drawing the model in 3D with a Wacom stylus using Paint 3D on Windows 10. I hear that now there is a 3D drawing program for the iPad (that is like Zbrush?) -- Nomad Sculpt -- (YouTube tutorial from Dave Reed).
No comments:
Post a Comment
Note: only a member of this blog may post a comment.