Wednesday, 30 November 2022

November 2022 Summary -- Overwhelmed by AI Art

Alexa, draw the model for me:

Midjourney prompt: "Life drawing by Alexa"

On November 5th AI Art made a quantum leap when Midjourney upgraded to "Version 4." I quickly became addicted and could barely keep up -- letting AI transform my figure drawings, for the good, bad and the unexpected:

My drawing photobashed with AI

AI preserved the balance and gesture
of my figure drawing


UPDATE  (Dec 10): Jazza just made a YouTube video about finessing AI to enhance your drawing:  

MY BEST AI CREATIONS
of November 2022































Download the AI Origami animals:
  • Stroked PNG
  • SVG vector files


I reduced the left image
 to a binary file (Floyd-Steinberg) on the right,
and created this diptych to be read as a hologram





Reverse Prompt --
 Img2Prompt generated text after I uploaded one of my drawings,
then I used that text prompt in Midjourney


PROGRESSIVE BANNERS



OBSERVATIONS

1)  Midjourney dramatically improved after they uploaded "Version 4" on November 5th.  Check out the before and after below, generated with similar prompts which asked that Lincoln hang out with Metallica:
The first image created in Midjourney on November 1st,
the second image created in Midjourney on November 5th

The other AI programs likewise are also taking great leaps.  The new Stable Diffusion 2.0 came out about November 24th, but seems to have taken a step backwards.

2)  Midjourney generated the best results when I uploaded two images to mix them, or used a famous artist's name in the prompt ("in the style of Giorgio de Chirico"). It seems like most of the community was focused on finessing word prompts to refine the images (like Prompt builder) which I did not use, except for the reverse prompt program img2prompt (YouTube explanation).

3) There is a tendency for Midjourney to default into a Barbie.  

Midjourney prompt: "3D model of Barbie created in Pepakura"


Vague prompts usually generate female figures (often only portraits) with similar proportions, in the face and body.  Even when I upload an image to Midjourney, it leans towards a Barbie figure.  I feel this probably is happening to all the other users as well, almost pushing us to all create the same idealized human archetype.

4)  I felt that Midjourney was also generating the same color chords, so I was really pleased to discover an AI Color program -- Palette FM



GOAL:  Ultimately I want AI to repair my figure drawings, or at least push them into new creative and unexpected styles. However in order to do that the AI has to be somewhat faithful to the drawings that I upload, and Midjourney is getting better at that.

LINKS

DOWNLOADS from my GDrive: 

YouTube Playlist:



AI Resources:





COLORS:





AI Thoughts:



(Midjourney -- Version 4):

OTHER:









Fake Henry Moores



How much influence do I have over AI when I upload my drawings? I uploaded several of the drawings I did in this month of November to Midjourney (version 4) with the prompt: "in the style of Henry Moore." Then I compared the generated results to see how faithful they were to my original drawing:


BEST FITS

These Henry Moore AI generations seem more faithful to the seed drawing I uploaded -- at least the AI preserved the flow of the figure. The pen + pencil drawings seem to translate pretty well:






***






***






***

















***






***







***








LESS FAITHFUL

These Henry Moore translations generated by Midjourney (version 4) are a little less faithful to the initial uploaded drawing, but interesting nevertheless.  Perhaps the starter drawings are a little less clear:







***

November 15, 2022





***







***





***







***







OTHER APPROCHES

I tried using the "Text 2 Dream" feature in Deep Dream Generator to transform my drawing into a Henry Moore, but it only generated a gesture from my thin lines:

Deep Dream Generator -- Text 2 Dream
prompt: "in the style of Henry Moore"


Sparse, and not impressive --
but Text 2 Dream did capture the gesture


I uploaded the best Midjourney result to DALL-E 2 and continued with "outpainting":





Then to complicate things by recursion, I uploaded the above DALL-E 2 outpainting images to Midjourney with the same prompt, and generate the four images below:

Midjourney prompt: "in the style of Henry Moore"