What's new

AI "art" can be fun

Eben Stone

Staff member
This is where I plan to share some of my AI art experiments.

If you're curious about which generator I use, lately it's been Microsoft Copilot Designer. You can use it without paying money, but it requires a Microsoft account. You can submit 10 prompts per day to the fast queue, but after that you get put into the slow queue and each submission takes about 10 minutes.

I've been trying to come up with something for Father's Day.

_f730ed6e-6b94-4855-ae97-ceffe685c28b.jpeg

Generator: Microsoft Copilot
Number of tries: 3 (technically only 2 different prompts, but I refreshed the page by accident and it generated a new set of images, so I'm counting that as a submission)
Prompt:
U. S. Marine coming home, meeting his young daughter on the tarmac, viewed from over his shoulder, with a woman holding a sign saying welcome home in the background
Observations:
The woman's shadow is good but not quite right.
 

Eben Stone

Staff member
I was inspired by this thread by @Albert V :

Thread 'Did Dolly, Jaws girlfriend from 007 movie Moonraker have braces?' Did Dolly, Jaws girlfriend from 007 movie Moonraker have braces? - https://www.badgerandblade.com/forum/threads/did-dolly-jaws-girlfriend-from-007-movie-moonraker-have-braces.655041/

Challenge: Dolly's high school yearbook photo:

OIG3.ejkKYUR6HU.jpeg


Generator: Microsoft Copilot.
Number of tries: 18 (I used all my free daily credits).
Prompt:
cute nerdy 1970's blonde 20 year old woman, blue eyes, with her hair in braids, white bows at the end of the braids, wearing white low cut European country dress, wearing small thin silver rimmed small round glasses, smiling with braces on her teeth, in the likeness of dolly from moonraker movie, Polaroid photo style
Observations: IMO, looks more like my niece than a high school Dolly. I couldn't get the glasses to be small and round. Overall, the most challenging so far. Not too terrible.
 

Eben Stone

Staff member
Thanks @Eben Stone,
I have seen lots of amazing AI pictures and always thought it was expensive software.
I have a Microsoft account so I will look for Copilot.
🍻
The link title says Bing but that's the link to Copilot's image generator.
 

Eben Stone

Staff member
_905e150b-d1ca-4476-8609-84f5f34677ae.jpeg

I tried to make something Halloween themed.
Generator: Microsoft Copilot.
Number of tries: 1.
Prompt:
spooky halloween female vampire with dark skin holding a straight razor and an evil revengeful smirk
Observations: finally, the image generator seems to know what a straight razor is. Sort of. I mean at least it doesn't look an amalgam of a toothbrush and a Mach 3.
 

Eben Stone

Staff member
_a4bf0ede-6caa-4608-a851-8d382a9159ff.jpeg

_bdb2b362-59e6-4d65-a6b5-94b5e767489f.jpeg

Generator: Microsoft Copilot
Number of tries: 6
Prompt:
young female santa elf creating a safety razor in Santa's workshop, holding up the razor inspecting it carefully, with a determined focused look on her face, while santa stands in the background with a watchful but approving look on his face
Observations: Typically, Copilot gives you four images every try. These were the two best images from my sixth try. Apparently someone was naughty...
 
I use DrawThings for AI art, but it requires a fairly high-end computer to work sufficiently fast. The app allows to use different Image generation models, mostly diffusion models. I found that I get better results when I add additional image components to my prompts. The AI works then a bit more precise. I haven't experimented with negative prompts (excluding things from the generated image). Here are some prompt components you could try for your next image generation session; just add those items as you like to your own prompt:

redshift style, LED lighting, 4k, UHD, [kind of] photography, volumetric lighting, photo realistic, sharp focus, masterpiece, highly detailed, volumetric light, establishing shot of [...], trending on pixiv, beautiful [kind of] painting by [artist], trending on artstation, unreal engine 5, cinematic lighting, by [name of artist], shot with [camera brand] camera, digital art, hyperrealistic, studio lighting

Items in [ ] brackets are variable. If you'd like to generate an image of a car for example, you would use "automotive photography". If you like to generate an image in the style of specific art direction and artist you would enter the corresponding art and artist, you can also specify that the image should be generated as shot with a Hasselblad 907X for example.

Generally the more details you give in your prompt the better the results will be. My images got a whole lot better once my prompts had these add-ons to the prompts. I still can't generate a good looking razor though. I think these models weren't really trained on how a safety razor looks.

Remember, these aren't intelligent systems, despite the moniker. Diffusion models work with joint probability. They model the distribution of the next signal to add to an image (or noise to remove) given the previous signal and the number of signals already added (or noise already removed). That's why each image in a batch of six images will always look (slightly) different.
 
Top Bottom