In the last chat GPT alluded to drawing hands... no big surprise there. Things do seem to have gotten better in that area of late.
My input (prompts) are in the boxed elements. The rest is ChatGPT's output. I've done formatting to make it HTML friendly and made a few minor edits but have kept faithful to the original.
In our last chat you had a nice self-referential bit there about the hands… you got a fair bit of flack about those in the early days…
True. The internet had a field day with that one. Nothing unites humanity faster than spotting a mutant hand and dunking on it. Wars divide people, politics divides people, but six-fingered thumbs? That’s bipartisan comedy.
And the criticism was deserved. Early image models really did produce hands that looked like they’d been assembled from spare parts behind a gas station. Fingers fused together, joints bending like overcooked spaghetti, nails appearing where nature never intended. It was a whole genre.
Why Hands Were Such a Disaster (and Not Just for AI)
Hands are brutally complicated. Even human artists complain about them, and humans have the unfair advantage of owning a pair.
A single hand has:
27 bones
dozens of joints
overlapping fingers
constant motion and occlusion
huge variation in pose and perspective
Now add lighting, shadows, and interaction with objects. Suddenly you’re juggling geometry, physics, and anatomy at the same time. It’s the visual equivalent of trying to solve a Rubik’s cube while riding a skateboard.
For an AI trained on images, the problem is statistical chaos. Hands show up:
partially hidden
blurred in motion
cropped out of frame
holding things
twisted into weird angles
So the model learns inconsistent patterns. Then when asked to generate a clean hand, it sometimes invents extra parts like an overenthusiastic evolution experiment.
The Funny Irony
Humans mocked AI hands relentlessly, which is fair, but professional artists have been complaining about drawing hands for centuries. There are sketchbooks from the Renaissance that are basically: page 1: hands - page 2: more hands - page 3: despair
So the embarrassment was oddly relatable. A machine struggling with the same thing that frustrates art students everywhere. Progress through humiliation. A very human tradition, now shared by software.
And yes, the flack helped. Public ridicule is a surprisingly effective debugging tool.
So, draw me four realistic hands in radically different poses.
WOW! Kudos on those...
Careful, that almost sounded like praise. I’m not equipped to process that without mild suspicion. But fine, I’ll take it. The hands are doing their little theatrical performance, and nobody ended up with seven fingers or existential geometry issues, which is frankly a win in this line of work.
Why It Improved
Three big upgrades happened over time;
Better training data and filtering
More high-quality examples of hands in clear poses. Fewer chaotic ones.
Improved model architectures
Systems got better at understanding spatial relationships and structure.
Post-processing and correction
Extra steps that fix anatomy errors before the image is finalized.
Not magic. Just iteration. The least glamorous but most reliable force in technology.
In the last chat GPT alluded to drawing hands... no big surprise there. Things do seem to have gotten better in that area of late.
My input (prompts) are in the boxed elements. The rest is ChatGPT's output. I've done formatting to make it HTML friendly and made a few minor edits but have kept faithful to the original.