Gemini Carbon naturally can paint a landscape well but struggles with faces. What is astounding about agents is how quickly they learn.
The first image is not good; the second image resulted from the following comment only:
"he looks like a potato"
"Fair call—hand-coding horizontal span intervals by eye gave him a rounded russet-potato silhouette and bumpy row back-runs with no real skull or hairline structure.
Rebuilt it properly from Antony Starr's actual facial anatomy using the skill's measured layering + shape-stroke pipeline (references/copying-a-reference.md) combined with hand-laid wet-in-wet washes"
https://t.co/4NGtFgumok
Gemini 4 Carbon is very good at painting.
I created a watercolouring skill which has no undo button, so the agent has to do a stroke, wet blend and then stand back and wait until it dries before assessing whether it's messed up or not.
@CtrlAltDwayne Have you tried getting angry with it?
Giving a still of the intended level of quality/ detail also allows it to do gradient descent and recursively self improve
@kevin_t_ngo Wonderful, love the pallet and style.
LLMs though, are always eager to spam a detail or object; i.e. almost too many kites or too many houses.
My theory is make the process onerous and they will natural find a stopping point. Here is my WIP painting skill's version of yours.
it seems that those not by charmed this ratchet rendition perhaps miss the point. Clankers can learn, and pretty quickly.
After giving it a few tutorials it's already materially better. Imagine if someone with actual painting talent was teaching it (i.e. not me).
At that point what does AI detection even look for? Your brush strokes are too artistic? Your composition is too on point?
This rig is intentionally annoying for agents, there is no undo button. It "paints" by writing a set of instructions for brush strokes with: direction, speed, twist, pressure with a given brush and pigment/medium ratio.
Then it has to dry, depending on how wet the paper is.
So far it seems OK on landscapes, but portraits (as you'd expect) are much harder and need more training.
As you can see in the simulation below the agent has to take a step back and watch the paint dry to see if it messed it up or not 🤣
Fable 4.2 really is great at painting and writing code. Currently working on a new rig which simulates realistic watercolours through fluid dynamics, gravity (angle of canvas), brush texture and bristle length (the list goes on)