Something I've been trying to pin down for a while is a depth of field effect.
The problem with Three.js's built in effect was that it was realistic to how real dof works and so was difficult to art direct.
Things in the foreground (like block manipulatives) would blur when we didn't want them to.
Finally landed on a custom shader that blurs anything behind a certain distance from the camera. We can control the blur amount, distance away from the camera the scene starts blurring, and the feather amount (how much blurrier things get the further away from the camera you move).
The result is a totally art directable effect that's easy to apply to all our scenes.
Had Fable 5 make me a browser port of Stronghold Crusader.
Impressively, it nailed the game logic in one shot (ai enemies, character health, pathfinding, farming, troop selection, map generation). This is the first model that one shotted anything of any complexity for me.
The weak spot is still graphics. I had to make a couple follow up prompts to improve the graphics and after switching to a sprite sheet approach it started to look pretty good. Maybe pre-rendered graphics will make a comeback?
@thdxr Consider asking it to break the work into commits first.
Iโve found that it makes it easier for me to vet the plan and makes the llm more reliable during implementation.
@HabsGifs The only thing I can think is that heโs thinking longer term and wanted to give struble some playoff experience.
But man not having Xhekaj in tonight is baffling.
Like the series against Washington last year the adjustments are coming too late.
@HabsGifs I actually feel like the man-to-man was just him implementing the simplest system he could with a young team.
Iโd be surprised if heโs a zealot about that system long term.
Everyone serious knows AGI isn't here, but I like doing experiments just to make sure.
Ran a 12 hour ralph loop that on each iteration:
Compares the existing scene with a target image (top right)
Picks the biggest difference between the two scenes
Prompts a coding agent to make a single change to bring it closer
Watching it struggle was pretty painful.
Next experiment will be similar but with separate asset generation and scene composition steps and the ability to 'steer' it as it goes so I can chime in with advice if I see it's stuck.
I experimented a while with creating our scenes with simple image gen models but I didn't like how static they looked.
However the image models are VERY good at greening out specific areas that you can key out with a shader to reveal some animated elements.
Add some god rays and fairy dust on top and you've got yourself a pretty good looking scene.