@TekonBurns the one-shot three.js scenes are pretty and also the easiest test to cherry-pick.
the one I'd trust more is round two: change one mesh or the lighting without the rest falling apart. that's closer to how you actually work than generating the first file.
@LatentOperatorX same model, three harnesses. that's the test that actually matters.
i wouldn't pick a winner on one run though. what I'd watch is whether the horse really jumps or just slides, and if the camera stays with it. that's usually where these break, not the mesh.
@fraxken a read-only viewer that eats glb, fbx and obj is the unsexy part everyone actually needs.
how painful was the .csmodel conversion compared to the usual three.js loaders?
unreal 5.9 exporting to the web again is a big signal. the browser is becoming a real game platform
meanwhile three.js devs: we've been here the whole time 😂
@anxuanng turning real CT scans into GLB for the browser is tanta roba. 20 MB for 7 bodies is pretty light too, curious how it runs on mobile with all the parts toggled
@10_X_eng an agent finding blender through mcp on its own is wild. ai to 3d to the browser is exactly the pipeline i want for three.js. do you export straight to glb from steve?
same models through claude code, omp and a tiny bash-tool script. scores basically the same, cost per task up to 3x apart.
the problem isn't which harness you use, it's how you use it.
still on omp bc i can bend it however i want https://t.co/e1hOFuAk7H
Researchers ran the same models through Claude Code, OpenCode and a 100-line Python script with one bash tool
Scores came out basically the same
Cost per task didn't, up to 3x apart
Most of my side projects are also 100 lines of Python and a bash tool
Nobody's written a paper about them
@nabendu82@itsfazley sol 6.1 for me, the lighting feels way more alive. would love to see the same prompt on a small local model too, curious how far the gap really is
@GithubProjects an agent that calls you instead of you babysitting it is the right direction. the approve-by-voice part is what i'd steal: i'm building a fully local bot that controls devices on my network, and confirming risky actions is the hardest UX problem
@ryantsai_16 wait, a whole destructible taipei in the browser? 😂 how are you keeping physics smooth when a full block comes down, instancing + sleeping bodies?
every "smart" vacuum uploads a map of your house to someone's server. this one runs fully local and you build it yourself. this is tanta roba https://t.co/p2BeS3iPbJ
@SelfHostedSI this is the part people sleep on. write the skill once, run it with claudio when you need the big brain and with opencode + a local qwen when you don't. localhost only 😂
@AlandysGame@threejs this is the right answer to that take. the problem isn't vibe coding, it's stopping at the first prompt. 33 days of actual iteration shows. curious how you handled netcode, did claude help there too or was that all you?
a $10 esp32 can run face detection, a noise monitor or a wall climbing robot, and the code is free. web devs sleeping on hardware is wild, it's just another api with wires
https://t.co/tFcp2pUEyP
Built a browser based game with a tuk-tuk that you can honk through a tiny Indiranagar.
Used @threejs + WebGL for the 3D rendering, JS for the game logic, OpenStreetMap data for the streets, and @GeminiApp for generating few 3D models.
Play it here: https://t.co/l3dBvW4SGI
@UchihaSomil@threejs@GeminiApp OSM data for the streets is such a smart move, real map basically for free. how did the gemini models hold up perf-wise once loaded in three.js? building a 3d car showroom myself and asset weight is always the real boss fight