Imagine a frontier model can program a custom sim engine which takes good advantage of the 72 GPUs of a NVL72 with 20TB memory …. and this 24/7 knows better physics, numerical analysis and low level optimization 🫠
Again, Fable 5.2 is insanely good at 3D modelling: here, it made its own take of that Lord of the Rings scene.
I let the model build its own engine for the water physics, simulating the river as a shallow-water sheet, then running its own 3D particle simulation (FLIP) at 6cm resolution to make the breaking wave
On the horses: anyone who tried to draw a galloping horse knows how hard it is: I gave the model an open static 3D model, then gave it a classic chronophotography of a gallop to build the rigging and animation from.
I had to correct a few things over a few days of iterating + waiting renderings, but it did well!
A friend was proudly sharing his vibe-coded pipeline tools, he is creating a Environment Variables hell but I think people have to learn this way these days.
So I let him 🔥 his tokens
* Domain Knowledge *
If a model knows about Apple ][ 0.25-track disk protection and can confidently write 6502 assembly to move the stepper motor, that's pretty good. I tried Qwen3.8-27b finally, impressed.
Just loaded 'Flint' ....
How can this be released ?? Ejected.
Now that AI can use Blender, I’m seeing more people use high-fidelity 3D to drive AI video.
IMO, that’s a bad idea if you want realism. Lower-fidelity input lets the model translate your intent better.
@Yokohara_h This is truly amazing time, and thank you so much for showing us all these truly unreal possibilities with these advanced models. A challenging time too, Yokohara san. 🙏
@Yuki_1t0 Ito san, thank you for sharing all these beautiful imageries on X and they definitely captured many CGI artists attentions worldwide. Your passion and talent are slowly opening many doors ahead.
@Yokohara_h incredible demo shows us how LLMs played direct operators in a mainstream CGI workflow setting. Soon, they use differentiable, procedural or SDF means to generate assets directly and DCC with GUI might become obsolete.
The 'real' energy cost is unsustainable ... 🙃
The next version of three-gpu-pathtracer with WebGPU support is published! Drop path tracing into any project for product viz, architectural renders, or generative art.
Check it out here & share what you're making ✨
https://t.co/OV0uLxj1Ob
#threejs#webgpu
This is NOT Jev.
Open source. Runs on your laptop. Decides in ~27 ms, about 200× faster than waiting on a hosted LLM.
Here it is playing Tetris by itself 👇
https://t.co/eq4gP53o2A