Alex Nichol flagged that UnconvAI's Un-0 image generator, built around simulated Kuramoto oscillators, still leans on pretrained DINOv2 weights for its drifting loss and a small convolutional decoder to turn latent phases into pixels, so the physics-based claim sits inside a familiar neural-network sandwich rather than replacing it outright.
@sedielem It's sad though that it still uses pretrained DINO weights and a convolutional decoder. Not really a "pure" departure from mainstream NNs
A drifting model using coupled oscillators 👀
It manages to produce half-decent ImageNet 64x64 samples, which may seem like a toy experiment nowadays, but was basically SoTA only about a decade ago.
Love the detailed blog post! Takes me back to my reservoir computing roots😁
The largest variant runs 16,384 oscillators whose learnable coupling matrix and frequencies do most of the work before the decoder finishes the job; total parameters land around 322 million and the best FID on class-conditional ImageNet 64×64 reaches 6.74.
UnconvAI released everything on day one so researchers can test whether the oscillator approach eventually delivers the hoped-for efficiency jump or stays limited by the hybrid training setup.
Positive users appreciate the ablations in the researcher's critique of the Un-0 Model because they demonstrate the effectiveness of coupled oscillators over pretrained neural nets.
No Digg Deeper questions have been answered for this story yet.
@unixpickle Fair, although that's true for any drifting model currently :) I do like their ablations which show that the coupled oscillators are actually doing quite a bit of meaningful work!
@sedielem It's sad though that it still uses pretrained DINO weights and a convolutional decoder. Not really a "pure" departure from mainstream NNs