Can a diffusion model generate the weights of a small neural... - Vibeus
Can a diffusion model generate the weights of a small neural network from a text description? NNGen is an experimental attempt to do that, and its latest test report is candid: the generated networks have not yet passed the project’s own end-to-end success criteria.
Instead of denoising an image, the prototype learns to denoise a “canvas” containing the weights of a fixed-size controller. A text description conditions the generator; the resulting small network is then run in an environment, for example to guide a pen along a shape. The author shared screenshots of a spiral prompt and a generated trace that fell below the threshold.
The repository describes a self-generated corpus of 2,916 small controllers and a 2.3-million-parameter generator. Its new experiment log reports 0 successes in 28 sampled trials under its strict criteria, which include accuracy, coverage, closure and pace. Some outputs traced part of a curve but did not complete it. An earlier apparent success on held-out shapes was retracted after the success criteria were made stricter.
The author reports that longer sampling and changes to guidance did not solve the problem, and that text conditioning had little effect in the tested setup. The log identifies distribution mismatch and overfitting as observed problems and proposes a larger training corpus and best-validation checkpoints as the next experiment; whether those changes will work remains open.
This is a small research prototype, not a general method for producing useful models from arbitrary descriptions. In the discussion, readers pointed to earlier work on hypernetworks and weight generation and raised the difficulty of obtaining suitable training data. The reported results are the project author’s; they have not been independently reproduced here.