Treating Weights as Images: a Wild Idea to Generate Model Weights with Diffusion
zsakib_ · x · 2026-11-06
Developer zsakib proposes a provocative idea: instead of training from scratch, teach models to "regurgitate" existing weights exactly, then use latent diffusion/flow matching to generate new model weights — or even "inpaint" weights into any model.
His reasoning:
- Weights can be treated as images, and generative models already produce pixels;
- Perturbing weights by a tiny epsilon barely changes behavior, suggesting the weight space is generable;
- Training from scratch wastes compute since early-training patterns are learned repeatedly across models;
- The idea extends: treat earlier layers as given "images" and generate later ones, like a video/world model.
He flags the odd, non-uniform dimensions of weight tensors as the main obstacle. It's an unverified open-ended speculation that sparked discussion on redundant training.
More from Models
- 12B open Nemotron trained on Surge RL data beats GPT-5.6 on SWE-Bench Pro V2 — echen · 2026-11-06
- Boson AI's Higgs Realtime Duplex: one 45B MoE that listens, watches, talks and renders its own face — smolix · 2026-11-06
- Boson AI launches Higgs Realtime Duplex: 45B MoE full-duplex omni model that listens, talks and works simultaneously — smolix · 2026-11-06
- Boson AI's 45B MoE duplex model tops Big Bench Audio at 98.9, beating TML and Google Live — smolix · 2026-11-06
- Artificial Analysis posts Intelligence Index v5 results; Nanbeige4.3-33B scores 20 — ArtificialAnlys · 2026-11-06
- Nanbeige4.3-33B hallucinates less than Qwen3.8 27B but answers fewer correctly — ArtificialAnlys · 2026-11-06