How far can a 1-bit neural network go at making pictures?
Phase 0 — Top: full-precision model · Bottom: 1-bit model. The shapes survive.
Most AI image generators run on big, high-precision numbers. MonoDiffuse asks the opposite question: what happens if you shrink the network down to a single bit?
We're pushing a 1-bit model — the smallest a network can possibly get — to its physical limit as an image generator that paints a picture out of pure random noise. It's a research project, built in two phases:
| Phase | The question | Status |
|---|---|---|
| Phase 0 | Can a 1-bit model learn the shape of real images without falling apart? | ✅ Done |
| Phase 1 — MonoDiffuse | Take it all the way: a fully 1-bit, color-image generator that's fast to run. | 🚧 In progress |
We compared a normal full-precision generator against a 1-bit version on handwritten digits. The short version: the tiny model gives up beauty, but keeps the meaning.
- It doesn't just memorize. The 1-bit model copies less from its training data — being small actually keeps it honest.
- It keeps the variety. No collapse into repeating the same few images.
- It loses fine texture. Results look more like clean stencils than photographs.
- But it stays readable. People (and a classifier) can still tell exactly what each image is — almost as easily as with the full-precision model.
There's also a fun quirk in how the two models "morph" one image into another: the big model cross-fades shapes, while the 1-bit model bends and curls them like reshaping a solid object.
Phase 1 takes the idea to its extreme and aims for color images, with three goals:
- Make everything 1-bit, not just part of the network.
- Start from pure random noise and let the model build a real picture out of it.
- Keep it fast — a lightweight "compressor" (a VAE) lets the model run quickly.
The bet: each piece of this has been tried separately before, but never all together. If it works, it's one of the smallest image generators ever built — and a clear answer to how far you can push a single bit.
pip install -r requirements.txt
# Reproduce the Phase 0 benchmark (trains both models, scores them)
python experiments/v0_mnist/rigorous_bench_trainer.py
# Generate samples from a saved model
python checkpoints/model_tester.pyPhase 1 lives in the monodiffuse/ package and is under active construction.
- Phase 0 (1-bit diffusion study) — original work by @nihal-gazi.
- Phase 1 (MonoDiffuse) — ongoing research building on that foundation.
MIT — free for academic and commercial use with attribution.