ElixpoResearch
← Research index

preprint / not peer reviewed

Native Binary-Weight Training Outperforms Zero-Shot Sign Projection in MNIST Diffusion

By

Ayushman Bhattacharya, Anwesha Chakraborty, Nihal Gazi

Published · MonoDiffuse-TR-2026-01 · doi:10.5281/zenodo.23424654

Abstract

A controlled study of one-bit-weight diffusion training that separates binary representational limits from post-training projection failure using matched residual U-Net experiments, repeated seeds, controls, hashes, and runnable artifacts.

Source code ↗Publication record ↗Read PDF ↗BibTeX ↓

Reproducibility

Evidence available for inspection

The repository provides exact commands, controls, three-seed artifacts, hashes, publication packaging, and a versioned citation record.

Authors and affiliations

  • Ayushman BhattacharyaPollinations AI
  • Anwesha ChakrabortyJIS University
  • Nihal GaziIEM

Artifact licenses

Citation record

Citation and persistent identifiers

Bhattacharya, Chakraborty, and Gazi. Native Binary-Weight Training Outperforms Zero-Shot Sign Projection in MNIST Diffusion. MonoDiffuse Technical Report 2026-01 (2026). https://doi.org/10.5281/zenodo.23424654

Download BibTeX →Open DOI record →