MonoDiffuse · not peer reviewed
Native Binary-Weight Training Outperforms Zero-Shot Sign Projection in MNIST Diffusion
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.
- Research status
- preprint
- Lifecycle
- active
- Relationship
- Elixpo-originated
- Registry state
- verified
- Reproducibility
- documented
- Repository license
- MIT
- Released
- 2026-10-17
Research context
Scope and public claim
MonoDiffuse studies conditional MNIST diffusion with conventional and W1A32 binary-neural-network variants. Its public record connects the source repository, MIT-licensed ONNX model artifacts, and the Diffuino interactive Hugging Face demonstration.
Elixpo-originated. Created or operated as part of Elixpo. Copyright and artifact licences still remain repository-specific.
Reproducibility
documented
The repository provides exact commands, controls, three-seed artifacts, hashes, publication packaging, and a versioned citation record.
Review statement
not peer reviewed
Elixpo reports the review state explicitly and does not imply external peer review where none is recorded.
Runnable demonstrations
Inspect the work interactively.
Licensing
software and model release
MIT
Repository-level evidence
GitHub currently detects an MIT repository license. Papers, models, and other research artifacts may publish additional artifact-specific terms.
License checked 2026-10-17 · registry status checked 2026-10-17
Citation
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