MonoDiffuse MNIST checkpoints

Compact class-conditioned MNIST diffusion checkpoints from the MonoDiffuse project.

Files

• mnist_conditional_fp32.onnx: reliable FP32 conditional generator used by the Diffuino appliance.
• mnist_conditional_w1a32.onnx: native binary-forward W1A32 research checkpoint. Convolutional weights are binarized during the forward graph; activations and boundary/normalization operations remain floating point.
• Matching .pt files: PyTorch checkpoints.
• Matching .json files: training history, configuration, provenance, and checkpoint SHA-256 values.
• mnist_classifier.onnx: held-out MNIST evaluator used only to name generated samples in the demo.

Both generators are small residual U-Nets trained for 20 epochs on MNIST with a 1,000-step linear diffusion schedule. They accept image, timestep, and label inputs. Images use shape (1, 1, 28, 28) and range [-1, 1]; labels are integers from 0 through 9.

Important scope

The W1A32 model is an experimental binary-forward checkpoint, not a packed binary execution engine. ONNX Runtime still executes ordinary floating-point operators around the exported sign/scale graph, so these files make no latency or energy-efficiency claim. The associated paper's scientific result is deliberately limited to controlled MNIST experiments.

Source, training code, evaluation artifacts, and the Arduino UNO Q deployment are available in the repository. The exact controlled-study artifact commit is 604a34d.

Authors

Ayushman Bhattacharya, Anwesha Chakraborty, and Nihal Gazi.