Image generation / Training experiment
Diffusion Explained
Build a small diffusion model and inspect how its generated images change through training.
- My contribution
- I implemented the PyTorch training experiment and explanatory material described in the portfolio.
- Tools & methods
- PyTorch · DDPM · DDIM · ONNX
The problem
What this project investigates
A compact implementation makes the steps of image diffusion easier to inspect: adding noise, learning to predict it and sampling new images.
The build
What I built
The project describes a DDPM trained on CIFAR-10 with a U-Net, a linear noise schedule, DDIM sampling and exponential-moving-average weights. Saved sample grids provide a record of training output.
Technical choices
Decisions behind the build
Keep the training example compact
The project prioritises an understandable PyTorch implementation.
Distinguish explanation from inference
The original page explicitly labels its denoising animation as a simulation.
Current status
What the evidence shows
Training samples and external artifact addresses are included in the repository. ONNX parity is reported in the description but was not rerun here. The sample image is evidence of output, not a quality benchmark.
Further work & evidence gaps
The registry still marks this work in progress. Browser inference support needs verification before a live-demo claim.