01 / IMPLEMENTATION & CONTRIBUTION
What the work involves
Course notebooks implement VAE objectives, DDPM training and sampling, classifier-free guidance, and simplified diffusion posterior sampling. Donald’s deblurring report compares six guidance scales and four measurement-noise settings.
Technical depth
Convolutional VAE with KL weighting and generation-distribution comparison; time-conditioned U-Net diffusion from supplied scaffolding; two-pass classifier-free guidance; Tweedie reconstruction; self-adjoint blur likelihood correction; six guidance scales and four noise settings.
The project family
uiuc/ece598/mp1uiuc/ece598/MP2uiuc/ece598/mp2uiuc/ece598/MP302 / RESULTS
What came out of it
Generated-image artifacts and a qualitative deblurring study show the tradeoff: insufficient guidance loses the input, excessive guidance introduces artifacts, and higher measurement noise degrades reconstruction.
03 / SUPPORTING EVIDENCE
Follow the source
Implementation notes, project records, and supporting artifacts.
Source context & project scope
Coursework using supplied architectures and OpenAI guided-diffusion/pretrained weights; those upstream models are not Donald-authored.
The DPS report is qualitative and based on its selected image setup, not broad image-restoration performance.
No aggregate PSNR/SSIM or reproducibility benchmark was established.
mp2 lowercase is a setup duplicate; group it with MP2 rather than count it as another project.
Implemented convolutional VAE reparameterization, loss and multiscale MMD evaluation.
SOURCE · 2026-09-17Implemented DDPM epsilon loss, classifier-free conditional/unconditional loss and samplers.
SOURCE · 2026-09-17Implemented Tweedie estimate and simplified observation-guidance update inside supplied diffusion code.
SOURCE · 2026-09-17Filled Donald-named November 17, 2025 report; six guidance strengths and four noise levels.
SOURCE · 2026-09-17