AI research / Research prototype

EDiME.

Look beyond a convincing image.

WHAT IT DOES

An explainable diffusion research pipeline exploring why synthetic medical images succeed or fail.

WORK IN THIS REPOSITORY

Conditional diffusion research and explicit evaluation of unsuccessful outcomes.

THE INTERESTING DECISION

FID, PSNR, and SSIM are considered alongside downstream classifier accuracy. That classifier is a proxy for utility, not proof of clinical validity.

START WITH THE EVIDENCE240.95

README-reported run; substantial room for improvement.

Inspect the source ↗

BUILT WITH

PyTorchDDPMGrad-CAMPathMNIST

PROJECT TYPE

Research prototype

START HEREREADME & setup

01 / THE IDEA

Look beyond a convincing image.

The problem

A generated pathology image can look plausible while failing to preserve meaningful structure. Image-quality scores alone do not explain what the model learned or where it breaks down.

The approach

A conditional UNet diffusion pipeline combines classifier-free guidance and EMA stabilization with Grad-CAM, Integrated Gradients, and downstream classifier evaluation on the PathMNIST benchmark.

02 / THE SYSTEM AT A GLANCE

Follow the flow.

01PathMNIST
02Conditional diffusion
03Generated samples
04Explanations
05Evaluation

A simplified architecture overview. The repository documents the implementation details.

03 / THE ENGINEERING DECISIONS

Why it's built this way.

Evaluate utility as well as appearance

FID, PSNR, and SSIM are considered alongside downstream classifier accuracy. That classifier is a proxy for utility, not proof of clinical validity.

Inspect explanations

Grad-CAM and Integrated Gradients help examine influential regions and understand failures in generated examples.

Report unsuccessful results

The README records low structural fidelity and mode collapse. The work includes a diagnosis of conditioning and noise-schedule issues, rather than a claim of deployment readiness.

04 / WHAT TO LOOK AT

Evidence over adjectives.

FID · epoch 299240.95

README-reported run; substantial room for improvement.

SSIM · epoch 2990.072

Low structural fidelity in the reported evaluation.

Classifier accuracy22.22%

Reported synthetic-image downstream accuracy, not clinical approval.

Capabilities and result snapshots follow the project's current README, reviewed October 10, 2026. Read the source

05 / THE HONEST PART

What I'd improve next.

Current boundaries

Research only; this is not a clinical tool.
64×64 resolution limits structural realism.
Mode collapse and imperfect evaluation proxies limit the usefulness of the reported model.

The next iteration

Retrain with timestep conditioning from initialization, balanced sampling, and a revised noise schedule.

Open my Software & AI CV

KEEP EXPLORING

Ask the documents

An answer with something behind it.