ICML 2026·Interactive Research Monograph

A Unified Framework for
Diffusion Model Unlearning
with f-Divergence

Nicola Novello, Federico Fontana, Luigi Cinque, Deniz Gündüz, Andrea M. Tonello

GitHub Code
ICML 2026

Presentation Deck

7 slides

7-slide keynote walk-through: context → gap → method → results → conclusions. Full-screen, keyboard-navigable with radar chart competitor analysis.

ContextGapMethodResultsProofLandscape

Educational Masterclass

8 sections

Step-by-step derivation from first principles. Every equation derived from scratch with scroll-triggered animations.

The Problem → Diffusion Primer → Objective → Gradients → Convergence

Interactive Arcade

3 games

Three research-grade games: gradient descent visualizer, drag-and-drop proof ordering, and local convergence simulator.

G1G2G3

Appendix Explorer

3 appendices

Formal proof side-by-side with PhD-level commentary. Every Jacobian block, every convergence bound, explained step-by-step.

Objective derivation · Closed-form Gaussians · Convergence proof

Competitive Landscape

3 charts

Interactive charts comparing f-DMU vs 11 baselines. Exact numbers from Tab. 2 and Tab. 3: Van Gogh and nudity erasure.

RadarVan GoghNudity SD1.5

The Anchors

3 anchor types

Visualise and compare the three anchor types: Near (→ Dalí), Superclass (→ generic painting), Null (→ ∅). Visual examples with real images from the paper.

Near → DalíSuperclassNull → ∅

Van Gogh or Not Van Gogh?

2 modes

Test your eye: spot the Van Gogh style among generated, erased, and f-DMU-modified images. Two modes: identify the style or guess the anchor used.

Van Gogh or not?Guess the anchorNear / Superclass / Null
0.063
H-DMU I2P rate
vs CAbl 0.120
1/f″(1)
Convergence speed
H² = 2, KL = 1, χ² = ½
11
DM unlearning methods
unified under f-DMU