-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
482 lines (433 loc) · 13.4 KB
/
index.html
File metadata and controls
482 lines (433 loc) · 13.4 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8"/>
<meta name="viewport" content="width=device-width, initial-scale=1.0"/>
<title>SAEmnesia: Erasing Concepts in Diffusion Models with Supervised Sparse Autoencoders</title>
<link rel="preconnect" href="https://fonts.googleapis.com"/>
<link href="https://fonts.googleapis.com/css2?family=DM+Serif+Display:ital@0;1&family=DM+Mono:wght@400;500&family=DM+Sans:ital,wght@0,300;0,400;0,500;1,300&display=swap" rel="stylesheet"/>
<style>
*, *::before, *::after { box-sizing: border-box; margin: 0; padding: 0; }
:root {
--bg: #f8f6f1;
--bg2: #f0ede6;
--ink: #1a1714;
--ink2: #4a4540;
--ink3: #8a8078;
--accent: #1c4f8a;
--accent2: #e8f0fa;
--rule: rgba(26,23,20,0.12);
--serif: 'DM Serif Display', Georgia, serif;
--sans: 'DM Sans', system-ui, sans-serif;
--mono: 'DM Mono', monospace;
}
body {
font-family: var(--sans);
background: var(--bg);
color: var(--ink);
line-height: 1.7;
font-size: 17px;
-webkit-font-smoothing: antialiased;
}
/* ── HERO ── */
.hero {
max-width: 860px;
margin: 0 auto;
padding: 80px 32px 64px;
text-align: center;
position: relative;
}
.venue-pill {
display: inline-flex;
align-items: center;
gap: 6px;
background: var(--accent2);
color: var(--accent);
font-family: var(--mono);
font-size: 13px;
font-weight: 500;
padding: 5px 14px;
border-radius: 100px;
border: 1px solid rgba(28,79,138,0.2);
margin-bottom: 28px;
letter-spacing: 0.02em;
}
.venue-dot {
width: 6px; height: 6px;
border-radius: 50%;
background: var(--accent);
flex-shrink: 0;
}
h1 {
font-family: var(--serif);
font-size: clamp(42px, 7vw, 68px);
font-weight: 400;
line-height: 1.08;
letter-spacing: -0.02em;
margin-bottom: 18px;
color: var(--ink);
}
h1 em {
font-style: italic;
color: var(--accent);
}
.subtitle {
font-size: 18px;
color: var(--ink2);
font-weight: 300;
max-width: 600px;
margin: 0 auto 36px;
line-height: 1.6;
}
/* authors */
.authors {
font-size: 15px;
color: var(--ink2);
margin-bottom: 8px;
}
.authors a {
color: var(--ink2);
text-decoration: underline;
text-underline-offset: 3px;
text-decoration-color: var(--rule);
}
.affil {
font-size: 13px;
color: var(--ink3);
font-family: var(--mono);
margin-bottom: 40px;
}
/* links row */
.links {
display: flex;
flex-wrap: wrap;
gap: 10px;
justify-content: center;
margin-bottom: 0;
}
.link-btn {
display: inline-flex;
align-items: center;
gap: 7px;
padding: 9px 20px;
border-radius: 8px;
font-size: 14px;
font-weight: 500;
text-decoration: none;
transition: opacity 0.15s, transform 0.15s;
border: 1px solid transparent;
cursor: default;
}
.link-btn:hover { opacity: 0.8; transform: translateY(-1px); }
.link-btn.primary {
background: var(--ink);
color: var(--bg);
border-color: var(--ink);
}
.link-btn.secondary {
background: transparent;
color: var(--ink2);
border-color: var(--rule);
}
.link-btn.disabled {
background: var(--bg2);
color: var(--ink3);
border-color: var(--rule);
cursor: default;
}
.link-btn.disabled:hover { opacity: 1; transform: none; }
.link-icon {
width: 16px; height: 16px;
opacity: 0.8;
flex-shrink: 0;
}
/* ── DIVIDER ── */
.rule {
max-width: 860px;
margin: 0 auto;
padding: 0 32px;
border-bottom: 1px solid var(--rule);
}
/* ── TEASER / COMING SOON BANNER ── */
.soon-section {
max-width: 860px;
margin: 64px auto;
padding: 0 32px;
}
.soon-card {
background: var(--bg2);
border: 1px solid var(--rule);
border-radius: 16px;
padding: 56px 48px;
text-align: center;
position: relative;
overflow: hidden;
}
.soon-card::before {
content: '';
position: absolute;
top: -40px; right: -40px;
width: 200px; height: 200px;
border-radius: 50%;
background: radial-gradient(circle, rgba(28,79,138,0.06) 0%, transparent 70%);
pointer-events: none;
}
.soon-label {
font-family: var(--mono);
font-size: 11px;
font-weight: 500;
letter-spacing: 0.12em;
text-transform: uppercase;
color: var(--ink3);
margin-bottom: 16px;
}
.soon-card h2 {
font-family: var(--serif);
font-size: 32px;
font-weight: 400;
color: var(--ink);
margin-bottom: 14px;
line-height: 1.2;
}
.soon-card p {
font-size: 16px;
color: var(--ink2);
max-width: 480px;
margin: 0 auto 28px;
font-weight: 300;
}
.watch-note {
display: inline-flex;
align-items: center;
gap: 8px;
font-family: var(--mono);
font-size: 13px;
color: var(--ink3);
background: var(--bg);
padding: 8px 16px;
border-radius: 8px;
border: 1px solid var(--rule);
}
/* ── ABSTRACT ── */
.section {
max-width: 720px;
margin: 0 auto 64px;
padding: 0 32px;
}
.section-label {
font-family: var(--mono);
font-size: 11px;
font-weight: 500;
letter-spacing: 0.12em;
text-transform: uppercase;
color: var(--ink3);
margin-bottom: 20px;
display: flex;
align-items: center;
gap: 12px;
}
.section-label::after {
content: '';
flex: 1;
height: 1px;
background: var(--rule);
}
.abstract-text {
font-size: 17px;
color: var(--ink2);
line-height: 1.8;
font-weight: 300;
}
/* ── KEY PROPERTIES ── */
.props-grid {
max-width: 860px;
margin: 0 auto 64px;
padding: 0 32px;
display: grid;
grid-template-columns: repeat(auto-fit, minmax(180px, 1fr));
gap: 16px;
}
.prop-card {
background: var(--bg);
border: 1px solid var(--rule);
border-radius: 12px;
padding: 24px 20px;
}
.prop-num {
font-family: var(--serif);
font-size: 38px;
font-weight: 400;
color: var(--accent);
line-height: 1;
margin-bottom: 8px;
}
.prop-label {
font-size: 13px;
font-weight: 500;
color: var(--ink);
margin-bottom: 4px;
}
.prop-desc {
font-size: 13px;
color: var(--ink3);
line-height: 1.5;
}
/* ── CITATION ── */
.citation-section {
max-width: 720px;
margin: 0 auto 80px;
padding: 0 32px;
}
.bibtex {
background: var(--ink);
color: #a8c4e0;
font-family: var(--mono);
font-size: 13px;
line-height: 1.7;
padding: 28px 28px;
border-radius: 12px;
overflow-x: auto;
white-space: pre;
position: relative;
}
.bibtex .kw { color: #e8c77a; }
.bibtex .str { color: #9de8b4; }
.copy-btn {
position: absolute;
top: 14px; right: 14px;
background: rgba(255,255,255,0.08);
border: 1px solid rgba(255,255,255,0.12);
color: #aaa;
font-family: var(--mono);
font-size: 12px;
padding: 5px 12px;
border-radius: 6px;
cursor: pointer;
transition: background 0.15s;
}
.copy-btn:hover { background: rgba(255,255,255,0.14); color: #eee; }
/* ── FOOTER ── */
footer {
border-top: 1px solid var(--rule);
padding: 32px;
text-align: center;
font-size: 13px;
color: var(--ink3);
font-family: var(--mono);
}
footer a { color: var(--ink3); text-decoration: underline; text-underline-offset: 3px; }
/* ── RESPONSIVE ── */
@media (max-width: 600px) {
.hero { padding: 56px 20px 48px; }
.soon-card { padding: 36px 24px; }
.props-grid { grid-template-columns: 1fr 1fr; }
.section, .citation-section, .soon-section { padding: 0 20px; }
}
</style>
</head>
<body>
<!-- HERO -->
<section class="hero">
<div class="venue-pill">
<span class="venue-dot"></span>
ICML 2026
</div>
<h1>SAE<em>mnesia</em></h1>
<p class="subtitle">
SAEmnesia: Erasing Concepts in Diffusion Models with Supervised Sparse Autoencoders
</p>
<!-- <p class="authors">
Riccardo Renzulli & EIDOSLab
</p> -->
<p class="affil">University of Turin</p>
<div class="links">
<a class="link-btn primary" href="https://github.com/EIDOSLAB/SAEmnesia" target="_blank">
<svg class="link-icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.8"><path d="M9 19c-5 1.5-5-2.5-7-3m14 6v-3.87a3.37 3.37 0 00-.94-2.61c3.14-.35 6.44-1.54 6.44-7A5.44 5.44 0 0020 4.77 5.07 5.07 0 0019.91 1S18.73.65 16 2.48a13.38 13.38 0 00-7 0C6.27.65 5.09 1 5.09 1A5.07 5.07 0 005 4.77a5.44 5.44 0 00-1.5 3.78c0 5.42 3.3 6.61 6.44 7A3.37 3.37 0 009 18.13V22"/></svg>
GitHub
</a>
<a class="link-btn primary" href="https://arxiv.org/abs/2509.21379">
<svg class="link-icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.8"><path d="M14 2H6a2 2 0 00-2 2v16a2 2 0 002 2h12a2 2 0 002-2V8z"/><polyline points="14 2 14 8 20 8"/><line x1="16" y1="13" x2="8" y2="13"/><line x1="16" y1="17" x2="8" y2="17"/><polyline points="10 9 9 9 8 9"/></svg>
Paper
</a>
<a class="link-btn disabled" href="#">
<svg class="link-icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.8"><rect x="3" y="3" width="18" height="18" rx="2"/><path d="M3 9h18M9 21V9"/></svg>
OpenReview (coming soon)
</a>
</div>
</section>
<div class="rule"></div>
<!-- COMING SOON CARD -->
<section class="soon-section">
<div class="soon-card">
<p class="soon-label">Status</p>
<h2>Project page coming soon</h2>
<p>
We are preparing the full page with figures, a method overview,
interactive demos, and benchmark results. Check back after the proceedings are published.
</p>
<div class="watch-note">
<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><path d="M1 12s4-8 11-8 11 8 11 8-4 8-11 8-11-8-11-8z"/><circle cx="12" cy="12" r="3"/></svg>
Star the repo to get notified
</div>
</div>
</section>
<!-- ABSTRACT -->
<section class="section">
<p class="section-label">Abstract</p>
<p class="abstract-text">
Concept unlearning in diffusion models is hampered by feature splitting, where concepts are distributed across many latent features, making their removal challenging and computationally expensive. We introduce SAEmnesia, a supervised sparse autoencoder framework that overcomes this by enforcing one-to-one concept-neuron mappings. By systematically labeling concepts during training, our method achieves feature centralization, binding each concept to a single, interpretable neuron. This enables highly targeted and efficient concept erasure. Compared to the state-of-the-art sparse autoencoder-based unlearning approach, SAEmnesia reduces hyperparameter search by 96.67% and achieves a 9.22% improvement on the UnlearnCanvas benchmark for objects. Our method also shows superior scalability in sequential unlearning, improving accuracy by 28.4% when removing nine objects, establishing a step forward for precise and controllable concept erasure. Moreover, SAEmnesia effectively suppresses nudity on the I2P benchmark and remains robust to adversarial attacks.
</p>
</section>
<!-- KEY PROPERTIES
<section style="max-width:860px;margin:0 auto 16px;padding:0 32px;">
<p class="section-label">Key properties</p>
</section>
<div class="props-grid">
<div class="prop-card">
<div class="prop-num">✦</div>
<p class="prop-label">Frozen backbone</p>
<p class="prop-desc">No fine-tuning of diffusion model weights — ever.</p>
</div>
<div class="prop-card">
<div class="prop-num">τ</div>
<p class="prop-label">Inference-time</p>
<p class="prop-desc">Unlearning applied via feature suppression at generation time.</p>
</div>
<div class="prop-card">
<div class="prop-num">↓</div>
<p class="prop-label">HPO reduction</p>
<p class="prop-desc">Substantially fewer hyperparameters to tune vs. SAeUron.</p>
</div>
<div class="prop-card">
<div class="prop-num">↑</div>
<p class="prop-label">Quantitative gains</p>
<p class="prop-desc">Improved effectiveness & efficiency scores over SAeUron.</p>
</div>
</div>
<div class="rule" style="margin-bottom:64px;"></div> -->
<!-- CITATION -->
<section class="citation-section">
<p class="section-label">Citation</p>
<div style="position:relative;">
<pre class="bibtex" id="bibtex"><span class="kw">@inproceedings</span>{saemnesia2026,
<span class="kw">title</span> = {<span class="str">SAEmnesia: Erasing Concepts in Diffusion Models with Supervised Sparse Autoencoders</span>},
<span class="kw">author</span> = {<span class="str">Enrico Cassano and Riccardo Renzulli and Marco Nurisso and Mirko Zaffaroni and Alan Perotti and Marco Grangetto</span>},
<span class="kw">booktitle</span> = {<span class="str">Forty-third International Conference on Machine Learning</span>},
<span class="kw">year</span> = {<span class="str">2026</span>},
<span class="kw">url</span> = {<span class="str">https://openreview.net/forum?id=Pa6EoOViOq</span>}
}</pre>
<button class="copy-btn" onclick="
navigator.clipboard.writeText(document.getElementById('bibtex').innerText);
this.textContent='Copied!';
setTimeout(()=>this.textContent='Copy',2000);
">Copy</button>
</div>
</section>
<!-- FOOTER -->
<footer>
<a href="https://github.com/EIDOSLAB" target="_blank">EIDOSLab</a> · University of Turin · 2026
·
<a href="https://github.com/EIDOSLAB/SAEmnesia" target="_blank">github.com/EIDOSLAB/SAEmnesia</a>
</footer>
</body>
</html>