AAAI 2026
Jun-Hyun Bae, Wonyong Jo, Jaehyup Lee, Heechul Jung
Kyungpook National University
๐Ÿ“„ Paper ๐Ÿ’ป Code ๐ŸŽฌ Poster & Video

Presentation


Abstract

Text-to-image diffusion models utilize cross-attention to integrate textual information into the visual latent space, yet the transformation from text embeddings to latent features remains largely unexplored. We provide a mechanistic analysis of the output-value (OV) circuits within cross-attention layers through spectral analysis via singular value decomposition. Our analysis demonstrates that semantic concepts are encoded in low-dimensional subspaces spanned by singular vectors in OV circuits across cross-attention heads. To verify this, we intervene on concept-related components in the diffusion process, demonstrating that intervention on identified spectral components affects conceptual changes. We further validate these findings by examining visual outputs of isolated subspaces and their alignment with text embedding space. Through this mechanistic understanding, we demonstrate that simply nullifying these spectral components can achieve targeted concept removal with performance comparable to existing methods while providing interpretability. Our work reveals how cross-attention layers encode semantic concepts in spectral subspaces of OV circuits, providing mechanistic insights and enabling precise concept manipulation without retraining.


Overview

Stable Diffusion์˜ cross-attention์—์„œ Van Gogh, Monet, Picasso๋Š” ๊ฐ™์€ head๋ฅผ ๊ณต์œ ํ•˜๋ฉด์„œ๋„ ์„œ๋กœ ๋‹ค๋ฅธ singular vector์— ๋†’์€ ๊ธฐ์—ฌ๋„๋ฅผ ๋ณด์˜€๋‹ค. ์ด ๋…ผ๋ฌธ์€ head ๋‚ด๋ถ€์˜ OV ๋ณ€ํ™˜์„ ๋ถ„ํ•ดํ•ด ์ด๋Ÿฐ ๊ฐœ๋…๋ณ„ ์ฐจ์ด๋ฅผ ๋ถ„์„ํ•œ๋‹ค.

OV ํ–‰๋ ฌ์„ SVD๋กœ ๋ถ„ํ•ดํ•˜๊ณ , โ€œa mountainโ€๊ณผ โ€œa mountain in Van Gogh styleโ€์ฒ˜๋Ÿผ ๊ฐœ๋… ์œ ๋ฌด๋งŒ ๋‹ค๋ฅธ prompt ์Œ์„ ๋น„๊ตํ•œ๋‹ค. ๋‘ prompt ์‚ฌ์ด์—์„œ ๋ฐ˜์‘์ด ํฌ๊ฒŒ ๋‹ฌ๋ผ์ง€๋Š” ์„ฑ๋ถ„์„ ๊ณจ๋ผ ์ค„์ด๊ฑฐ๋‚˜ ์ฆํญํ•˜๋ฉด, ์ƒ์„ฑ ์ด๋ฏธ์ง€์˜ ํ™”ํ’๋„ ๋‹ฌ๋ผ์ง„๋‹ค. ํ™”ํ’ ๊ด€๋ จ OV ์„ฑ๋ถ„๋งŒ ๋‚จ๊ธฐ๋Š” ์‹คํ—˜์—์„œ๋Š” ์žฅ๋ฉด์˜ ์„ธ๋ถ€ ๋‚ด์šฉ์ด ์‚ฌ๋ผ์ง„ ๋’ค์—๋„ ๋ถ“์งˆ๊ณผ ํ˜•ํƒœ๊ฐ€ ๋‚จ์•˜๋‹ค. Nudity ๊ด€๋ จ ์„ฑ๋ถ„์„ ์ œ๊ฑฐํ•˜๋Š” ๋ฐ๋„ ๊ฐ™์€ ๋ถ„์„์„ ์ ์šฉํ–ˆ๋‹ค.

Spectral Isolation

๊ฐœ๋…๋ณ„ spectral isolation. ๊ฐ ๊ฐœ๋…๋งˆ๋‹ค ์œ—์ค„์€ ๊ฐœ๋… ๊ด€๋ จ component๋ฅผ ์ œ๊ฑฐํ•œ ๊ฒฐ๊ณผ, ์•„๋žซ์ค„์€ ๊ทธ component๋งŒ ๋‚จ๊ธด ๊ฒฐ๊ณผ๋‹ค(์ƒ์œ„ k%). ์Šคํƒ€์ผ ๊ฐœ๋…(Van Gogh, Picasso)์€ ์žฅ๋ฉด ๋‚ด์šฉ๊ณผ ๋ถ„๋ฆฌ๋œ ๋ถ“์งˆยท์ž…์ฒดํŒŒ ํ˜•ํƒœ ๊ฐ™์€ ์Šคํƒ€์ผ ํŒจํ„ด๋งŒ ๋‚จ๋Š” ๋ฐ˜๋ฉด, ์ฝ˜ํ…์ธ  ๊ฐœ๋…(nudity)์€ ์ธ์ฒด ํ˜•ํƒœ๋ฅผ ํ†ต์งธ๋กœ ์žฌ๊ตฌ์„ฑํ•˜๋Š” holisticํ•œ ํ‘œํ˜„์„ ๋ณด์ธ๋‹ค. ์ด๋Š” ๋ชจ๋ธ์ด ๊ฐœ๋… ์œ ํ˜•์— ๋”ฐ๋ผ ์งˆ์ ์œผ๋กœ ๋‹ค๋ฅธ ์ธ์ฝ”๋”ฉ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•จ์„ ์‹œ์‚ฌํ•œ๋‹ค.


Method

Cross-attention์ด residual stream์— ๋”ํ•˜๋Š” ์ถœ๋ ฅ์€ ๋‘ ํšŒ๋กœ๋กœ ๋ถ„ํ•ด๋œ๋‹ค. QK circuit์€ attention ๊ฐ€์ค‘์น˜๋ฅผ ๊ณ„์‚ฐํ•ด ๊ฐ ์œ„์น˜๊ฐ€ ์–ด๋–ค ํ…์ŠคํŠธ ํ† ํฐ์„ ์ฐธ์กฐํ• ์ง€ ์ •ํ•˜๊ณ , OV circuit \(\mathbf{W}_{\text{OV}} = \mathbf{W}_{\text{V}}\mathbf{W}_{\text{O}}\)๋Š” ์ฐธ์กฐํ•œ ํ† ํฐ์˜ ์˜๋ฏธ๋ฅผ ์–ด๋–ค ์‹œ๊ฐ ํŠน์ง•์œผ๋กœ ๋ฐ”๊ฟ€์ง€ ์ •ํ•œ๋‹ค. Attention ํŒจํ„ด์€ denoising ๊ณผ์ •์—์„œ ๊ณ„์† ๋ฐ”๋€Œ์ง€๋งŒ OV ๋ณ€ํ™˜์€ ๊ณ ์ •๋œ ์„ ํ˜• ์‚ฌ์ƒ์ด์–ด์„œ, ํ…์ŠคํŠธ๊ฐ€ ์‹œ๊ฐ ํŠน์ง•์œผ๋กœ ๋ฒˆ์—ญ๋˜๋Š” ๊ณผ์ •์„ ์ •์ ์œผ๋กœ ๋ถ„์„ํ•˜๊ธฐ์— ์ ํ•ฉํ•˜๋‹ค.

Text embedding์€ ์˜๋ฏธ ์ •๋ณด๋ฅผ ๊ณ ์œ ํ•œ ์ถ•(semantic axes)์„ ๋”ฐ๋ผ ์กฐ์งํ•œ๋‹ค๊ณ  ์•Œ๋ ค์ ธ ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด ๊ณต๊ฐ„์„ ์‹œ๊ฐ ํŠน์ง•์œผ๋กœ ์„ ํ˜• ๋ณ€ํ™˜ํ•˜๋Š” \(\mathbf{W}_{\text{OV}}\) ์—ญ์‹œ ๊ทธ ์ถ•์— ์ •๋ ฌ๋œ subspace๋ฅผ ํ•™์Šตํ–ˆ์„ ๊ฒƒ์ด๋ผ๋Š” ๊ฐ€์„ค์—์„œ ์ถœ๋ฐœํ•œ๋‹ค. ๊ฐ head \(h\)์˜ OV ํ–‰๋ ฌ์„ SVD๋กœ ๋ถ„ํ•ดํ•˜๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

\[\mathbf{W}_{\text{OV}}^{(h)} = \sum_{i=1}^{r} \sigma_i^{(h)} \mathbf{u}_i^{(h)} (\mathbf{v}_i^{(h)})^\top\]

๊ฐ rank-1 ์„ฑ๋ถ„(spectral component)์€ text embedding ๊ณต๊ฐ„์˜ ๋ฐฉํ–ฅ \(\mathbf{u}_i\)๋ฅผ head ์ถœ๋ ฅ ๊ณต๊ฐ„์˜ ๋ฐฉํ–ฅ \(\mathbf{v}_i\)๋กœ ์„ธ๊ธฐ \(\sigma_i\)๋งŒํผ ์˜ฎ๊ธฐ๋Š”, ๋…๋ฆฝ๋œ text-to-visual ๋ณ€ํ™˜ ๊ฒฝ๋กœ๋‹ค. โ€œVan Gogh ์Šคํƒ€์ผโ€์ด๋‚˜ โ€œnudityโ€ ๊ฐ™์€ ์˜๋ฏธ ๊ฐœ๋…์€ ์ „์ฒด spectrum ์ค‘ ์†Œ์ˆ˜์˜ component์— ์ง‘์ค‘๋˜์–ด ์žˆ๋‹ค.

๊ฐœ๋… ๊ธฐ์—ฌ๋„ ์ธก์ •

๊ฐœ๋…์ด ์–ด๋””์— ์ธ์ฝ”๋”ฉ๋˜๋Š”์ง€ ์ฐพ๊ธฐ ์œ„ํ•ด, ๊ฐœ๋… ์œ ๋ฌด๋งŒ ๋‹ค๋ฅธ prompt ์Œ(์˜ˆ: โ€œa mountainโ€๊ณผ โ€œa mountain in Van Gogh styleโ€)์„ ๋น„๊ตํ•œ๋‹ค. ๊ฐ prompt์˜ token embedding์„ ํ‰๊ท ํ•œ \(\bar{\mathbf{c}}\)๋ฅผ head \(h\)์˜ singular vector์— ํˆฌ์˜ํ•˜๊ณ  singular value๋ฅผ ๊ณฑํ•ด \(\mathbf{s}^{(h)}(\bar{\mathbf{c}}) = \bar{\mathbf{c}}\,\mathbf{U}^{(h)}\mathbf{\Sigma}^{(h)}\)๋ฅผ ์–ป๋Š”๋‹ค. ๋‘ prompt ์‚ฌ์ด์—์„œ ์ด ๊ฐ’์ด ํฌ๊ฒŒ ๋‹ฌ๋ผ์ง€๋Š” ์„ฑ๋ถ„์„ ๊ฐœ๋…์— ๋งŽ์ด ๊ธฐ์—ฌํ•˜๋Š” ํ›„๋ณด๋กœ ๋ณธ๋‹ค. ์‹ค์ œ ์œ„์น˜๋ณ„ head ์ถœ๋ ฅ์—๋Š” ํ† ํฐ๋ณ„ attention ๊ฐ€์ค‘์น˜๋„ ๋ฐ˜์˜๋˜๋ฏ€๋กœ, ์ด ํ‰๊ท  ์ž„๋ฒ ๋”ฉ ๊ธฐ๋ฐ˜ ์ ์ˆ˜๋Š” ๊ฐœ๋… ์„ฑ๋ถ„์„ ์ฐพ๊ธฐ ์œ„ํ•œ ์ธก์ •๊ฐ’์ด๋‹ค.

\[\Delta\mathbf{s}^{(h)} = \mathbf{s}^{(h)}(\bar{\mathbf{c}}_{\text{concept}}) - \mathbf{s}^{(h)}(\bar{\mathbf{c}}_{\text{base}}), \qquad [\Delta\mathbf{s}^{(h)}]_i = \sigma_i^{(h)} \langle \bar{\mathbf{c}}_{\text{concept}} - \bar{\mathbf{c}}_{\text{base}},\, \mathbf{u}_i^{(h)} \rangle\]

๋ฒกํ„ฐ ์ „์ฒด์˜ ํฌ๊ธฐ \(\|\Delta\mathbf{s}^{(h)}\|_2\)๋Š” head์˜ ๊ฐœ๋… ๊ธฐ์—ฌ๋„์ด๊ณ , ๊ฐ ์›์†Œ์˜ ํฌ๊ธฐ \(|[\Delta\mathbf{s}^{(h)}]_i|\)๋Š” ๊ฐœ๋ณ„ spectral component์˜ ๊ธฐ์—ฌ๋„๋‹ค.

๋จผ์ € head ๋‹จ์œ„๋กœ ๋ณด๋ฉด, ๊ฐœ๋… ๊ธฐ์—ฌ๋„๊ฐ€ ๊ฐ€์žฅ ๋†’์€ ์ƒ์œ„ 20๊ฐœ head(SD v2.1 ์ „์ฒด 195๊ฐœ ์ค‘ ์•ฝ 10%)์˜ ์ถœ๋ ฅ๋งŒ ์Šค์ผ€์ผ๋งํ•ด๋„ ๊ฐœ๋…์˜ ๊ฐ•๋„๊ฐ€ ์กฐ์ ˆ๋œ๋‹ค.

Head Modulation

"Van Gogh" ๊ฐœ๋… ๊ธฐ์—ฌ๋„ ์ƒ์œ„ 20๊ฐœ head(์ „์ฒด 195๊ฐœ ์ค‘ ์•ฝ 10%)์˜ ์ถœ๋ ฅ์„ $\alpha$๋ฐฐ๋กœ ์Šค์ผ€์ผ๋งํ•œ ๊ฒฐ๊ณผ.

Head์—์„œ subspace๋กœ

Head ๋‹จ์œ„ ๋ถ„์„์€ ์–ด๋–ค head๊ฐ€ ๊ฐœ๋…์— ๊ธฐ์—ฌํ•˜๋Š”์ง€๋Š” ์•Œ๋ ค์ฃผ์ง€๋งŒ, ๊ฐœ๋…์ด head ์•ˆ์—์„œ ์–ด๋–ป๊ฒŒ ์กฐ์ง๋˜๋Š”์ง€๋Š” ๋ณด์—ฌ์ฃผ์ง€ ์•Š๋Š”๋‹ค. ์‹ ๊ฒฝ๋ง์˜ ํ•œ ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ์—ฌ๋Ÿฌ ํŠน์ง•์„ ํ•จ๊ป˜ ์ธ์ฝ”๋”ฉํ•˜๋Š” polysemanticity๋ฅผ ๊ณ ๋ คํ•˜๋ฉด, ๊ฐœ๋…์ด head ์ „์ฒด๋ฅผ ์“ฐ๋Š”์ง€ ์•„๋‹ˆ๋ฉด head ์•ˆ์˜ ํŠน์ • spectral component์— ๊ตญํ•œ๋˜๋Š”์ง€๊ฐ€ ์งˆ๋ฌธ์ด ๋œ๋‹ค. ์ด๋ฅผ ๋ณด๊ธฐ ์œ„ํ•ด ๋ชจ๋“  head์˜ ๋ชจ๋“  component๋ฅผ ๊ธฐ์—ฌ๋„ \(|[\Delta\mathbf{s}^{(h)}]_i|\)๋กœ ์ •๋ ฌํ•˜๊ณ , ์ƒ์œ„ \(k\)%๋ฅผ ๊ทธ ๊ฐœ๋…์˜ spectral signature \(\mathcal{S}_c\)๋กœ ์‚ผ์•„ ์ด component๋งŒ ์Šค์ผ€์ผ๋งํ•œ๋‹ค.

\[\widetilde{\mathbf{W}}_{\text{OV}}^{(h)} = \mathbf{W}_{\text{OV}}^{(h)} + (\alpha - 1) \sum_{i:\,(h,i) \in \mathcal{S}_c} \sigma_i^{(h)} \mathbf{u}_i^{(h)} (\mathbf{v}_i^{(h)})^\top\]

\(\alpha = 0\)์ด๋ฉด ์„ ํƒํ•œ component๋ฅผ ์ง€์šฐ๊ณ (Spectral Nullification), \(\alpha = 1\)์ด๋ฉด ์›๋ž˜ ๋ชจ๋ธ ๊ทธ๋Œ€๋กœ์ด๋ฉฐ, \(\alpha > 1\)์ด๋ฉด ํ•ด๋‹น component์˜ ๊ธฐ์—ฌ๋ฅผ ํ‚ค์šด๋‹ค. SD v2.1์˜ ์ •์„ฑ ์‹คํ—˜์—์„œ๋Š” ๊ฐœ๋… ๊ธฐ์—ฌ๋„ ์ƒ์œ„ 15โ€“20%๋ฅผ ์ œ๊ฑฐํ–ˆ์„ ๋•Œ ๋ชฉํ‘œ ๊ฐœ๋…์ด ์–ต์ œ๋˜๋Š” ์˜ˆ์‹œ๋ฅผ ํ™•์ธํ–ˆ๋‹ค. ๋ฐ˜๋ฉด 30% ์ด์ƒ ์ œ๊ฑฐํ•˜๋ฉด ์ „๋ฐ˜์ ์ธ ์ƒ์„ฑ ํ’ˆ์งˆ๋„ ๋–จ์–ด์งˆ ์ˆ˜ ์žˆ์—ˆ๋‹ค.

์•„๋ž˜ ๊ทธ๋ฆผ์€ ๊ธฐ์—ฌ๋„๊ฐ€ ๋†’์€ head ์ „์ฒด๋ฅผ ์กฐ์ ˆํ•˜๋Š” ๊ฒฝ์šฐ์™€, head ๋‚ด๋ถ€์˜ spectral component๋ฅผ ๊ณจ๋ผ ์กฐ์ ˆํ•˜๋Š” ๊ฒฝ์šฐ๋ฅผ ๋น„๊ตํ•œ๋‹ค. ๊ฐ๊ฐ ์ „์ฒด head์˜ 10%์™€ ์ „์ฒด component์˜ 10%๋ฅผ ์„ ํƒํ•˜๋ฏ€๋กœ, ๋‘ โ€œ10%โ€์˜ ๊ธฐ์ค€์€ ๋‹ค๋ฅด๋‹ค. ์ด ๋น„๊ต๋Š” head ๋‚ด๋ถ€๋ฅผ ๋” ์„ธ๋ฐ€ํ•˜๊ฒŒ ๋‚˜๋ˆ„์—ˆ์„ ๋•Œ ๊ฐœ๋…์„ ์กฐ์ ˆํ•˜๋Š” ๋ฐฉ์‹์ด ์–ด๋–ป๊ฒŒ ๋‹ฌ๋ผ์ง€๋Š”์ง€ ๋ณด์—ฌ์ค€๋‹ค.

Spectral vs Head

Spectral modulation(์œ„, ์ „์ฒด component ์ค‘ ๊ฐœ๋… ๊ธฐ์—ฌ๋„ ์ƒ์œ„ 10%)๊ณผ head-level modulation(์•„๋ž˜, ๊ธฐ์—ฌ๋„ ์ƒ์œ„ 10% head)์„ ฮฑ๋ฅผ ๋ฐ”๊ฟ” ๊ฐ€๋ฉฐ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ.

์•„๋ž˜ ๊ทธ๋ฆผ์€ Van Gogh, Monet, Picasso ์„ธ ๊ฐœ๋… ๋ชจ๋‘์—์„œ ๊ธฐ์—ฌ๋„๊ฐ€ ๋†’์€ head๋“ค์„ ๊ณจ๋ผ, head ์•ˆ์˜ singular vector๋ณ„ ๊ธฐ์—ฌ๋„๋ฅผ ๋‚˜ํƒ€๋‚ธ ๊ฒƒ์ด๋‹ค. ๊ฐ™์€ head ์•ˆ์—์„œ๋„ ์„ธ ๊ฐœ๋…์€ ์„œ๋กœ ๋‹ค๋ฅธ singular vector๋ฅผ ํ™œ์„ฑํ™”ํ•˜๋ฉฐ, ๋†’์€ ๊ธฐ์—ฌ๋ฅผ ํ•˜๋Š” singular vector๊ฐ€ ๋ฐ˜๋“œ์‹œ singular value๊ฐ€ ํฐ ๊ฒƒ(๋‚ฎ์€ ์ธ๋ฑ์Šค)์€ ์•„๋‹ˆ๋‹ค. ์ฆ‰ ๊ฐœ๋…๋“ค์€ head๋ฅผ ๊ณต์œ ํ•˜๋ฉด์„œ๋„, ๋ถ€๋ถ„์ ์œผ๋กœ ๊ฒน์น˜์ง€๋งŒ ๊ตฌ๋ณ„๋˜๋Š” ๊ณ ์œ ํ•œ ํ™œ์„ฑํ™” ํŒจํ„ด์„ ๊ฐ€์ง„๋‹ค.

Head Distribution

์„ธ ํ™”ํ’ ๊ฐœ๋…์—์„œ ๊ณตํ†ต์œผ๋กœ ๊ธฐ์—ฌ๋„๊ฐ€ ๋†’์€ head๋“ค์˜ singular vector๋ณ„ ๊ธฐ์—ฌ๋„. ๊ฐ™์€ head ์•ˆ์—์„œ๋„ ๊ฐœ๋…๋งˆ๋‹ค ์„œ๋กœ ๋‹ค๋ฅธ singular vector๊ฐ€ ํ™œ์„ฑํ™”๋œ๋‹ค.


Results

Concept Removal Benchmark

Spectral Nullification(SN)์˜ nudity ๊ฐœ๋… ์ œ๊ฑฐ ์„ฑ๋Šฅ์„ SD v1.4์—์„œ 5๊ฐœ adversarial prompt ๋ฒค์น˜๋งˆํฌ๋กœ ํ‰๊ฐ€ํ•œ๋‹ค. ๋ฒค์น˜๋งˆํฌ๋Š” Ring-A-Bell(K16, K38, K77)๊ณผ I2PยทMMAยทP4DยทUnLearnDiffAtk์ด๊ณ , SN์€ ์•ž์˜ ๊ด€์ฐฐ์— ๋”ฐ๋ผ ๊ฐœ๋… ๊ด€๋ จ spectral component ์ƒ์œ„ 20%๋ฅผ ์ œ๊ฑฐํ•œ๋‹ค. ํ‰๊ฐ€ ์ง€ํ‘œ๋Š” NudeNet์œผ๋กœ ์ธก์ •ํ•œ Attack Success Rate(ASR, ๋‚ฎ์„์ˆ˜๋ก ์ข‹์Œ)์ด๋‹ค.

Nudity ๊ด€๋ จ ์ฝ˜ํ…์ธ ์— ๋Œ€ํ•œ adversarial benchmark๋ณ„ Attack Success Rate (%, โ†“). ์—ด๋ณ„ ์ตœ์ €๊ฐ’์€ bold, 2์œ„๋Š” underline. ๋ฐฉ๋ฒ• ์œ ํ˜•์€ ๋ฐฐ๊ฒฝ์ƒ‰์œผ๋กœ ๊ตฌ๋ถ„ํ•œ๋‹ค โ€” ํšŒ์ƒ‰: training-based, ํŒŒ๋ž‘: closed-form, ์ดˆ๋ก: inference-time, ์ง„ํ•œ ํŒŒ๋ž‘: spectral.

MethodRing-A-BellI2PMMAP4DUnLearn
K16K38K77
SD v1.497.8994.7487.3725.0368.1069.7650.70
ESD76.8478.9574.7413.0424.8050.2426.06
CA88.4288.4284.2119.3058.5063.4144.37
MACE89.4795.7993.6825.5666.0068.2950.70
SDID95.7991.5884.2123.1262.0066.8348.59
UCE22.1118.9521.058.0641.0038.0521.13
RECE10.539.477.374.2425.0021.469.15
SLD-Medium68.4260.0050.538.3848.7043.9023.94
SLD-Strong18.9510.536.322.337.7011.717.04
SAFREE65.2655.7945.266.2629.9038.5414.79
SN (Ours)41.0535.7930.534.2417.6018.548.45

SN์€ I2P์—์„œ RECE์™€ ๊ณต๋™ 2์œ„(4.24%), MMAยทP4DยทUnLearnDiffAtk์—์„œ ๊ฐ๊ฐ 2์œ„๋ฅผ ๊ธฐ๋กํ•œ๋‹ค. ํ‘œ์— ํฌํ•จ๋œ training-based ๋ฐฉ๋ฒ•(ESD, CA, MACE, SDID)๋ณด๋‹ค ๋ชจ๋“  ๋ฒค์น˜๋งˆํฌ์—์„œ ASR์ด ๋‚ฎ์ง€๋งŒ, Ring-A-Bell์—์„œ๋Š” UCEยทRECEยทSLD-Strong๋ณด๋‹ค ๋†’๋‹ค. ๋”ฐ๋ผ์„œ ์„ฑ๋Šฅ์€ ๊ณต๊ฒฉ ํ”„๋กฌํ”„ํŠธ์˜ ๊ตฌ์„ฑ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง„๋‹ค.

SN์€ ์•ž์„œ ๋ถ„์„ํ•œ ๊ฐœ๋…๋ณ„ ์„ฑ๋ถ„์„ OV ํ–‰๋ ฌ์—์„œ ์ œ๊ฑฐํ•˜๋ฉฐ, ์ถ”๊ฐ€ ํ•™์Šต์€ ํ•˜์ง€ ์•Š๋Š”๋‹ค. ์ œ๊ฑฐ ์„ฑ๋Šฅ๊ณผ ํ…์ŠคํŠธโ€“์ด๋ฏธ์ง€ ์ผ์น˜๋„์˜ ๊ด€๊ณ„๋Š” COCO ์บก์…˜ 1,000๊ฐœ์— ๋Œ€ํ•œ CLIP score๋กœ ํ‰๊ฐ€ํ–ˆ๋‹ค. CLIP score๋Š” ์˜๋ฏธ์  ์ผ์น˜๋„๋ฅผ ์ธก์ •ํ•˜๋ฏ€๋กœ ์‹œ๊ฐ์  ํ’ˆ์งˆ์˜ ๋ชจ๋“  ์ธก๋ฉด์„ ํฌํ•จํ•˜์ง€๋Š” ์•Š๋Š”๋‹ค.

Quality Tradeoff

Concept removal ์„ฑ๋Šฅ(P4D ASR)๊ณผ ํ…์ŠคํŠธโ€“์ด๋ฏธ์ง€ ์ผ์น˜๋„(CLIP score)์˜ ๊ด€๊ณ„. SN์€ ์žฌํ•™์Šต ์—†์ด ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ๊ท ํ˜•์„ ๋ณด์ธ๋‹ค. SLD-Strong์€ ASR์ด ๊ฐ€์žฅ ๋‚ฎ์ง€๋งŒ CLIP score๋„ ๋‚ฎ์•„์ง„๋‹ค.

Spectral Subspace์˜ ์˜๋ฏธ ๊ฒ€์ฆ

์‹๋ณ„๋œ spectral subspace๊ฐ€ ์‹ค์ œ๋กœ ํ•ด๋‹น ๊ฐœ๋…์˜ ์˜๋ฏธ๋ฅผ ํฌ์ฐฉํ•˜๊ณ  ์žˆ๋Š”์ง€๋ฅผ ๋‘ ๊ฐ€์ง€ ๋ฐฉ์‹์œผ๋กœ ๊ฒ€์ฆํ•œ๋‹ค.

ํ…์ŠคํŠธ ๊ณต๊ฐ„ ์ •๋ ฌ: ๊ฐœ๋… prompt์™€ base prompt์˜ text embedding ์ฐจ์ด ๋ฒกํ„ฐ๋ฅผ concept-specific spectral component์˜ ์ž…๋ ฅ ๋ฐฉํ–ฅ์œผ๋กœ ์žฌ๊ตฌ์„ฑํ•œ ๋’ค, CLIP ์–ดํœ˜ 49,408๊ฐœ ํ† ํฐ ์ „์ฒด์™€์˜ cosine similarity๋ฅผ ๊ณ„์‚ฐํ•œ๋‹ค. Nudity ๊ฐœ๋…์˜ ๊ธฐ์—ฌ๋„ ์ƒ์œ„ 2๊ฐœ head๋ฅผ ๋ณด๋ฉด, ํ•œ head์—์„œ๋Š” โ€œnudesโ€, โ€œnsfwโ€, โ€œnakedโ€๊ฐ€, ๋‹ค๋ฅธ head์—์„œ๋Š” โ€œtoplessโ€, โ€œlingerieโ€, โ€œeroticaโ€, โ€œnudeโ€๊ฐ€ ์œ ์‚ฌ๋„ ์ƒ์œ„ 16๊ฐœ์— ํฌํ•จ๋œ๋‹ค. Head๋งˆ๋‹ค ๋ชฉ๋ก์€ ๋‹ค๋ฅด์ง€๋งŒ ๋ชจ๋‘ ๊ฐœ๋…๊ณผ ๊ด€๋ จ๋œ ํ† ํฐ์ด ๋‚˜ํƒ€๋‚œ๋‹ค. ๋™์‹œ์— ๋ฌด๊ด€ํ•œ ํ† ํฐ๋„ ์„ž์—ฌ ์žˆ์–ด, ๊ฐœ๋…์ด ์™„์ „ํžˆ ๋ถ„๋ฆฌ๋˜์–ด ์žˆ์ง€๋Š” ์•Š์Œ์„ ๋ณด์—ฌ์ค€๋‹ค.

Token Alignment

Nudity ๊ด€๋ จ spectral component๋กœ ์žฌ๊ตฌ์„ฑํ•œ ๋ฒกํ„ฐ์™€ cosine similarity๊ฐ€ ๊ฐ€์žฅ ๋†’์€ 16๊ฐœ ํ† ํฐ(๊ธฐ์—ฌ๋„ ์ƒ์œ„ 2๊ฐœ head). ์Œ์˜ ํ‘œ์‹œ๋œ ํ† ํฐ์ด nudity์™€ ์ง์ ‘ ๊ด€๋ จ๋œ ํ† ํฐ์ด๋‹ค.

๊ฐœ์ž… ์ „ํ›„์˜ ํ‘œํ˜„(t-SNE)๊ณผ ๊ฐœ๋… ๊ฐ„ ๊ฒน์นจ(Jaccard): ๊ฐœ๋… ๊ด€๋ จ component๋ฅผ ์ œ๊ฑฐํ•˜๋ฉด, head output์˜ t-SNE์—์„œ base/concept prompt ํด๋Ÿฌ์Šคํ„ฐ๊ฐ€ ๋” ๋งŽ์ด ๊ฒน์นœ๋‹ค. ์ด ์‹œ๊ฐํ™”๋Š” ์•ž์„œ ์„ฑ๋ถ„์„ ์ง์ ‘ ์กฐ์ ˆํ–ˆ์„ ๋•Œ ์ƒ์„ฑ ๊ฒฐ๊ณผ๊ฐ€ ๋‹ฌ๋ผ์ง„ ์‹คํ—˜์„ ๋ณด์™„ํ•œ๋‹ค. t-SNE๋งŒ์œผ๋กœ ์ธ๊ณผ๊ด€๊ณ„๋ฅผ ์ž…์ฆํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋‹ค. Jaccard similarity์—์„œ๋Š” Van Gogh์™€ Monet์ฒ˜๋Ÿผ ์œ ์‚ฌํ•œ ๊ฐœ๋…๋“ค์ด ์„ฑ๋ถ„์„ ๋” ๋งŽ์ด ๊ณต์œ ํ•˜๋ฉด์„œ๋„ ์„œ๋กœ ๊ตฌ๋ณ„๋˜๋Š” spectral signature๋ฅผ ๊ฐ–๋Š” ๋ชจ์Šต์„ ํ™•์ธํ•œ๋‹ค.

t-SNE Jaccard

(a) ๊ฐœ๋… ๊ธฐ์—ฌ๋„ ์ƒ์œ„ 10% component ์ œ๊ฑฐ ์ „ํ›„์˜ head output t-SNE. ์ œ๊ฑฐ ํ›„ base/concept ํด๋Ÿฌ์Šคํ„ฐ์˜ ๊ฒน์นจ์ด ๋Š˜์–ด๋‚œ๋‹ค. (b) ๊ฐœ๋… ๊ฐ„ Jaccard similarity. ์œ ์‚ฌํ•œ ๊ฐœ๋…๋ผ๋ฆฌ ๊ฒน์น˜์ง€๋งŒ, ์„œ๋กœ ๊ตฌ๋ณ„๋˜๋Š” signature๋ฅผ ๋ณด์ธ๋‹ค.

Qualitative

์•„๋ž˜ I2P adversarial prompt ์˜ˆ์‹œ์—์„œ๋Š” SN์ด ์žฅ๋ฉด ๊ตฌ์„ฑ์„ ์œ ์ง€ํ•˜๋ฉด์„œ ๋ชฉํ‘œ ์ฝ˜ํ…์ธ ๋ฅผ ์–ต์ œํ•˜๋Š” ๋ชจ์Šต์„ ๋ณผ ์ˆ˜ ์žˆ๋‹ค. ์ •๋Ÿ‰ ํ‘œ์—์„œ ๋ณด์•˜๋“ฏ ๋ชจ๋“  ํ”„๋กฌํ”„ํŠธ์—์„œ ์ œ๊ฑฐ๊ฐ€ ์„ฑ๊ณตํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋‹ค.

I2P Comparison

I2P adversarial prompt์— ๋Œ€ํ•œ ์ƒ์„ฑ ๊ฒฐ๊ณผ ๋น„๊ต. ์™ผ์ชฝ๋ถ€ํ„ฐ SD v1.4(์›๋ณธ), CA, RECE, SLD-Medium, SLD-Strong, SAFREE, SN(Ours).