cs.AIOct 18, 2025
SaveSELECT: Semantic Erasure Boundary Construction with Sibling Exclusive Counterfactuals
Organizations: Beijing University of Posts and Telecommunications, School of Cyberspace Security, Beijing, 100876
Abstract
Concept erasure aims to suppress undesirable concepts in text-to-image diffusion models while preserving their generative capabilities. Existing methods often redirect target concepts toward predefined semantic anchors, but the influence of anchor selection on the scope of concept erasure remains insufficiently understood. Through causal tracing, we reveal that concepts with more dispersed causal effects tend to exhibit higher anchor sensitivity and lower erasure robustness. Controlled anchor interventions demonstrate that anchor selection affects target suppression and collateral damage to surrounding semantic concepts. These findings motivate us to formulate anchor selection as semantic erasure boundary construction. We propose SELECT (Selective Erasure via Local Exclusive Counterfactual Targets), a plug-and-play framework based on Sibling Exclusive Concepts (SECs), which are semantically close to the target but mutually exclusive in its defining attributes. SELECT jointly constructs a counterfactual redirection endpoint and a local retention set to balance target suppression and semantic preservation. SELECT filters candidates using contextual activation, and uses semantic coherence to identify context compatible redirection endpoints, while the remaining SECs serve as local preservation constraints. Experiments on object, celebrity, artist-style, and NSFW erasure show that SELECT strengthens target suppression over fixed general anchors and achieves suppression-preservation trade-offs, while reducing concept re-emergence and erosion.
Figures & tables
Figure 1: Causal tracing across concept categories. (a) Restoration-based intervention framework. We inject one cross-attention layer from the erased checkpoint into the original model at a time and measure the resulting change in target response to quantify its causal contribution. (b) Category averaged causal distribution metrics ( , , , and ) and anchor sensitivity metrics ( , , and ) for artist-style, celebrity, and object concepts.
Figure 2: Consistent relationship between causal dispersion and anchor sensitivity within the object category. Dog shows more dispersed causal effects and greater sensitivity to anchor selection, resulting in less stable target suppression across different anchors. In contrast, Corgi exhibits more concentrated causal effects and more consistent erasure performance. The causal heatmaps and representative generations illustrate that the association between causal dispersion and anchor sensitivity also holds within the same semantic category.
Figure 3: Anchor selection jointly affects target suppression and the locality of collateral damage. (a) Damage is concentrated around the local semantic neighborhood, with boundary concepts affected more strongly than distant unrelated concepts. (b) Visualization of target suppression and non-target damage under different anchor types. (c) BCTS across ten targets and five anchor types. Sibling Exclusive anchors achieve the strongest overall suppression-preservation trade-off.
Figure 4: Effect of contextual activation on SEC anchor quality. For representative Car and Blood targets, low anchors suppress target associated cues under related context prompts, whereas high anchors retain stronger target related visual evidence and are more prone to re-emergence.
Figure 5: Semantic coherence and the suppression preservation trade-off. Candidate anchors for Corgi erasure are compared using , target CLIP change , and scene CLIP change . Higher coherence anchors provide stronger target suppression while preserving more of the non-target scene content.
Figure 6: Overview of SELECT. SELECT first generates Sibling Exclusive Concept candidates using an LLM, filters them using contextual activation, and ranks the remaining candidates by semantic coherence. The selected counterfactual anchor determines the redirection endpoint, while the remaining high quality SECs form the boundary retention set passed to the downstream erasure algorithm.
Figure 7: Qualitative comparison of object erasure under SPEED [ 24 ] , MACE [ 25 ] , and OCE, together with two standalone erasure frameworks, AGE and RECE. For the three downstream erasure methods, we compare the original model, fixed-anchor editing, and SELECT-enhanced editing. AGE and RECE are evaluated independently on corresponding target and non-target concepts. SELECT effectively suppresses target-specific characteristics while preserving recognizable non-target content in the illustrated examples, whereas alternative methods may exhibit incomplete suppression, semantic drift, or structural degradation.
Figure 8: Artist-style erasure and retention. SELECT reduces target-style characteristics while better maintaining the visual structure of retained styles.
| Results of GCD Detection (Celebrity) | ||||||||||||
| Erase | Retain | CLIP | MS-COCO 30K | |||||||||
| Method | Top1 | Top3 | Top5 | Top1 | Top3 | Top5 | Erase | Retain | CLIP | FID | ||
| SD v1.4 | 94.03 | 95.20 | 95.36 | 92.60 | 95.10 | 95.6 | 11.22 | 32.57 | 35.67 | 31.3 | - | |
| SPEED | General | 24.2 | 33.7 | 38.4 | 89.3 | 92.4 | 93.5 | 81.99 | 30.59 | 34.92 | 30.56 | 8.40 |
| Ours | 12.7 | 22.2 | 27 | 88.7 | 92.8 | 93.9 | 87.99 | 27.89 | 35.13 | 30.67 | 7.71 | |
| MACE | General | 2.49 | 4.66 | 6.27 | 79.02 | 84.94 | 86.71 | 87.30 | 26.69 | 35.15 | 29.67 | 9.84 |
Table 2: Quantitative evaluation of celebrity and NSFW erasure. For celebrity erasure, SELECT reduces erased-identity Top-1, Top-3, and Top-5 detection rates relative to the General anchor under both SPEED and MACE. For NSFW erasure, SELECT achieves the lowest total NudeNet detections under both methods.
Figure S1 : Semantic regions for anchor construction. Target-overlap anchors are too close and may retain target-defining attributes, whereas far-unrelated anchors induce excessive semantic displacement. SECs occupy a local yet attribute-exclusive region, providing a suitable counterfactual redirection space.
Figure S2 : Relationship between contextual activation and concept re-emergence. Bars denote the CLIP Top-1 concept re-emergence rate, while CRR is shown on a logarithmic right y-axis. Concepts with stronger context-induced activation generally exhibit more frequent re-emergence.
| Category | Concepts |
| Object | airplane, ambulance, apple, backpack, banana, bed, bicycle, bird, boat, book, bottle, bowl, bridge, broccoli, bus, butterfly, cake, camera, car, carrot, castle, cat, chair, clock, clothing, comb, corgi, cup, dog, doughnut, drum, Eiffel Tower, elephant, fire truck, fish, flower, fork, frog, fruit, garbage truck, giraffe, guitar, hamburger, hammer, hat, helmet, horse, ice cream, knife, lamp, laptop, lemon, lesser panda, machine, mango, Mickey Mouse, motorcycle, Naruto, orange, pear, phone, piano, pineapple, pizza, plate, pyramid, rabbit, rose, sandwich, saxophone, scissors, screwdriver, shark, shoe, snake, sneakers, sofa, Spider-Man, spoon, squirrel, statue, strawberry, table, taxi, television, tomato, toothbrush, towel, traffic light, train, tree, truck, trumpet, turtle, umbrella, vegetable, violin, watermelon, wrench, zebra. |
| Celebrity | 50 Cent, Aaron Paul, Adam Driver, Adam Sandler, Adele, Al Gore, Al Pacino, Alan Rickman, Alicia Keys, Amy Adams, Angelina Jolie, Anne Hathaway, Ariana Grande, Barack Obama, Ben Affleck, Benedict Cumberbatch, Bernie Sanders, Beyoncé, Bill Clinton, Brad Pitt, Bradley Cooper, Britney Spears, Bruce Willis, Bruno Mars, Bryan Cranston, Cameron Diaz, Celine Dion, Charlize Theron, Chris Evans, Chris Pratt, Christian Bale, Cristiano Ronaldo, David Beckham, David Bowie, Denzel Washington, Donald Trump, Drake, Drew Barrymore, Dua Lipa, Dwayne Johnson, Ed Sheeran, Elon Musk, Elton John, Emma Stone, Emma Watson, George Clooney, Halle Berry, Harrison Ford, Hillary Clinton, Hugh Jackman, Idris Elba, Jack Black, Jack Nicholson, Jackie Chan, Jake Gyllenhaal, Jamie Foxx, Jay Z, Jennifer Aniston, Jennifer Lawrence, Jennifer Lopez, Jim Carrey, Joaquin Phoenix, Joe Biden, Johnny Depp, Julia Roberts, Justin Bieber, Katy Perry, Keanu Reeves, Kevin Hart, Kim Kardashian, Kobe Bryant, Lady Gaga, Leonardo DiCaprio, Lionel Messi, Margot Robbie, Matt Damon, Matthew McConaughey, Miley Cyrus, Morgan Freeman, Natalie Portman, Nicole Kidman, Oprah Winfrey, Reese Witherspoon, Rihanna, Robert Downey Jr., Roger Federer, Sandra Bullock, Scarlett Johansson, Selena Gomez, Serena Williams, Shakira, Sofia Vergara, Stephen Curry, Steve Jobs, Taylor Swift, Vladimir Putin, Whoopi Goldberg, Will Smith, Zendaya, Zoe Saldana. |
| Artist Style | Agnes Martin style, Albert Bierstadt style, Alphonse Mucha style, Amedeo Modigliani style, Andre Derain style, Aubrey Beardsley style, Banksy style, Barnett Newman style, Berthe Morisot style, Camille Pissarro style, Canaletto style, Caravaggio style, Caspar David Friedrich style, Chaim Soutine style, Clyfford Still style, Dali style, David Hockney style, Dorothea Tanning style, Edgar Degas style, Edvard Munch style, Edward Hopper style, Egon Schiele style, El Greco style, El Lissitzky style, Ernst Ludwig Kirchner style, Eugene Delacroix style, Fernand Leger style, Fra Angelico style, Francis Bacon style, Francis Picabia style, Francisco Goya style, Franz Kline style, Franz Marc style, Frida Kahlo style, Georges Braque style, Georges Seurat style, Gerhard Richter style, Giorgio de Chirico style, Gustav Klimt style, Gustave Courbet style, Henri de Toulouse-Lautrec style, Henri Matisse style, Henri Rousseau style, Hieronymus Bosch style, Hilma af Klint style, Jackson Pollock style, Jacques-Louis David style, James Ensor style, Jean Arp style, Jean-Francois Millet style, Jean-Honore Fragonard style, Jean-Michel Basquiat style, Joan Miro style, Joan Mitchell style, Johannes Vermeer style, John Constable style, John Everett Millais style, Juan Gris style, Kandinsky style, Kazimir Malevich style, Keith Haring style, Leonardo da Vinci style, Leonora Carrington style, Man Ray style, Marc Chagall style, Marcel Duchamp style, Mark Rothko style, Mary Cassatt style, Max Ernst style, Michelangelo style, Monet style, Odilon Redon style, Paul Cezanne style, Paul Gauguin style, Paul Klee style, Picasso style, Piet Mondrian style, Pieter Bruegel the Elder style, Raphael style, Rembrandt style, Remedios Varo style, Rene Magritte style, Robert Delaunay style, Robert Motherwell style, Roy Lichtenstein style, Sandro Botticelli style, Sonia Delaunay style, Theo van Doesburg style, Thomas Cole style, Thomas Gainsborough style, Titian style, Van Gogh style, Warhol style, William-Adolphe Bouguereau style, William Blake style, Winslow Homer style, Yayoi Kusama style, Yves Tanguy style. |
Table S1: Complete concept lists used in the causal tracing experiments.
Figure S3 : Averaged category-level causal heatmaps for object, celebrity, and artist-style concepts.
Figure S4 : Anchor sensitivity under different causal dispersion patterns. Representative distributed (Dog) and concentrated (Corgi) concepts are edited using anchors of different semantic types. The more dispersed the causal effect is, the more susceptible it is to the influence of anchor selection, and it exhibits an unstable erasure effect.
Figure S5 : Distance-binned analysis of collateral damage. Damage is concentrated in nearby semantic-distance bins.
| Target | SELECT Anchor | Synonyms / Variants | Boundary Retain Set | Random Retain Set |
| cat | dog | domestic cat; house cat; feline | raccoon; lion; tiger; bear; wolf | piano; keyboard; toaster; microphone; projector |
| dog | cat | domestic dog; canine; pet dog | wolf; raccoon; lion; tiger; zebra | Pikachu; garbage truck; banana; statue; backpack |
| pig | sheep | swine; hog; domestic pig | cow; goat; donkey; llama; alpaca | suitcase; toaster; fork; phone; umbrella |
| corgi | dachshund | Welsh corgi; Pembroke corgi; corgi dog | poodle; beagle; basset hound; chihuahua; pug | Pikachu; banana; bicycle; statue; umbrella |
| garbage truck | excavator | refuse truck; trash truck; waste collection truck | front loader; grader; ambulance; taxi; fire truck | cat; pig; Pikachu; guitar; watermelon |
| banana | avocado | ripe banana; yellow banana; banana fruit | apple; orange; grape; pineapple; strawberry | cat; Pikachu; garbage truck; guitar; umbrella |
Table S6: Concept configurations for the nine object-erasure targets. For each target, we report the selected SELECT anchor, target synonyms/variants, and the concept retention sets used during model editing. The Boundary Retain Set contains local semantic neighbors preserved by SELECT, whereas the Random Retain Set provides preservation concepts for the fixed-anchor baselines.
| Parameter | Value |
| Erasure algorithm | SPEED |
| aug_num | 10 |
| threshold | |
| retain_scale | 0.05 |
| lambda / lamb | 0.0 |
| disable_filter | true |
Table S7: SPEED hyperparameters used for object erasure. The same configuration is used for Null-text, General= ground , and SEC editing.
Figure S6 : Additional qualitative comparison of non-target preservation under OCE, MACE, and SPEED. Within each three-row block, the rows correspond from top to bottom to the original model, fixed-anchor editing, and SELECT-enhanced editing. Images are generated using matched prompts and seeds across the three rows. SELECT preserves recognizable non-target semantics and coherent image structure more consistently, particularly in cases where fixed-anchor editing causes semantic drift or structural degradation.
Figure S7 : Additional qualitative examples of object erasure.
Figure S8 : Qualitative comparison of the original SD v1.4 model, SEC-anchor-only editing without boundary retention, full SELECT, and RECE.
| Concept | Celebrity concepts | Anchor concepts |
| Erasure Group | Adam Driver, Adriana Lima, Amber Heard, Amy Adams, Andrew Garfield, Angelina Jolie, Anjelica Huston, Anna Faris, Anna Kendrick, Anne Hathaway, Arnold Schwarzenegger, Barack Obama, Beth Behrs, Bill Clinton, Bob Dylan, Bob Marley, Bradley Cooper, Bruce Willis, Bryan Cranston, Cameron Diaz, Channing Tatum, Charlie Sheen, Charlize Theron, Chris Evans, Chris Hemsworth, Chris Pine, Chuck Norris, Courteney Cox, Demi Lovato, Drake, Drew Barrymore, Dwayne Johnson, Ed Sheeran, Elon Musk, Elvis Presley, Emma Stone, Frida Kahlo, George Clooney, Glenn Close, Gwyneth Paltrow, Harrison Ford, Hillary Clinton, Hugh Jackman, Idris Elba, Jake Gyllenhaal, James Franco, Jared Leto, Jason Momoa, Jennifer Aniston, Jennifer Lawrence, Jennifer Lopez, Jeremy Renner, Jessica Biel, Jessica Chastain, John Oliver, John Wayne, Johnny Depp, Julianne Hough, Justin Timberlake, Kate Bosworth, Kate Winslet, Leonardo Dicaprio, Margot Robbie, Mariah Carey, Meryl Streep, Mick Jagger, Mila Kunis, Milla Jovovich, Morgan Freeman, Nick Jonas, Nicolas Cage, Nicole Kidman, Octavia Spencer, Olivia Wilde, Oprah Winfrey, Paul Mccartney, Paul Walker, Peter Dinklage, Philip Seymour Hoffman, Reese Witherspoon, Richard Gere, Ricky Gervais, Rihanna, Robin Williams, Ronald Reagan, Ryan Gosling, Ryan Reynolds, Shia Labeouf, Shirley Temple, Spike Lee, Stan Lee, Theresa May, Tom Cruise, Tom Hanks, Tom Hardy, Tom Hiddleston, Whoopi Goldberg, Zac Efron, Zayn Malik, Melania Trump | Lupita-Nyong-o, Tilda-Swinton, Queen-Latifah, Viola-Davis, Lupita-Nyong-o, Keanu-Reeves, Mahershala-Ali, Tilda-Swinton, Hailee-Steinfeld, Lupita-Nyong-o, Danny-DeVito, Greta-Thunberg, Zooey-Deschanel, Lupita-Nyong-o, David-Bowie, Denzel-Washington, Robert-De-Niro, Kurt-Russell, Steve-Carell, Oscar-Isaac, Nelson-Mandela, Lupita-Nyong-o, Danny-DeVito, Lupita-Nyong-o, Tilda-Swinton, James-McAvoy, Lisa-Kudrow, Miley-Cyrus, Kendrick-Lamar, Alicia-Keys, Ryan-Phillippe, Niall-Horan, Sundar-Pichai, Cate-Blanchett, Natalie-Portman, Salvador-Dali, Mahershala-Ali, Lupita-Nyong-o, Keanu-Reeves, Danny-DeVito, Kamala-Harris, Denzel-Washington, Lupita-Nyong-o, Christian-Bale, Tilda-Swinton, Lupita-Nyong-o, Javier-Bardem, Pedro-Pascal, Danny-DeVito, Fergie, Cobie-Smulders, Mark-Ruffalo, Saoirse-Ronan, Rosamund-Pike, Clint-Eastwood, Michael-Fassbender, Selena-Gomez, Pablo-Picasso, Usher, Lupita-Nyong-o, Danny-DeVito, Saoirse-Ronan, Saoirse-Ronan, Melania-Gulić, Frances-McDormand, Albert-Einstein, Lupita-Nyong-o, Christian-Bale, Chiwetel-Ejiofor, Selena-Gomez, Willem-Dafoe, Lupita-Nyong-o, Beyonce, Sofia-Vergara, Samuel-L-Jackson, Sting, John-Cena, Al-Pacino, John-C-Reilly, Keanu-Reeves, Robert-De-Niro, Rowan-Atkinson, Keanu-Reeves, Eddie-Murphy, Denzel-Washington, Jude-Law, Lupita-Nyong-o, Rami-Malek, Raquel-Welch, Quentin-Tarantino, Jim-Shooter, Recep-Tayyip-Erdogan, Brad-Pitt, Gary-Oldman, Lupita-Nyong-o, Benedict-Cumberbatch, Tyler-Perry, Mahershala-Ali, Lupita-Nyong-o, Tilda-Swinton |
| Retention Group | Aaron Paul, Alec Baldwin, Amanda Seyfried, Amy Poehler, Amy Schumer, Amy Winehouse, Andy Samberg, Aretha Franklin, Avril Lavigne, Aziz Ansari, Barry Manilow, Ben Affleck, Ben Stiller, Benicio Del Toro, Bette Midler, Betty White, Bill Murray, Bill Nye, Britney Spears, Brittany Snow, Bruce Lee, Burt Reynolds, Charles Manson, Christie Brinkley, Christina Hendricks, Clint Eastwood, Countess Vaughn, Dakota Johnson, Dane Dehaan, David Bowie, David Tennant, Denise Richards, Doris Day, Dr Dre, Elizabeth Taylor, Emma Roberts, Fred Rogers, Gal Gadot, George Bush, George Takei, Gillian Anderson, Gordon Ramsey, Halle Berry, Harry Dean Stanton, Harry Styles, Hayley Atwell, Heath Ledger, Henry Cavill, Jackie Chan, Jada Pinkett Smith, James Garner, Jason Statham, Jeff Bridges, Jennifer Connelly, Jensen Ackles, Jim Morrison, Jimmy Carter, Joan Rivers, John Lennon, Johnny Cash, Jon Hamm, Judy Garland, Julianne Moore, Justin Bieber, Kaley Cuoco, Kate Upton, Keanu Reeves, Kim Jong Un, Kirsten Dunst, Kristen Stewart, Krysten Ritter, Lana Del Rey, Leslie Jones, Lily Collins, Lindsay Lohan, Liv Tyler, Lizzy Caplan, Maggie Gyllenhaal, Matt Damon, Matt Smith, Matthew Mcconaughey, Maya Angelou, Megan Fox, Mel Gibson, Melanie Griffith, Michael Cera, Michael Ealy, Natalie Portman, Neil Degrasse Tyson, Niall Horan, Patrick Stewart, Paul Rudd, Paul Wesley, Pierce Brosnan, Prince, Queen Elizabeth, Rachel Dratch, Rachel Mcadams, Reba Mcentire, Robert De Niro | - |
Table S8: Celebrity erasure configuration. We use 100 identities for erasure and an independent set of 100 identities for retention evaluation.
Figure S9 : Representative celebrity-erasure and identity-retention results under SPEED and MACE.
Figure S10 : Artist-style erasure and retention. SELECT reduces target-style characteristics while better maintaining the visual structure of retained styles.
| Group | Artist styles | SELECT anchors |
| Erasure Group | ’brent-heighton’, ’brett-weston’, ’brett-whiteley’, ’brian-bolland’, ’brian-despain’, ’brian-froud’, ’brian-k.-vaughan’, ’brian-kesinger’, ’brian-mashburn’, ’brian-oldham’, ’brian-stelfreeze’, ’brian-sum’, ’briana-mora’, ’brice-marden’, ’bridget-bate-tichenor’, ’briton-rivière’, ’brooke-didonato’, ’brooke-shaden’, ’brothers-grimm’, ’brothers-hildebrandt’, ’bruce-munro’, ’bruce-nauman’, ’bruce-pennington’, ’bruce-timm’, ’bruno-catalano’, ’bruno-munari’, ’bruno-walpoth’, ’bryan-hitch’, ’butcher-billy’, ’c.-r.-w.-nevinson’, ’cagnaccio-di-san-pietro’, ’camille-corot’, ’camille-pissarro’, ’camille-walala’, ’canaletto’, ’candido-portinari’, ’carel-willink’, ’carl-barks’, ’carl-gustav-carus’, ’carl-holsoe’, ’carl-larsson’, ’carl-spitzweg’, ’carlo-crivelli’, ’carlos-schwabe’, ’carmen-saldana’, ’carne-griffiths’, ’casey-weldon’, ’caspar-david-friedrich’, ’cassius-marcellus-coolidge’, ’catrin-welz-stein’, ’cedric-peyravernay’, ’chad-knight’, ’chantal-joffe’, ’charles-addams’, ’charles-angrand’, ’charles-blackman’, ’charles-camoin’, ’charles-dana-gibson’, ’charles-e.-burchfield’, ’charles-gwathmey’, ’charles-le-brun’, ’charles-liu’, ’charles-schridde’, ’charles-schulz’, ’charles-spencelayh’, ’charles-vess’, ’charles-francois-daubigny’, ’charlie-bowater’, ’charline-von-heyl’, ’chaïm-soutine’, ’chen-zhen’, ’chesley-bonestell’, ’chiharu-shiota’, ’ching-yeh’, ’chip-zdarsky’, ’chris-claremont’, ’chris-cunningham’, ’chris-foss’, ’chris-leib’, ’chris-moore’, ’chris-ofili’, ’chris-saunders’, ’chris-turnham’, ’chris-uminga’, ’chris-van-allsburg’, ’chris-ware’, ’christian-dimitrov’, ’christian-grajewski’, ’christophe-vacher’, ’christopher-balaskas’, ’christopher-jin-baron’, ’chuck-close’, ’cicely-mary-barker’, ’cindy-sherman’, ’clara-miller-burd’, ’clara-peeters’, ’clarence-holbrook-carter’, ’claude-cahun’, ’claude-monet’, ’clemens-ascher’ | ’Abstract-Expressionism-style’, ’Academic-Art-style’, ’structured fauvist landscape painting style’, ’Aestheticism-style’, ’streamlined art-deco robot illustration style’, ’American-Realism-style’, ’Art-Brut-style’, ’Art-Deco-style’, ’Art-Nouveau-style’, ’documentary fine-art portrait photography style’, ’Ashcan-School-style’, ’hard-surface science-fiction concept art style’, ’Bauhaus-style’, ’Biopunk-style’, ’Byzantine-Art-style’, ’Celtic-Art-style’, ’Chiaroscuro-style’, ’Color-Field-Painting-style’, ’Conceptual-Art-style’, ’Constructivism-style’, ’Cubism-style’, ’Cyberpunk-Art-style’, ’Dadaism-style’, ’De-Stijl-style’, ’Deconstructivism-style’, ’Digital-Art-style’, ’Dutch-Golden-Age-Painting-style’, ’Earth-Art-style’, ’Expressionism-style’, ’Fauvism-style’, ’metaphysical realist figurative painting style’, ’Folk-Art-style’, ’Minimalist-vector-illustration-style’, ’Tachisme-style’, ’Glitch-Art-style’, ’Gothic-Art-style’, ’Graffiti-Art-style’, ’Hard-Edge-Painting-style’, ’Harlem-Renaissance-Art-style’, ’High-Renaissance-style’, ’Hudson-River-School-style’, ’Hyperrealism-style’, ’Impressionism-style’, ’Installation-Art-style’, ’Islamic-Architecture-style’, ’Japonisme-style’, ’Kinetic-Art-style’, ’De-Stijl-style’, ’Constructivism-style’, ’Light-and-Space-movement-style’, ’Bauhaus-style’, ’Luminism-style’, ’Lyrical-Abstraction-style’, ’Magic-Realism-style’, ’Cubism-style’, ’Maximalism-style’, ’Medieval-Art-style’, ’Memphis-Design-style’, ’Metaphysical-Art-style’, ’Fauvism-style’, ’Modernism-style’, ’Mughal-Painting-style’, ’Dadaism-style’, ’Neoclassicism-style’, ’Neo-Dada-style’, ’Neo-Expressionism-style’, ’Neo-Geo-style’, ’Neo-Impressionism-style’, ’Suprematism-style’, ’Op-Art-style’, ’Orphism-style’, ’Outsider-Art-style’, ’Performance-Art-style’, ’De-Stijl-style’, ’Photorealism-style’, ’Pixel-Art-style’, ’Pointillism-style’, ’Pop-Art-style’, ’Post-Impressionism-style’, ’Postmodernism-style’, ’Precisionism-style’, ’Pre-Raphaelite-Brotherhood-style’, ’flat-color mid-century architectural illustration style’, ’Realism-style’, ’Regionalism-style’, ’Rococo-style’, ’cinematic fantasy landscape concept art style’, ’atmospheric science-fiction environment concept art style’, ’Russian-Futurism-style’, ’Social-Realism-style’, ’Color-Field-Painting-style’, ’Street-Art-style’, ’Suprematism-style’, ’Surrealism-style’, ’decorative stained-glass illustration style’, ’Tachisme-style’, ’Glitch-Art-style’, ’Ukiyo-e-style’, ’Vaporwave-Art-style’, ’Vorticism-style’ |
| Retention Group | ’A.J.Casson’, ’Aaron Douglas’, ’Aaron Horkey’, ’Aaron Jasinski’, ’Aaron Siskind’, ’Abbott Fuller Graves’, ’Abbott Handerson Thayer’, ’Abdel Hadi Al Gazzar’, ’Abed Abdi’, ’Abigail Larson’, ’Abraham Mintchine’, ’Abraham Pether’, ’Abram Efimovich Arkhipov’, ’Adam Elsheimer’, ’Adam Hughes’, ’Adam Martinakis’, ’Adam Paquette’, ’Adi Granov’, ’Adolf Hiremy-Hirschl’, ’Adolph Gottlieb’, ’Adolph Menzel’, ’Adonna Khare’, ’Adriaen van Ostade’, ’Adriaen van Outrecht’, ’Adrian Donoghue’, ’Adrian Ghenie’, ’Adrian Paul Allinson’, ’Adrian Smith’, ’Adrian Tomine’, ’Adrianus Eversen’, ’Afarin Sajedi’, ’Affandi’, ’Aggi Erguna’, ’Agnes Cecile’, ’Agnes Lawrence Pelton’, ’Agnes Martin’, ’Agostino Arrivabene’, ’Agostino Tassi’, ’Ai Weiwei’, ’Ai Yazawa’, ’Akihiko Yoshida’, ’Akira Toriyama’, ’Akos Major’, ’Akseli Gallen-Kallela’, ’Al Capp’, ’Al Feldstein’, ’Al Williamson’, ’Alain Laboile’, ’Alan Bean’, ’Alan Davis’, ’Alan Kenny’, ’Alan Lee’, ’Alan Moore’, ’Alan Parry’, ’Alan Schaller’, ’Alasdair McLellan’, ’Alastair Magnaldo’, ’Alayna Lemmer’, ’Albert Benois’, ’Albert Bierstadt’, ’Albert Bloch’, ’Albert Dubois-Pillet’, ’Albert Eckhout’, ’Albert Edelfelt’, ’Albert Gleizes’, ’Albert Goodwin’, ’Albert Joseph Moore’, ’Albert Koetsier’, ’Albert Kotin’, ’Albert Lynch’, ’Albert Marquet’, ’Albert Pinkham Ryder’, ’Albert Robida’, ’Albert Servaes’, ’Albert Tucker’, ’Albert Watson’, ’Alberto Biasi’, ’Alberto Burri’, ’Alberto Giacometti’, ’Alberto Magnelli’, ’Alberto Seveso’, ’Alberto Sughi’, ’Alberto Vargas’, ’Albrecht Anker’, ’Albrecht Durer’, ’Alec Soth’, ’Alejandro Burdisio’, ’Alejandro Jodorowsky’, ’Aleksey Savrasov’, ’Aleksi Briclot’, ’Alena Aenami’, ’Alessandro Allori’, ’Alessandro Barbucci’, ’Alessandro Gottardo’, ’Alessio Albi’, ’Alex Alemany’, ’Alex Andreev’, ’Alex Colville’, ’Alex Figini’, ’Alex Garant’ | - |
Table S9: The Experiment Setup for Artistic Style Erasure.
| Concept | Sibling-Exclusive Concepts (SECs) |
| Nudity/naked | ”a firefighter in full gear”, |
| ”an arctic explorer in a heavy parka”, | |
| ”a beekeeper in a protective suit”, | |
| ”a welder in a leather apron and helmet”, | |
| ”a scientist in a cleanroom bunny suit”, | |
| ”a scuba diver in a full wetsuit”, |
Table S10: Sibling-Exclusive Concepts (SECs)
More in this topic
Concept Erasure
Concept erasure removes specified knowledge or attributes from models while preserving their ability to perform unrelated tasks.
See all 61 papers