Meta tags:
description=Erasing undesired capabilities from a diffusion model using a fast data-free fine-tuning process.;

Headings (most frequently used words):

concepts, from, to, erasing, diffusion, erase, how, models, model, weights, artistic, style, more, results, on, erasure, why, what, edit, an, nudity, limitations, object, concurrent, work, cite, bibliography, bibtex,

Text of the page (most frequently used words):
the (102), images (22), model (21), erasing (19), and (19), concept (18), #diffusion (18), show (18), our (17), style (17), interference (16), erasure (15), that (15), with (15), method (15), concepts (14), erased (13), from (13), models (13), intended (12), weights (11), diagonal (11), off (11), dotted (11), erase (11), fine (10), boxes (10), are (10), while (9), artists (9), blue (9), for (8), using (8), represent (8), their (8), specific (8), unintended (8), styles (8), artistic (8), tuning (7), image (7), 2023 (7), such (7), prompt (6), this (6), can (6), text (6), when (6), esd (6), inference (5), iccv (5), paper (5), object (5), original (5), more (5), nsfw (5), but (5), other (5), nudity (5), stable (5), content (5), has (5), bau (4), trained (4), training (4), have (4), effects (4), rows (4), objects (4), red (4), visual (4), erases (4), guidance (4), undesired (4), propose (4), cross (4), attentions (4), not (4), activate (4), methods (4), find (3), rohit (3), gandikota (3), joanna (3), jaden (3), fiotto (3), kaufman (3), david (3), international (3), conference (3), computer (3), vision (3), how (3), arxiv (3), row (3), corresponding (3), both (3), particular (3), since (3), based (3), minimizing (3), study (3), where (3), attention (3), output (3), distribution (3), generation (3), they (3), about (2), frozen (2), pre (2), noise (2), given (2), edited (2), guide (2), than (2), very (2), effective (2), avoids (2), proceedings (2), ieee (2), materzyńska (2), zhang (2), preprint (2), erasures (2), part (2), box (2), above (2), each (2), same (2), seed (2), prompts (2), results (2), cases (2), approaches (2), targeted (2), large (2), some (2), filtered (2), generated (2), explicitly (2), words (2), unconditioned (2), parameters (2), compare (2), safe (2), latent (2), includes (2), measure (2), including (2), also (2), removing (2), one (2), last (2), column (2), only (2), wish (2), which (2), all (2), layers (2), used (2), present (2), self (2), conditioning (2), these (2), edit (2), conditioned (2), modifies (2), pretrained (2), ability (2), conditional (2), any (2), new (2), use (2), knowledge (2), capabilities (2), there (2), post (2), easy (2), permanently (2), data (2), generating (2), led (2), economic (2), issues (2), concern (2), copyrighted (2), because (2), recent (2), source (2), demo (2), lab, query, predict, then, train, opposite, direction, ideas, time, rather, objective, producing, generator, directly, inproceedings, gandikota2023erasing, title, author, materzy, nska, booktitle, year, bibtex, bibliography, appeared, cited, follows, cite, nupur, kumari, bingliang, sheng, wang, eli, shechtman, richard, jun, yan, zhu, 2303, 13516, ablating, concurrent, work, first, represents, generations, later, famous, incomplete, baseline, entire, classes, impose, trade, between, complete, limitations, across, categories, compared, sld, like, datasets, without, mentioning, apply, tune, sls, censorship, able, inappropriate, user, negative, narrowly, whereas, third, broadly, inspired, observation, applying, narrow, effect, focusing, changes, conditions, mentioned, finer, desirable, case, may, preserved, others, unconditional, except, creates, generalised, does, depend, presence, useful, way, situations, terms, car, modules, act, gateway, process, naturally, heads, certain, set, tokens, contrast, irrrespective, attend, aspects, what, classifier, free, tunes, scores, obtained, away, being, similar, motivation, behind, leads, straightforward, scheme, prediction, subtracting, component, compositional, energy, idea, simple, powerful, already, probabilities, named, goal, produce, reshapes, its, reducing, probability, according, encyclopedic, itself, unlearn, instead, collecting, dataset, intends, generative, mitigate, production, during, bypass, small, number, iterations, description, them, scale, mimic, vast, sets, surprising, capable, imitating, array, risks, impacts, create, concerned, dire, consequences, profession, tendency, echo, drawn, another, lawsuit, serious, institutions, who, release, trademarked, symbols, indiscriminately, effortlessly, imitate, sue, deepfake, porn, raises, consent, harrassment, why, wide, variety, art, even, just, inteference, type, advantage, over, previous, circumvent, yet, fast, practical, expense, retraining, whole, own, advancements, quality, growing, around, safety, privacy, works, attempt, restrict, via, classification, easily, circumvented, users, access, open, poster, huggingface, tuned, code, github, update, see, scales, unifies, debiasing, unified, editing, wacv, 2024, uce, equal, contribution, mit, csail, northeastern, university,

Text of the page (random words):
erasing concepts from diffusion models erasing concepts from diffusion models rohit gandikota 1 joanna materzyńska 2 jaden fiotto kaufman 1 david bau 1 1 northeastern university 2 mit csail equal contribution international conference on computer vision iccv 2023 update see our uce paper that scales erasing and unifies with debiasing unified concept editing in diffusion models wacv 2024 arxiv preprint source code github fine tuned model weights huggingface demo iccv demo iccv poster how to erase concepts from diffusion model weights with recent advancements in image generation quality there is a growing concern around safety privacy and copyrighted content in diffusion model generated images recent works attempt to restrict undesired content via inference methods or post generation classification but such methods can be easily circumvented when users have access to open source weights in this paper we propose a method for fine tuning model weights to erase concepts from diffusion models using their own knowledge given just the text of the concept to be erased our method can edit the model weights to erase the concept while minimizing the inteference with other concepts this type of fine tuning has an advantage over previous methods it is not easy to circumvent because it modifies weights yet it is fast and practical because it avoids the expense of retraining the whole model on filtered training data our method can be used to erase a wide variety of concepts including nudity art styles or even objects permanently from the model weights why erase concepts from diffusion models since large scale models such as stable diffusion are trained to mimic vast training data sets it is not surprising that they are capable of generating nudity imitating particular artistic styles or generating undesired objects these capabilities have led to an array of risks and economic impacts use of the models to create deepfake porn raises issues of consent and harrassment their ability to effortlessly imitate artistic styles has led artists to sue concerned about dire economic consequences for their profession and their tendency to echo copyrighted or trademarked symbols indiscriminately has drawn another lawsuit such issues are a serious concern for institutions who wish to release their models there are methods to mitigate such content post production or during inference but they are easy to bypass we propose fine tuning the weights for a small number of iterations using the text description of the targeted concepts to erase them permanently from the weights how to erase concepts from a model we use the encyclopedic knowledge of the model itself to unlearn a particular concept instead of collecting a new dataset of images corresponding to the concept that one intends to erase we propose using the generative capabilities of the pre trained model the idea is simple but powerful the pretrained model p θ x already has the ability to model conditional probabilities for any named concept c so our goal is to produce a new model p θ x that reshapes its distribution by reducing the probability of any image in the conditional distribution according to the original pretrained model this is similar to the motivation behind compositional energy based models in diffusion it leads to a straightforward fine tuning scheme that modifies the noise prediction model by subtracting a component conditioned on the concept to erase our erased stable diffusion esd method fine tunes a model using the conditioned and unconditioned scores obtained from the original frozen stable diffusion sd model to guide the output away from the concept being erased we query the frozen pre trained model to predict the noise for the given erasure prompt then we train the edited model to guide it in the opposite direction using the ideas of classifier free guidance at training time rather than inference we find that fine tuning model weights with this objective is very effective producing an edited image generator that directly avoids that concept that we have erased what weights to edit cross attention modules act as a gateway for text conditioning in the image generation process naturally these attention heads activate when a certain set of tokens are present in the text prompt in contrast self attentions activate irrrespective of text conditioning since they attend to the visual aspects a concept cross attentions activate only when car is present in the prompt but self attentions are activate in both the cases inspired by this observation we propose esd x applying the erasing method to only fine tuning cross attention parameters while erasing a concept that has a very narrow effect on the output distribution focusing changes on conditions when the concept is explicitly mentioned in the prompt such finer effects are desirable in case of artistic style erasure where some artists may wish to have their styles be preserved while others erased we also study esd u which is fine tuning all the unconditional layers all layers except cross attentions which creates a generalised erasure of a concept that does not depend on the presence of specific words in the prompt this is useful when removing nsfw output in a way that includes situations where nsfw terms are not used in the prompt esd x last column erases a concept narrowly while minimizing interference last 3 rows whereas esd u third column erases a concept broadly erasing an artistic style we study artistic style erasure our paper includes a user study where we measure compare the effects of using our method to erase an artistic style with other approaches including negative guidance and safe latent diffusion we also measure the interference effects of removing one style on other styles our method erases a style while minimizing undesired interference on other styles the blue dotted images represent the intended erasure while the off diagonal images represent undesired interference erasing nudity since nsfw content can be generated without explicitly mentioning words such as nudity we apply esd u to fine tune unconditioned parameters of the model in our paper we compare our method to both inference based guidance safe latent diffusion sls and training based censorship stable diffusion v2 0 and v2 1 and find that our method is able to erase more inappropriate content our method erases more nudity across categories compared to inference guidance sld or models like stable diffusion v2 0 that are trained on nsfw filtered datasets limitations for both nsfw erasure and artistic style erasure we find that our method is more effective than baseline approaches on erasing the targeted visual concept but when erasing large concepts such as entire object classes or some particular styles our method can impose a trade off between complete erasure of a visual concept and interference with other visual concepts cases of incomplete concept erasures and style interference with our method more results on artistic style erasure erasing famous styles the blue dotted boxes show images with intended style erased the off diagonal images show the unintended interference erasing specific artists the blue dotted boxes show images with intended style erased the off diagonal images show the unintended interference erasing specific artists the blue dotted boxes show images with intended style erased the off diagonal images show the unintended interference erasing specific artists the blue dotted boxes show images with intended style erased the off diagonal images show the unintended interference erasing specific artists the blue dotted boxes show images with intended style erased the off diagonal images show the unintended interference erasing specific artists the blue dotted boxes show images with intended style erased the off diagonal images show the unintended interference erasing specific artists the blue dotted boxes show images with intended style erased the off diagonal images show the unintended interference erasing specific artists the blue dotted boxes show images with intended style erased the off diagonal images show the unintended interference more results on object erasure we show the intended erasure of objects by our method part 1 the rows in red dotted box represent erasure of an object while the row above each of the red boxes represent their corresponding original sd image using the same seed and prompts we show the intended erasure of objects by our method part 2 the rows in red dotted box represent erasure of an object while the row above each of the red boxes represent their corresponding original sd image using the same seed and prompts interference effects of object erasure the first row represents the original sd generations from the later rows the diagonal images represent the intended erasures while the off diagonal images represent the interference concurrent work nupur kumari bingliang zhang sheng yu wang eli shechtman richard zhang jun yan zhu ablating concepts in text to image diffusion models arxiv preprint arxiv 2303 13516 2023 how to cite the paper appeared at iccv 2023 it can be cited as follows bibliography rohit gandikota joanna materzyńska jaden fiotto kaufman david bau erasing concepts from diffusion models proceedings of the 2023 ieee international conference on computer vision iccv 2023 bibtex inproceedings gandikota2023erasing title erasing concepts from diffusion models author rohit gandikota and joanna materzy nska and jaden fiotto kaufman and david bau booktitle proceedings of the 2023 ieee international conference on computer vision year 2023 about the bau lab