Ablating Concepts in Text-to-Image Diffusion Models

Kumari, Nupur; Zhang, Bingliang; Wang, Shengyu; Shechtman, Eli; Zhang, Richard; Zhu, Jun‐Yan

doi:10.48550/arxiv.2303.13516

Cited by 2 publications

(5 citation statements)

References 51 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…2. In Ablation (Kumari et al 2023), the erasing concept is mapped to a broader "anchor" concept. While their loss is effective, the effect of it leaks to nearby concepts, similar to Dreambooth (Ruiz et al 2023).…”

Section: Related Workmentioning

confidence: 99%

“…Baselines. We compare the performance of our method with four different latest concept-erasing fine-tuning methods: ESD, SDD, "Ablating" (Kumari et al 2023), and Forget-Me-Not. Because of the applicability and the utility of a sexual-content censored model, our experiments are centered around erasing "nudity".…”

Section: Experiments Settingsmentioning

confidence: 99%

“…Still, alongside a high false positive rate, their near-unpredictable masking rate of images limits their applicational range, especially when the application relies on a continuous stream of data (Rando et al 2022). Given these complications, both computation-wise and performance-wise, of retraining these "foundational" models with a heavily curated dataset, researchers have proposed to directly fine-tune foundation models such as Stable Diffusion to erase target concepts (Gandikota et al 2023;Kumari et al 2023;Zhang et al 2023).…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models

Hong,

Lee,

Woo

2024

AAAI

View full text Add to dashboard Cite

Text-to-Image models such as Stable Diffusion have shown impressive image generation synthesis, thanks to the utilization of large-scale datasets. However, these datasets may contain sexually explicit, copyrighted, or undesirable content, which allows the model to directly generate them. Given that retraining these large models on individual concept deletion requests is infeasible, fine-tuning algorithms have been developed to tackle concept erasing in diffusion models. While these algorithms yield good concept erasure, they all present one of the following issues: 1) the corrupted feature space yields synthesis of disintegrated objects, 2) the initially synthesized content undergoes a divergence in both spatial structure and semantics in the generated images, and 3) sub-optimal training updates heighten the model's susceptibility to utility harm. These issues severely degrade the original utility of generative models. In this work, we present a new approach that solves all of these challenges. We take inspiration from the concept of classifier guidance and propose a surgical update on the classifier guidance term while constraining the drift of the unconditional score term. Furthermore, our algorithm empowers the user to select an alternative to the erasing concept, allowing for more controllability. Our experimental results show that our algorithm not only erases the target concept effectively but also preserves the model’s generation capability.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: Experiments Settingsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models

Hong,

Lee,

Woo

2024

AAAI

View full text Add to dashboard Cite

show abstract

“…At the same time, methods based on machine unlearning have also shown promising results. Kumari et al (2023) train the hidden state of sentences containing specified concept be closer to those without such concept. This can remove the model's capability to generate specific concept.…”

Section: Related Workmentioning

confidence: 99%

“…However, this method relies on a filtering model pre-trained on specific concepts, which makes it challenging to detect open-vocabulary visual concepts. Finally, the third approach, refining models (Gandikota et al 2023; Kumari et al 2023), aims at fine-tuning the whole or the part of models to learn and satisfy the administrator's requirements, thus enhancing the model's ability to adhere to the desired guidelines and produce content that aligns with the established rules and policies. However, these methods are often limited by the biases of tuning data, making it difficult to achieve open-vocabulary capabilities.…”

Section: Introductionmentioning

confidence: 99%

ORES: Open-Vocabulary Responsible Visual Synthesis

Ni,

Wu,

Wang

et al. 2024

AAAI

View full text Add to dashboard Cite

Avoiding synthesizing specific visual concepts is an essential challenge in responsible visual synthesis. However, the visual concept that needs to be avoided for responsible visual synthesis tends to be diverse, depending on the region, context, and usage scenarios. In this work, we formalize a new task, Open-vocabulary Responsible Visual Synthesis (ORES), where the synthesis model is able to avoid forbidden visual concepts while allowing users to input any desired content. To address this problem, we present a Two-stage Intervention (TIN) framework. By introducing 1) rewriting with learnable instruction through a large-scale language model (LLM) and 2) synthesizing with prompt intervention on a diffusion synthesis model, it can effectively synthesize images avoiding any concepts but following the user's query as much as possible. To evaluate on ORES, we provide a publicly available dataset, baseline models, and benchmark. Experimental results demonstrate the effectiveness of our method in reducing risks of image generation. Our work highlights the potential of LLMs in responsible visual synthesis. Our code and dataset is public available in https://github.com/kodenii/ORES.

show abstract

Ablating Concepts in Text-to-Image Diffusion Models

Cited by 2 publications

References 51 publications

All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models

All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models

ORES: Open-Vocabulary Responsible Visual Synthesis

Contact Info

Product

Resources

About