ReFACT: Updating Text-to-Image Models by Editing the Text Encoder

We thank you for your comments and for striving to improve our work.

1. The closed form solution is indeed based on Meng et al.’s, which was shown to be an effective mathematical formulation. However, as we detail in section 3, adapting the method to the text-to-image generation setting required several non-trivial innovations that are specific to our work: (1) the choice to optimize the [eos] token based on CLIP’s training objective; (2) the choice of similarity function for the loss; and (3) most importantly, we introduced a contrastive approach for optimizing v*, which has led to significant improvements over the direct optimization approach in all metrics, as we discuss in the Appendix A. Note that the contrastive approach is not specific to our setting and could potentially also improve the results for Meng et al.’s setting in editing language models.

2. Regarding the amount of negative examples (now referred to as “contrastive examples” to distinguish from negative prompts in the test set) - using zero negative samples reduces to the direct optimization approach, which we discuss in Appendix A. As we mentioned, this variation was much less effective than the contrastive approach, i.e. an approach that does utilize negative examples. We ablated over the amount of negative examples and found 20 to be a reasonable number, far less than typically used in contrastive learning approaches [1, 2, 3]. We note that the negative prompts are not method specific as they are taken from the MS-COCO dataset. Moreover, the negative prompts are not specific to each of the editing cases, meaning that the same prompts are used throughout all of the editing cases we present in the paper.

3. We would like to clarify that ReFACT does not perform fine-tuning. Given k* and v*, ReFACT uses a closed-form solution to edit the specific layer weights.

4. We argue that currently - the alternative to editing for correcting outdated associations is to re-train the entire model, which requires far more resources. Our method is aimed for model providers and developers. As such, model providers that are aware of something that changed in the world (e.g., a new president is elected in the US), can use ReFACT to keep their model up-to-date, and avoid costly retraining and data curation. For example, without our editing approach, one would need to remove all photos of Donald Trump with descriptions of “the president of the US”, or similar, from the training data, and perform full training from scratch. With editing, only a description of the fact to edit is required in the form of an editing prompt (e.g., “The President of the United States”), a source prompt (e.g., “Donald Trump”) and a target prompt (e.g., “Joe Biden”).

5. As for unseen visual concepts - we discuss in section 7 the combination of personalization methods with ReFACT to achieve editing with novel entities and demonstrate a proof-of-concept in appendix I.

We hope that our response answered your concerns, and are happy to discuss further.

[1] Khosla, Prannay, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. “Supervised contrastive learning.”

[2] He, Kaiming, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. "Momentum contrast for unsupervised visual representation learning."

[3] Chen, Ting, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. "A simple framework for contrastive learning of visual representations."

Thank you for your review and for taking the time to improve our work.

1. As for the positive and negative prompts - during editing, only three prompts are used: the editing prompt (e.g., “The Prince of Wales”), the source prompt (e.g., “Prince Charles”) and the target prompt (e.g., “Prince William”). To edit the model weights, we embed the given editing prompt into different prompt templates (e.g., “A photo of the President of Wales”) and get k*. To calculate v*, we use negative examples (prompts) from the MS-COCO dataset. During evaluation, we use a completely different set of negative prompts. Thus, a prompt like “The Prince of Wales drinking coffee” is completely unseen. Given this prompt, the model edited by our method generates correct images, containing Prince William. Further examples are shown in Figure 17, which we added to the paper. Additionally, we have updated the paper and clarified the terminology by now referring to the negative examples used for editing as “contrastive examples” and distinguishing them from the “negative prompts” which are a part of the dataset, to avoid confusion.

2. The main results in the paper refer to the case of a single edit. In this case, a new “clean” pre-trained stable-diffusion instance is edited, and used to generate the images for evaluation. Using ReFACT to perform multiple edits, i.e. editing several associations on the same instance of the model is possible and discussed in section 5.4 and appendix H.

3. The FID and CLIP scores are indeed calculated on the MS-COCO dataset as is standard practice [1, 2, 3, 4] and as described in the last paragraph in section 4.3.

We hope our response addressed your concerns and would be happy to provide any further clarifications. If our response is satisfactory, would you consider increasing your score?

[1] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models.

[2] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding.

[3] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents.

[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. eDiff-I: Text-to-image diffusion models with an ensemble of expert denoisers

We thank you for the thorough and thoughtful review. We address your concerns and questions below.

1. As for the baseline you discussed - this is indeed one of the approaches used in our experiments, referred to as the “Oracle”, which we discuss briefly in section 4.2. While the oracle performs well, it does not achieve the goals of editing: the model itself still encodes outdated or incorrect associations and the extraction of the updated facts is left to the end user. The oracle is an unedited model, which generates a variation of the editing prompt that contains the target , in a similar manner to your suggestions. For example, if we wish to edit “The President of the United States” from “Donald Trump” to “Joe Biden”, an oracle prompt would be “Joe Biden as The President of the United States”. Our method, on the other hand, is aimed for model providers to ensure the outdated or incorrect facts are no longer generated by the model, without requiring the end user to perform prompt engineering. The performance of the Oracle baseline is detailed in Table 1. We have added clarification on this approach in the paper by adding more details to section 4.2.

2. As for generalization - Figure 5 provides some examples of prompts that use synonyms and different phrasing of the editing prompt, such as prime minister/PM, cat/kitten, apple/granny smith and The tower of Pisa/Pisa tower. We have added additional discussion on this point in section 5.1.

3. As for specificity - In the case of “The President of the United States”, our test set includes negative prompts such as “A politician” and “A congressman” which are similar in nature to your suggestion. However, we find your question interesting and thus performed further tests with the prompts “A president” and “The President”, shown in Figure 19 in the Appendix. Before editing, using 25 different seeds, only 4 and 5 images (respectively, for each prompt) generated images of Donald Trump. After editing, the images only show minor changes and remain mostly the same as they were before edits, while seeds that portrayed Donald Trump before editing now portray a different person, representing a generic notion of “president” (not Joe Biden).

4. As for scaling our method to perform multiple edits - we performed experiments on editing up to 90 facts by applying ReFACT sequentially to the same instance of the model. Our results, discussed in section 5.4 and in appendix H, show that using ReFACT for sequential edits work almost as well as single edits in all three metrics.

5. As for context driven edits - this is indeed a very interesting use case. The context is encoded in the key k*, thus the edit is context dependent. This can be seen, for example, on the right in Figure 6, where we edit Ice cream to be strawberry ice cream. In this case images of ice are left unaffected, which shows that the context of “cream” leads to a specific edit. In the case you suggested, we would have to better define what we expect the model to do in this case - if leaves are edited to be purple, do we still expect them to turn brown in the fall? Our experiment, added in Figure 17, shows that after editing “leaves” to “purple leaves” the prompt “an autumn leaf” indeed generates images of leaves that are purple, but a warmer autumn-like purple compared to the prompt “leaves”.

We hope that we addressed your questions and concerns and are happy to continue the discussion.

We thank you for the thorough review and for taking the time to consider our work and for your valuable comments. We provide answers to your questions and address your concerns below.

1. As for the dataset size - indeed our evaluation data, consisting of our newly collected dataset ROAD and the prior dataset TIME, contains around 200 samples. Nonetheless, we argue that this test size is sufficient for showing improvement over the previous method and general effectiveness of our method. Furthermore, each test sample is composed of 11 different prompts: an editing prompt, five positive prompts and five negative prompts. Each prompt in the test sample is used for image generation with 25 different seeds, thus achieving stable results over each test sample. Overall, we generated more than 50k images for evaluation.

2. Variety of prompts: ROAD captures both roles and appearances, which add variety to TIME dataset’s implicit assumptions. In ROAD there are a variety of roles including politicians, musicians and tv characters, and appearances for objects such as recognized brands, plant varieties, and car models, as discussed in section 4.1. We also find that our method is robust to synonyms and different phrasings of the editing prompt, such as prime minister/PM, cat/kitten, apple/granny smith and The tower of Pisa/Pisa tower. See Figure 5. We added an additional discussion on this point in section 5.1

3. As we discussed in section 2, personalization methods -- including Textual Inversion -- target a different task with different goals compared to editing. The main difference is that personalization methods add a special token (e.g., [v]) to distinguish the specific entity (“A [v] dog”) from the general class (“A dog”). Editing, however, should persistently alter the representation of the entity (e.g., “The President of the United States”) without requiring a specific token, and without preserving the original association (“Donald Trump”). Thus the two tasks are fundamentally different. Nonetheless, for comparison, we adapted Dreambooth, a personalization method which achieved superior performance on personalization datasets compared to Textual Inversion, to perform a variation of personalization that is related to editing (though not achieving the same goal). As we discussed in Section 6 and Appendix I, we found that it leads to inferior performance compared to ReFACT, results in images that are less diverse and demonstrates catastrophic forgetting.

We hope that we addressed your concerns, and are happy to continue a fruitful discussion.