Precise Parameter Localization for Textual Generation in Diffusion Models

Have the authors tested whether fine-tuning only the localized layers with LoRA introduces any long-term risks, such as overfitting or loss of generalization? Are there any benchmarks or datasets where this approach might underperform relative to fine-tuning all cross-attention layers?

As presented in Figure 5 of our submission, fine-tuning only the localized layers with LoRA introduces a very small decrease in the general performance of the model as measured by Precision and Recall. This is in contrary to the fine-tuning of all cross-attention layers, where we can observe higher overfitting. To answer this question further, we ran an additional experiment that we included in Appendix A8. There, we show that our approach with single-layer fine-tuning outperforms full fine-tuning in additional scenarios with various amounts of data samples.

Why were SimpleBench and CreativeBench chosen as the primary benchmarks?

In our experiments, we follow benchmarks introduced with recent work by Yang et al. (NeurIPS 2023) [1] whose goal was to increase text generation capabilities of DMs. Thus, the benchmarks assess accurately the performance of text-to-image models in generating visually accurate text.

Would the method be equally effective on more complex, real-world benchmarks that include diverse text styles, such as cursive or stylized fonts?

In our evaluation, we focus solely on the quality of the generated text and the background. The evaluation is performed entirely with the generations and does not require access to any real-world images. Therefore, it will not benefit from more complex benchmarks. Nevertheless, we are grateful for your suggestions to look into additional features of generated texts. We run additional experiments that evaluate how our approach can be extended to this setup. In Appendix A.9 of the updated submission, we include the study on the localization of parameters controlling the style of the generated text. Our experiments indicate that the layer we localize for textual content does not control the style of the generated text. Furthermore, we attempt to find the parameters responsible for this text property and show that the style is controlled by at least seven layers of cross-attention in Stable Diffusion 3. This indicates that not all visual aspects of an image can be controlled using a single layer, making our contribution more unique.

Could the approach for localizing parameters influencing text generation extend to other applications? The paper would benefit from a discussion of possible applications beyond text generation and toxicity prevention, potentially sparking ideas for future work or new applications of the method.

Thank you for this valuable suggestion. Even though the main contribution of our work is the localization of a small part of the model responsible for the generation of textual content, we also introduced three applications where our observations can be used in practice. However, we agree that the list of potential benefits of our findings is not limited to the evaluated use cases. In particular, we think that we can easily adapt our method to the task of text removal or text addition, including an extension to other textual features such as styles or fonts for which we included initial experiments in the Appendix. As an interesting future work, we envision analyses in other domains, such as video generation, where layer specialization might be especially valuable considering the problem of consistency between video frames.

References:

[1] "GlyphControl: Glyph Conditional Control for Visual Text Generation" Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, Kai Chen. NeurIPS 2023.

Different attention layers tend to focus on varying aspects, such as content or style. This has already been explored in previous works. This paper applies this feature to textual generation, yet more clarification is needed to highlight the novelty of this contribution.

Our work is based on previous studies that tackled the problem of layer specialization. We overview such publications in the section "Interpretability of diffusion models" of the related work. Comparing to those approaches, we provide several new insights that were not studied before:

• To the best of our knowledge, this is the first work focusing on the localization of layers responsible for textual content generation.
• Contrary to recent studies, we extend our analyses over several large-scale models with various architectures (U-Net, Diffusion Transformers), training paradigms (diffusion, flow matching), and different text encoders (CLIP, T5). We show that our hypothesis holds independently of those design choices.
• On top of our main observation, we introduce several applications that directly benefit from the localization of text-specific layers. This includes parameter-efficient fine-tuning and safeguarding text-to-image models, which, to the best of our knowledge, were not associated with layers specialization before.

Even though this submission does not include a new algorithm, we still believe we have brought insights valuable to the ML community.

From some visualizations, it appears that the proposed approach not only alters the content of the text but also changes its style, such as the font. Are both style and content localized within the same layer?

We thank the reviewer for this suggestion, which poses an interesting research question. For all of our experiments in the original submission, we did not specify the style of the text in the source or the target prompt. Hence, we did not expect the textual style to be preserved. However, in Appendix A.9 of the updated submission, we conduct an additional experiment showing that the generated text style is not localized together with its content within the same layer. In the case of Stable Diffusion 3, textual style is, in fact, controlled by at least seven layers of cross-attention, which indicates that not all the visual aspects of an image can be controlled using a single layer.

There are many papers in text-to-image generation/editing working on cross-attention layers and manipulating the denoising steps through keys and values. The authors should consider doing more comprehensive survey and here are some of them:

We appreciate the Reviewer for highlighting these important works. We already included them in the updated version of our submission. Additionally, we provide a brief review below to illustrate their connection to our submission. While highly relevant, we note that these works are not directly comparable to our approach.

Multi-Concept Customization of Text-to-Image Diffusion In this work, the authors present an efficient way of customization of text-to-image diffusion models by fine-tuning a subset of cross-attention layer parameters. While their approach demonstrates that targeting the key and value matrices in all the cross-attention layers is sufficient to introduce new concepts, we reveal that fine-tuning those matrices in fewer than 5% of cross-attention layers (see Table 1 of submission) enables better quality of the generated text.

TokenFlow: Consistent Diffusion Features for Consistent Video Editing TokenFlow presents a framework that enables video editing using text-to-image diffusion models. Specifically, the authors introduce a method of editing the keyframes by extension of self-attention layers in which the keys from all timeframes are concatenated in order to encourage the frames to share a global appearance. The presented solution offers an effective approach to the semantic video edition, that's why it significantly differs from our work which targets image editing.

Localizing object-level shape variations with text-to-image diffusion models Prompt-Mixing enables users to explore different shapes of objects in an image. In order for objects to stay in the same positions but change their appearance, the method operates in the inference time and, in different denoising timestep intervals, injects different prompts into the cross-attention layers. In our work, we use a similar injection mechanism that we apply only to the selected text-controlling layers. We evaluate the effect of injection at different denoising steps in section A.2 in the Appendix, where we show that starting the patching from later timestep can positively affect both the quality of generated text and better preserve other visual attributes.

Grounded text-to-image synthesis with attention refocusing Cross-Attention Refocusing (CAR) is a calibration technique enabling better attending of tokens representing objects to image regions. By performing multiple intermediate latent optimizations by using CAR loss and Self-Attention Refocusing loss, authors achieve better controllability of the layout of generated objects. Similar to our work, CAR focuses on cross-attention maps but aims to strengthen attention to the correct token while reducing it elsewhere.

Plug-and-play diffusion features for text-driven image-to-image translation Plug-and-Play is an effective image-to-image translation method. In this work, authors show that in the denoising procedure, one can extract spatial features from the U-Net decoder's Residual Blocks and their following self-attention layers, obtaining encodings of the composition of the image. Next, by passing a different prompt during the denoising procedure for the same initial Gaussian noise, one can inject previously extracted features and obtain generations, differing in image attributes specified in the condition. In this work, we show that by focusing on text-related features, we can perform a precise edition by targeting a single attention layer.

Is there any specific reason to classify the target text into different buckets and use the most frequent texts as the validation set? It seems other buckets are not used.

In our experiments, we use all four out of four buckets from both benchmarks (SimpleBench and CreativeBench). For fairness of results, we use a bucket of the most frequent words as a validation set: on this set, we localize attention layers responsible for the content of the generated text and find the starting timesteps $t_s$ for patching. Then, we evaluate the performance of our method on the remaining three buckets, which constitute the test set. We described the details of these splits in lines 146-147 of our submission. We also note that we evaluate our fine-tuning and editing performance on test split in lines 349 and 408, respectively.

Lines 144 and 145 look like new paragraph names.

We thank the reviewer for pointing it out. We updated the text in this paragraph to eliminate confusion.

We are delighted that you found our submission worthy of acceptance. We would be happy to address any further questions and hope for your continued support during the discussion period.

The author mentioned "We focus on LoRA for SDXL since this model has a significantly lower text generation quality compared to other studied DMs". Can fine-tuned SDXL outperform SOTA methods on generation task with commonly used benchmark? If it can not, why not fine-tuning one of the SOTA models to show that the proposed method also works and can be used to improve not only SDXL? Only fine-tuning a model with lower performance is not convincing.

We evaluate both the base SDXL model and the SDXL model fine-tuned with LoRA on a benchmark for text editing, as shown in Table 3 of our submission. The improvement in the OCR F1 metric for the LoRA fine-tuned model over the base model (0.43 vs. 0.34) is solely attributable to the fine-tuning of localized cross-attention layers, demonstrating its effectiveness.

However, even the fine-tuned SDXL model achieves significantly lower results on the OCR F1 metric compared to the DeepFloyd IF (0.70) and SD3 (0.68) models. Following the reviewer's suggestion, we also attempted fine-tuning localized layers for these models, but we were not able to improve their performance. We hypothesize that this is due to the use of the T5 text encoder in both DeepFloyd IF and SD3, which proved to significantly enhance models' ability to generate high-quality text in images, as studied in [1].

[1] Esser, Patrick, et al. "Scaling rectified flow transformers for high-resolution image synthesis." Forty-first International Conference on Machine Learning. 2024.

More qualitative results (especially with more words to be generated) and human evaluation are suggested.

To answer this question, we conducted an additional study with the SD3 model, measuring how our localization-based text edition performs across texts with multiple words. We present the results of this comparison in the table below. We note that editing longer texts results in slightly higher values of the background preservation loss, which is understandable given that proportionally longer texts usually occupy more space in the image. Regarding the alignment between the generated text and the target prompt, we notice slightly worse performance for a larger number of words. Although the OCR F1-Score increases with the number of words, this metric is not comparable for texts of different lengths, and we find CLIP-T to be more reliable in this case. In Appendix A.11, we present example results of editing text for images with varying text lengths. The generations indicate that our method is also applicable in such a scenario.

Since MARIO-10M dataset is utilized in training, why is MARIO-Eval Benchmark not directly used to evaluate the generation performance in the paper?

Our evaluation of the model's fine-tuning does not require access to any real images, and therefore, we do not use images provided in the MARIO Eval benchmark. Instead, we rely solely on prompts, ensuring they are distinct from those used during fine-tuning. To measure the quality of the generated text, we utilize an external OCR to extract text from the generated images and calculate how closely it matches the text specified in the prompts.

For this evaluation, we use a test set of 300 prompts sourced from both SimpleBench and CreativeBench. We selected these benchmarks to align our assessment with prior works aiming to improve text generation quality in text-to-image diffusion[1].

[1] "GlyphControl: Glyph Conditional Control for Visual Text Generation" Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, Kai Chen. NeurIPS 2023.

How is the editing performance compared to other methods like textDiffuser?

In the text edition experiment, we compare our method's performance against Prompt-to-Prompt, which, similarly to us, uses only source and target prompts to modify an image generated by the diffusion model. The TextDiffuser method strongly relies on the Stable Diffusion 1.5 backbone with an additional layout prediction model. At the same time, our work utilizes newer backbone models but does not require any additionally trained models that improve the quality of generated text. Therefore, we believe that there is no possibility of a fair comparison between those two techniques.

If there are any further questions we are more than happy to answer them. If there are not, in light of those clarifications, we ask the Reviewer to consider raising the score.

Dear Reviewer RLCH, with the discussion period concluding on December 2nd AoE, we would appreciate you reviewing our responses and confirming whether they adequately address your questions and concerns about the submission.