Not Just Pretty Pictures: Toward Interventional Data Augmentation Using Text-to-Image Generators

We thank Reviewer 3cd6 for their detailed summary and feedback. In particular, we appreciate their efforts in finding an important typo in our analysis, which we have now corrected (see Q1). We address their concerns below.

Q1: Figure 3 shows Text2Image is the best at RRSF, but statements on page 8 seem to contradict this.

We thank the reviewer for pointing out the typo. We fixed it in the revised draft (in red). Text2Image is indeed the most effective approach.

Q2.1: The effectiveness of Text2Image “undermine[s] the idea proposed in the paper”

Please refer to General Response (B).

Q2.2: “Where is the novelty” relative to [1]?

The goal of [1] is to show:

• it is possible to fine-tune stable diffusion to produce a class-conditional generative model whose synthetic samples mimic the training distribution (goal: reduce gap between synthetic and real)
• they test the performance on i.i.d samples with respect to the training set

Our goal is to:

• provide an extensive analysis of how to best use zero-shot generative techniques to simulate non-i.i.d. test domains (with no form of fine-tuning!)
• we test the performance on non-i.i.d. test domains, showing zero-shot generated samples can be used to reduce bias and improve generalisation across a wide variety of domains
• draw conclusions about how to most effectively exploit zero-shot generative capabilities of SD Thus, while [1] is an interesting related work (which is why we cite it in both Sections 1 and 2), the authors’ purpose, experimental setting, and results are all distinct from ours. For additional context, please consult our list of contributions in General Response (C).

Q3: “To further show that the generalization improves for unseen target domains, can the authors also show results for Cifar10-C and ImageNet-C datasets?”

We would like to point out that CIFAR-10-C and ImageNet-C are synthetic datasets obtained by applying some well-known parametric and handcrafted transformations to the images. Several of these transformations try to approximate the observation of different weather conditions or lighting conditions. We have experimented on 7 benchmarks, among which we have two large scale dataset (NICO++ and DomainNet), which represent some of these transformations and are made of naturalistic data – real photos captured in, e.g., dim light, autumn weather, etc.

However, as requested, we have also performed a comparable SDEdit experiment (using handcrafted prompts) on CIFAR-10-C:

Cifar-10-C Result with ResNet-18

It is important to note that Stable Diffusion (SD) is trained on web-scraped data, and as such we do not expect it to perform these kinds of artificial transformations as well as the naturalistic ones found in, e.g., NICO++. All the same, it appears that SD is still able to synthesize augmented data that can improve performance even on these unseen transformations. However, we note that, as these parametric transformations are artificially produced and mathematically well-understood (with implementations available in open-source libraries), it is obviously more efficient to target them with parametric augmentations at training time than with SD.

We appreciate the reviewer taking time to respond to our rebuttal.

P1: the fact that the "Text2Image" baseline performs better/comparable to SDEdit variant undermines the novelty of the paper

We respectfully disagree, because our novelty is not to establish SDEdit as the best performing method, it is to study a range of image synthesis and editing techniques (including both SDEdit and Text2Image) in the context of SDG and RRSF. We make a more detailed argument on this point in General Response (B) (Conditioning and Interventions); but in sum: it is a contribution, not a contradiction, that our investigation explores both ways to simulate interventions and finds Text2Image to often outperform SDEdit.

P2.1: [with respect to] [1], I feel the main idea is still the same

The main similarity between [1] and our work is the use of foundational generative models to generate synthetic data to train classifiers. (We do not claim novelty on this point, and in Section 2, we provide a detailed discussion of additional works that are also similar to our work in this respect.) Below, we note several key differences between our work and [1].

In [1], the authors show:

• Assuming access to all images of ImageNet, Imagen can produce the best approximation of ImageNet if fine-tuned on it (according to FID, IS and CAS scores)
• Adding the so-produced approximation of ImageNet to the ImageNet training set improves classification performance on ImageNet itself.
• Although the authors talk of performing augmentations in the abstract, the generated data is explicitly generated to resemble ImageNet. No deliberate manipulation of the image features is performed. The test set is i.i.d. with respect to the dataset.

In our work:

• Assuming no access to non-i.i.d. test domains training data, Stable Diffusion can approximate them zero-shot if some meta-data is accessible. No image is accessed from the non-i.i.d. test domains. This is a significantly more difficult problem that cannot be solved with more available i.i.d. training data (like [1] does). If, similarly to [1], we assumed full access to the test domain (non-i.i.d., in our case), the problem would not even require a generative model to be solved with perfect accuracy on PACS/Office-Home.
• Adding the zero-shot generated non-i.i.d. data to i.i.d. data, we improve the performance on non-i.i.d. test data (not accessible for any form of fine-tuning) and reduce the impact of biases on model's predictions.
• Our experiments in the main paper and rebuttal (Table 2 in the paper, Q2 response to reviewer MWyF) show exact knowledge of the meta-data is not required.
• We ablate and analyse which conditioning, prompting and filtering techniques produce the best performance, and compare with a retrieval baseline pointing out weaknesses and strengths of all approaches in different settings, as already extensively discussed in General Response (C) (Novelty and Contribution).

P2.2: the experimental setting and results are different to [1], I feel the main idea is still the same"

The different experimental setting is not a minor difference. The availability of i.i.d. training data and testing only on i.i.d. training data vs. the non-availability of non-i.i.d. data and testing on non-i.i.d. test data is the fundamental difference that distinguishes in-domain generalization (which [1] addresses) and out-of-domain generalization and reducing the reliance on spurious features (which we address). The two fields are fundamentally different areas of research: solutions and conclusions that hold in one do not necessarily generalize to the others. Indeed, under the assumptions of our setting, [1] would not be applicable (as there would be no non-i.i.d. data to fine-tune on).

Having carefully read [1], it is unclear why the novelty of the results of our work are diminished by paper [1], as they operate in very different settings with different methodologies, goals, and contributions.

We thank Reviewer nrcC for acknowledging the clear strengths and contribution of our work. We respond to their remaining concerns below.

Q1.1: “Not very soundness from technical side.”

Our experiments use several well-known and widely-utilized methodologies (e.g., Stable Diffusion-based image generation techniques including Text2Image, SDEdit, ControlNet, etc.). We are confident in the soundness of our implementation, which we have anonymized and made available at https://tinyurl.com/49sb675n. We are happy to engage in further discussion or experimentation to address any specific technical concerns the reviewer might have.

Q1.2: “No novelty in the model is presented.”

Our goal in this work is not to develop a novel methodology or model for performing interventional data augmentation (IDA). Rather, it is to:

1. Analyse the extent to which generative diffusion models, when used in a zero-shot setting, can serve as a general-purpose IDA mechanism.
2. For each application case (SDG, RRSF), discern the factors that contribute to their superior or inferior effectiveness (conditioning, filtering, prompting). No previous work has conducted the necessary study to address these questions, and our extensive findings, as described in General Response (C), are indeed novel.

Q2: Have you compared with GANs?

We thank the reviewer for this suggestion. In response, we carry out our SDG PACS experiment using VQGAN-CLIP [1] as a comparison and include the result in the selected table below and Table 3. We observe it underperforms with respect to Stable Diffusion. While this should not discourage future work from considering newly developed GANs, this reflects the superior generative capabilities of diffusion models established in the current literature [2].

SDG PACS result with ResNet-18 for VQGAN-CLIP comparison

[1] Crowson, Katherine, et al. "Vqgan-clip: Open domain image generation and editing with natural language guidance." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2022.

[2] Rombach, Robin, et al. "High-resolution image synthesis with latent diffusion models." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022.

Q3: “[P]lease provide more information on the prompts used for generating images on the three RRSF tasks [...] How much effort was spent on engineering them? How were they tuned?”

We asked a human participant to write up a set of prompts that would manipulate a specific feature of an image. We provided them only with information from metadata contained in the respective benchmark papers and the following high-level instructions:

• For the background-bias prompts, we asked them to specify common image backgrounds randomly.
• For the texture-bias prompts, we asked them to reason about what kind of transformations would change the texture of an object. Later, we asked the same participant to add a few additional prompts following the same instructions (see Appendix F.1).
• For the demographic bias, using the meta-data of the task was enough to create “Minimal” prompts (i.e., considering all possible combinations of gender/color), so no participant was required.

The prompts were not engineered nor tuned on any metric: our goal was to evaluate how well these systems perform “off-the-shelf” (i.e., where no time was spent engineering or tuning them). This illustrates the usefulness, flexibility, and simplicity of using off-the-shelf T2I generators for RRSF.

Q4: “Please spell out [the takeaways …] What are the implications for evaluation of generators and SDG/RRSF research?”

Please refer to our answers to General Response (B) (for context on our use of “conditioning” and “IDA”) and General Response (C) (for our response to this question).

We thank Reviewer bsQg for their comprehensive feedback, openness to improving their rating, and for clearly articulating their questions in detail. We believe the feedback helps us to further clarify the purpose of our work, and hope our response will stimulate an engaging and fruitful discussion.

Q1.1: “Premise Compromised?” Core premises and goals are unclear.

Please refer to General Response (A).

Q1.2 Is Text2Image interventional? Is it conditional?

Please refer to General Response (B).

Q1.3: Clearly state the contributions.

Please refer to General Response (C).

Q2.1 Why is sketch-domain performance “better without simulated target domain”?

The distribution of sketches produced by Stable Diffusion is substantially different from those represented in PACS (see Appendix I.5, Figure 16 of our updated draft). Since tuning the prompts to align the Stable Diffusion and PACS distributions is out of the scope of our analysis (see our response to Q3 below), this behavior is not unexpected.

Q2.2 “[I]nclude the baselines: (a) ERM trained on all but target domain, (b) ERM trained on all the domains, (c) ERM trained only on the target domain.”

We perform the requested experiments and report the results in the Table below (and in our updated manuscript, Table 14). Before consulting the results, please note the following:

1. Training on the target domains (b,c) is nonstandard in the context of SDG, as they are i.i.d. to the test sets. Thus, (b,c) do not constitute a baseline but an empirical “upper bound” on the best possible interventional augmentation strategy (i.e., where the generator can directly sample from a perfect approximation of the target distribution).
2. Experiment (a) constitutes a typical Multi-Domain Generalization (MDG) setting that is not comparable to the SDG we focus on, as the reported metrics for the two kinds of experiments are different:
   • In MDG, ERM is trained on all but the target domain and the performance on the unseen target domain is reported (column name is unseen target domain)
   • In SDG, the average of the performance across 3 unseen target domains is reported for each known source domain (column name is the only known training domain). Therefore the MDG results cannot be compared with the ones in Table 2 of the paper, so for a fair comparison with (a), we use SDEdit in the context of MDG (i.e., SDEdit(MDG)) using minimal prompts.

Our results are as follows:

• Unsurprisingly, (b,c) achieve near-perfect performance, confirming that the train and test datasets are nearly perfectly i.i.d.
• The performance improvement of SDEdit (MDG) with respect to (a) indicates the effectiveness of IDA also extends to the case where multiple training domains are available.

Q2.3: Why is SDEdit so much better than target-agnostic augmentation[s]?

Our results indicate that the target-agnostic augmentations are:

• Not as effective as commonly believed: most of them have been tested on benchmarks crafted from CIFAR and ImageNet but be less effective in other cases. On these datasets, a few of the methods we consider have already been shown not to be significantly more effective than ERM in the DomainBed benchmark (https://tinyurl.com/yeyjzvd9)
• Not as “target-agnostic” as commonly believed: our RRSF experiments clearly show they explicitly target specific types of invariances while neglecting others. On the other hand, SDEdit can manipulate shape, texture and other factors in more sophisticated ways, as captured by the 4 SDG and 3 RRSF benchmarks we experiment with.

Q2.4.1 Concerns about implementation.

We have integrated all baselines in the same codebase and ensured no method has an unfair edge over the others, using the same evaluation code across all methods. We have anonymized our codebase and made it available for review at https://tinyurl.com/49sb675n, and will release the de-anonymized codebase upon acceptance.

Q2.4.2 Can you experiment with using OfficeHome prompts on PACS?

We appreciate the interesting suggestion for this experiment. We added the prompts for Officehome to the ones of PACS to augment the PACS dataset. We include the results in the table below (and in Table 4 of the revised draft) indicating this method as SDEdit(PO-M). Note that adding these “irrelevant” prompts further improves performance.

Dear Reviewer bsQg,

We sincerely thank you for raising your rating. Thank you for recognizing our contribution in evaluating T2I models for SDG and RRSF and experiments and acknowledging the experiments are sound. We have indeed experimented with large-scale and challenging datasets with a wide variety of classes like NICO++ (61K imgs, 60 classes) and DomainNet (410K imgs, 345 classes; one of the most challenging benchmarks in the domain due to its outstanding complexity and scale) to an extent that is rarely seen in RRSF and SDG literature.

Regarding the tension between the following two statements:

1. “A general IDA mechanism […] has often been conjectured as a theoretical solution” (in the abstract)
2. “Text2Image (conditioning only on text) and Retrieval generally outperform SDEdit […] indicating that augmenting the original image is not necessary.” (in our General Response)

Specifically, the tension is this: our empirical findings contrast with earlier theoretical literature, in that we generally find that “augmenting the original image is not necessary.” In other words, when we empirically test the strength of the conjectured theoretical solution (e.g., an approach like SDEdit), we find that there are alternatives (e.g., an approach like Text2Image) that perform even better without direct augmentation (i.e., not requiring sample-level alignment, per General Response (B)). This is a key contribution of our work: it unearths an empirical limitation of the theoretical literature, pointing to the need for additional theoretical analysis in the context of IDA.

We will make this key point of tension and the associated empirical contribution clearer in our abstract and introduction. We will also strive for greater clarity on the remaining points you raised and will remove ambiguous statements.

About the limitations of prompting, we had already included a failure case in Appendix (on 2 datasets from the WILDS dataset mentioned by the reviewer, please refer to Figure 14) but rest assured we will emphasise this aspect more in the main paper.

Thank you again for your careful analysis of our response, we hope these lines address the remaining concerns about the framing of our work. Given the little time before the end of the discussion period, we will not be able to update the draft immediately to reflect these changes, but all the points discussed above will be modified considering your suggestions in the final draft.

Best Regards,

All Authors

We thank Reviewer MWyF for their detailed feedback. The potential weaknesses they outline are important to address, and we believe that our responses below are sufficient to address each concern.

Q1.1a: Assumption that “the target domains can be easily described.”

We agree that, in principle, not all domain shifts can be easily described via language; however our work clearly targets those that can. These represent a wide variety of problems of interest (we use 7 of the most popular benchmarks in the domain-shift literature). Indeed, reference [1] suggested by the reviewer provides further evidence that even when they are not easily describable, language-driven generation can achieve impressive performance on 3 additional benchmarks. In contrast, none of the non-textual baselines are able to outperform ERM on more than a small subset of the 7 benchmarks considered by our work, clearly indicating they are less general.

Q1.1b: Assumption that “the number of domains [is] known.”

While it is true that our primary SGD experiments assume that the number of domains is known, we have several key experiments (experiments in Table 2, the additional NICO++ we include in response to Q2 for SGD, and all RRSF experiments) that do not.

Q1.2: Human handcrafted prompts required? Expensive to attain?

For SDG, we experiment extensively with two prompting strategies that do not require human input: language enhancement (LE) and minimal (M). The same is very likely possible for RRSF, but for this task, we elected to focus on ablating the conditioning mechanism instead of the prompting strategy, as our results from SDG showed this to be more important. Regarding expense, we note that the handcrafted prompts took only a single participant 2-5 minutes per benchmark. (across all 3 RRSF benchmarks).

Q1.3: [1] does not “require knowledge of the domains.”

We thank the reviewer for mentioning [1]. However, we kindly note that : [1] assumes access to images of several domains to create the captions. Indeed, this is much more information than we access for, e.g., Minimal (M) prompts, which are produced with meta-data only. However, this requires explicit access to the target domain images, which is unavailable in our domain generalization setting. We will discuss the work in the revised draft.

[1] Diversify Your Vision Datasets with Automatic Diffusion-based Augmentation. NeurIPS’23

Q2: Table 2 seems to say that SD is biased to generate some domains that happen to be aligned with the target for this dataset, but may not be the case for other datasets.

We perform the experiment of Table 2 for NICO++, which has domains that are significantly not aligned (see the column names in the Table below). This clearly indicates the intervention performed aids generalization per-se as the improvement is substantially maintained when the target domain is not included.

SDG NICO++ Result with ResNet-50 for comparing the effect of accessing synthetic target domain or not

Q3.1: Why “different number of additional data for generated images and baseline[s]”?

The difference is caused by the current cost of generating samples with Stable Diffusion (SD). Since the process is expensive, we perform it offline before training:

• the augmentation baselines produce $D*E$ augmentations where $E= 50$ (the number of epochs), and $D$ is the size of the training set.
• the SD-based approaches generate $D*(C-1)$ samples, where $C$ is the number of domains. $C=4$ or $C=6$ across the SDG experiments.
• Since $(C-1)$ is much lower than 50, SD produces much less augmentations The superior performance of using fewer SD-based augmentations indicates interventions targeting the right environmental variables are significantly more effective than generic augmentations whose contribution to generalization is unclear.

Q3.2: How much does performance change with amounts of data similar to cited works?

In Appendix I.3 we already report this result for PACS. The conclusion is that a few additional data samples can help, but too much data can degrade the performance. The fact that few additional samples are enough to outperform all SOTA baselines indicates the slow generation is counterbalanced by the effectiveness of using just a few generated samples.

Thanks for the clarifications and additional experiments!

About Table 2, my point was more general, if the target domain were from another dataset or with substantial shifts, it is unlikely that the transformed training data would help. Thus, it seems to be with some assumptions that "Describing the Target Domain Is Not Necessary".

I have read the rebuttal and I appreciate the discussions. However, I am inclined to keep my score.