TL;DR: Concise Summary of Our Revisions

Our work demonstrates that competitive text-to-image models can be trained on a small, accessible dataset like ImageNet, challenging the need for massive commercial datasets.

Key Contributions & New Experiments:

• Data-Efficiency: We introduce a new metric, $\text{score}_{\text{eff}} = \frac{\text{score}}{\log(\text{dataset size})}$, to show our models achieve superior performance relative to their training data size, outperforming massive baselines.
• Generalizability: We successfully applied our approach to other small datasets, finding that ImageNet surprisingly yields better results than LAION-POP, confirming its value as a standardized and accessible training source.
• Multi-object Capabilities: We will add more examples of complex, multi-object scenarios to the final paper to better showcase the model's capabilities.
• Novel Concepts: New experiments show our model can generate high-quality images for concepts not in ImageNet (e.g., "iPad") by leveraging the text encoder and reinterpreting existing visual patterns.
• CutMix Limitations: We acknowledge potential caption artifacts from our CutMix strategy and address this with a "Crop" alternative and fine-tuning. We also note that CutMix improves spatial relationship learning.

In short, our results reinforce that accessible, academic-scale datasets are a viable path for training powerful and reproducible text-to-image models.

Addressing weaknesses:

More data seems to reach better absolute performance

We agree that large-scale commercial models such as SD3 and Janus currently achieve the highest absolute performance, and we do not claim to outperform them. Our goal is not to replace such models, but rather to challenge the prevailing assumption that only massive datasets and models can yield strong results.

Our work demonstrates that with reproducible, resource-efficient training on a small and well-known dataset like ImageNet, it is still possible to achieve competitive performance.

Regarding data efficiency:

To make this comparison more explicit, we introduce a data-efficiency-aware metric, defined as:

$\text{score}_{\text{eff}} = \frac{\text{score}}{\log(\text{dataset size})}$

This metric accounts for both performance and dataset size. As shown in Table D.1, our models outperform many large-scale baselines on this adjusted score, reinforcing our central claim: smaller datasets and models can yield excellent performance per unit of data.

Table D.1: Comparison with SOTA models on DPG Bench adjusted w.r.t $\text{score}_\text{eff}$

What about using another data-source than Imagenet with the same amount of data?

To evaluate the generalizability and robustness of our approach, we trained models on other small-scale datasets using the same pipeline (captioning, augmentation, architecture, and compute budget). Specifically, we experimented with:

• ~$1$M samples from DataComp

• $470$k images from LAION-POP (as suggested by Reviewer tjpV)

• ImageNet combined with $120$k COCO images to reduce object-centric bias

The GenEval results are shown in Tables D.2 (copied from Table C.1 in the answer to Reviewer he9B) and D.3:

Table D.2 (copied from C.1): Comparison on GenEval Benchmark between different training datasets with the DiT-I architecture.

Table D.3: Comparison on GenEval Benchmark between ImageNet-1.2M and Datacomp-1M, using CAD-I architecture

Key Takeaways:

• ImageNet+COCO improves compositionality (e.g., Two Objects and Counting), suggesting that reducing ImageNet’s object-centric bias helps.

• ImageNet outperforms LAION-POP when used for training from scratch, likely due to its more balanced and concept-uniform distribution.

• DataComp-1M lags behind significantly despite similar scale, possibly due to noise, and e-commerce bias.

A promising direction for future work is to explore combinations of high-quality academic datasets (e.g., ImageNet, COCO, CC12M) to build a robust yet reproducible training set, extending what we began with ImageNet+COCO.

Move details about CutMix from Appendix E-1 to the main paper.

Thank you for the suggestion, we will do in the camera ready version.

Include results showcasing more complex multi-object scenarios to better demonstrate the model’s text-to-image (T2I) generation capabilities.

Thank you for the suggestion, we will add a full page of such examples in the camera ready version. Note that we already have such images in the paper, for example the Kangaroo with sun-glasses in front of Sydney's Opera House in the teaser depict a complex interaction with 3 concepts that are probably not seen together in ImageNet. Same for the Corgy with the pizza or some of the DPGBench prompts in Appendix A.2.

Adressing questions:

Can we apply this to other datasets?

See the answer to the corresponding weakness.

Are there any limitations to the CutMix strategy and how are they addressed?

The main limitation in the CutMix strategy is that some of the captions are easily distinguishable from non-CutMix captions (e.g., “The left part of the image [...]. The right part of the image [...]”). In turns, the model learns these characteristics and has a chance to produce artifacts when prompted with a CutMix caption (which is not the case for human made captions). This is the reason for also proposing the Crop strategy which does not suffer from such bias. In any case, fine-tuning without CutMix augmentation gets rid of any artifacts. This is what we do for higher resolution models (512, 1k).

How does the model handle concepts that are rare or absent in ImageNet?

This is a question we also found fascinating and that **we have explored in depth since the submission deadline••. We wish we could show images as in the previous NeurIPS instances, but alas. To test this, we tried prompts containing objects that did not exist at the time of ImageNet’s creation such as an iPad or virtual reality goggles like the HoloLens. The images we get are high quality (about the same as for regular prompts), and the new concept is hallucinated from patterns that are in the training set. For example, the iPad is made from a shortened CRT monitor and the HoloLens are made from a scuba-diving mask with plastic inserts on the side. We believe this is thanks to the text encoder that knows about these concepts and produces features that the diffusion model has to interpret as visual structures. It could be an indication that the closest visual structure to a HoloLens is a scuba-diving mask. We will definitely add examples of such cases in the final version.

How does CutMix contribute to learning spatial relationships? Could the approach be extended to involve more than two objects?

CutMix does improve a bit the two-objects category on GenEval (see Table 2), which could indicate better spatial relationship learning. Merging more than 2 images is definitely possible, but it would then become tricky to design a strategy that produces images without harmful artifacts. This is an interesting research direction.

TL;DR

We thank the reviewer for their valuable feedback. We have addressed the weaknesses and questions raised by the reviewer, and we summarize our key points and new experimental findings below:

• Methodological Novelty: We acknowledge that our core techniques (long captions and image augmentations) are not novel. Our core contribution is demonstrating that a reproducible, resource-efficient training on a small, open dataset like ImageNet can achieve state-of-the-art performance, matching or exceeding models trained on orders of magnitude more data. Our work advocates for a standardized, accessible, and reproducible training paradigm for text-to-image generation.

• New Experiments:

  • Compositional Capabilities: We trained a new model on a combined ImageNet+COCO dataset to test for object-centric bias. The results show a slight improvement on multi-object compositional tasks like counting and two-object generation, suggesting that while our augmentation strategy is effective, having a more diverse dataset still offers benefits.
  • Scaling to Higher Resolutions: We fine-tuned our model to $1024 \times 1024$ resolution on both a high-resolution ImageNet subset and LAION-POP. Both models successfully generate high-quality images. The LAION-POP model showed a noticeable improvement in aesthetic scores, suggesting that ImageNet may have limitations for scaling to higher resolutions due to its lack of diverse high-res content.
  • Improving Multi-Object Composition: We believe that further improving multi-object composition is a crucial open question. While our current approach helps, a dedicated research direction could focus on developing cost-effective data augmentation strategies that combine objects and generate appropriate captions without relying on large, noisy datasets. This is a very interesting direction for future work.

Adressing Weaknesses:

Novel idea but lack of methodogical novelty

Indeed, we fully acknowledge that using long captions and image augmentations is not novel, and we do not claim these techniques as our contribution.

The central goal of our paper is to challenge the “bigger is better” paradigm by showing that reproducible, resource-efficient training on smaller datasets, specifically ImageNet, can still yield strong performance; comparable to state-of-the-art models trained on orders of magnitude more data.

Our main contribution is to demonstrate that it is possible to:

• Train models from scratch on a small, open, and well-known dataset (ImageNet);

• Avoid reliance on massive web-scraped datasets;

• Match or exceed the performance of popular models like SDXL on key compositional benchmarks;

• All while using orders of magnitude less data and compute.

Our intent is to advocate for a standardized, accessible, and reproducible training setup for text-to-image generation. We believe that promoting reproducibility, accessibility, and sustainability in generative model research (in particular text-to-image generation) is important and timely message for the community.

Can we improve on Imagenet object centric bias?

We train two additional models: one on ImageNet$+$COCO and one on LAION-POP (as suggested by Reviewer tjpV) to assess whether having a dataset without the object centric bias helps. The GenEval results are:

Table C.1: Comparison on GenEval Benchmark between different training datasets with the DiT-I architecture.

As we can see, adding COCO to the training mix does not change the overall results very much, which indicates that the augmentation strategies are enough to get general capabilities. The model trained with COCO performs better on the Two Objects task (0.59 vs. 0.56) and Counting (0.42 vs. 0.38), suggesting that reducing the object-centric bias of ImageNet helps improve multi-object compositional reasoning.

LAION-POP is either too small or not diverse enough to lead to satifying results. More advanced capabilities (like high quality faces or aesthetics images) could be acquired after pretraining on ImageNet by fine-tuning on a dedicated dataset (see Table A.3 in the answer to Reveiwer tjpV)

Can we scale further than 512x512 resolution?

To explore higher resolutions, we fine-tuned two models at $1024\times1024$ resolution:

• One on a high-resolution subset of ImageNet, and

• One on LAION-POP, as suggested by Reviewer tjpV.

Both models successfully generate high-quality $1024\times1024$ images, with distinct biases: the ImageNet model tends to produce photographic outputs, while the LAION-POP model exhibits a stronger aesthetic bias.

As shown in Table C.2 (copied from Table A.3 in the answer to Reviewer tjpV), the ImageNet-1024 model achieves aesthetic scores comparable to its $512\times 512$ counterpart, while the LAION-POP-1024 model shows a noticeable improvement in aesthetic metrics. This suggests a potential limitation of ImageNet for high-resolution generation: its relative lack of high-res content may hinder performance scaling at $1024\times 1024$ and beyond.

Table C.2 (copied from A.3): Comparison on Aesthetic metrics between ImageNet and LAION-POP for finetuning at $1024\times 1024$ . We start our finetuning from the previously reported DiT-I Crop $512\times 512$ resolution checkpoint. PickScore and Aesthetic Scores are based on images generated using Partiprompts whereas HPSv2.1 and ImageReward are calculated using the respective benchmark prompts.

We see this as an interesting direction for future work, particularly in combining data-efficient training with selective use of high-resolution content.

Addressing questions:

How to further improve multi-object composition without increasing data?

This is a very interesting open question. We show that mixing images from the training set can improve two object composition (see GenEval score in Table 2). This strategy could be explored by mixing more objects. However, this requires generating captions for the specific mix of objects, and it also requires designing a strategy that avoids creating bad images (cutting objects, floating objects, nonsensical mixes, etc). Solving multi-object composition augmentations without a high caption generation cost is a very interesting future research direction. This would be part of an entire field of research that could be built on top of our proposition of imagenet-based pretraining of T2I models, with fair, reproducible and accessible research.

TL;DR

We appreciate the reviewers' comments and have significantly updated our submission.

Key Updates & Contributions:

• Public Release: All code and data will be made public upon acceptance.
• Simplicity & Reproducibility: Our core contribution is a reproducible and accessible framework for T2I research. We demonstrate that impressive results are achievable with simple setups (ImageNet, straightforward methods), challenging the "bigger is better" paradigm and lowering the entry barrier.
• New SOTA Baselines: We've added comprehensive comparisons against PixArt-$\Sigma$, Janus, FLUX, SANA, and SD3 on GenEval and DPGBench (Tables B.1, B.2 from reviewer tjpV).
  • Our models outperform PixArt-$\Sigma$ on GenEval at $512\times 512$ resolution, particularly in compositional tasks (Two Objects, Position, Color Attribution).
  • Though lagging behind large-scale commercial models (SANA, SD3, FLUX), our models achieve competitive results with $1000\times$ fewer images and $10\times$ fewer parameters than SDXL.
• CutMix Effectiveness: While not always improving FID, CutMix significantly enhances compositional understanding (e.g., Two Objects, Position scores in Table 2) and mitigates overfitting, leading to improved training stability (Figure 3).
• Robustness Across Paradigms: We demonstrate the generalizability of our training setup by evaluating Flow Matching models (Table B.3). They achieve competitive results comparable to PixArt-$\Sigma$ and SD v2.1, proving our approach is robust across different generative modeling paradigms.
• New Crop Augmentation Experiment: We added experiments with the Crop augmentation strategy on CAD-I to complete our ablation studies (Table B.4).
• Overfitting Mitigation: Our augmentations (especially CutMix) effectively delay overfitting on ImageNet, enabling strong performance with significantly less data and compute than state-of-the-art models.

Addressing Reviewer Feedback

We appreciate your thorough review and valuable comments. We've addressed your points below.

1. Typos and Style Discrepancies

We apologize for the typos and inconsistencies in style. We will meticulously proofread and correct these in the camera-ready version of the paper.

2. Broken Code and Data Links

We understand your frustration regarding the non-functional links. These links were placeholders, included to estimate space requirements and maintain anonymity during the review process. All code and data are prepared and will be made publicly available upon the paper's acceptance.

3. Simplicity of the Solution

We appreciate your feedback on the simplicity of our solution. Our primary contribution lies in offering the community a reproducible and easily accessible framework for Text-to-Image (T2I) research. The simplicity of our approach is, in fact, a deliberate design choice rather than a limitation.

We aim to demonstrate that even with a straightforward dataset like ImageNet and a simple setup, it's possible to achieve impressive results that might otherwise be considered unfeasible. While more complex methods could potentially yield incrementally better performance, we believe our current setup effectively highlights the value and potential of this research direction. Our work underscores that significant insights can be gained without resorting to overly intricate solutions, thereby lowering the barrier to entry for other researchers.

4. More Modern baselines

Thank you for the helpful suggestion.

We have updated Tables 5 (GenEval) and 6 (DPGBench) in the main paper to include comparisons with more recent models, including PixArt-$\Sigma$, Janus, FLUX, SANA, and SD3. Please refer to Tables B.1 and B.2 from reviewer tjpV for a preview.

At $512\times 512$ resolution, our models outperform PixArt-$\Sigma$ on the GenEval benchmark, particularly in the Two Objects, Position, and Color Attribution sub-tasks. However, we acknowledge that our models still fall short of the strongest commercial baselines such as SANA and SD3.

On the DPGBench benchmark, our best model again achieves comparable performances to PixArt-$\Sigma$ at $512\times 512$ resolution and Janus, but lag behind SANA, SD3 and FLUX, which use larger-scale training and commercial infrastructure.

5. The effectiveness of the proposed methods is not robust across different models. For example, the CutMix is not effective on CAD-I as in Table 3.

We acknowledge that CutMix does not consistently improve FID metrics, particularly for the CAD-I model in Table 3. However, we believe its contribution remains essential for two key reasons:

1. Improved Compositionality:

While FID remains similar, CutMix significantly enhances compositional understanding. For example, in Table 2, the Two Objects score for CAD-I improves from 0.60 to 0.68, and the Position score improves from 0.26 to 0.35 with CutMix. These gains indicate stronger multi-object reasoning and spatial understanding.

2. Overfitting Mitigation:

As shown in Figure 3, CutMix helps delay overfitting during training. The FID curve stays lower for longer, and GenEval scores degrade more slowly, demonstrating improved training stability and generalization under limited data.

6. More fundamental methods need to be evaluated, such as SiT and traditional U-Net.

Thank you for the suggestion.

In response, following SiT, we trained and evaluated Flow Matching models under the exact same conditions as our diffusion-based models. These models use the CAD-I architecture, and were trained with identical datasets, compute budgets, and training procedures.

As shown in Table B.3, while Flow Matching models slightly lag behind diffusion-based models in overall performance due to undertraining, they still achieve competitive results, comparable to PixArt-$\Sigma$ (0.52) and Stable Diffusion v2.1 (0.50) on the GenEval benchmark. This demonstrates that our proposed training setup, based on ImageNet with long captions and augmentations, is robust and generalizable across different generative modeling paradigms. We believe with more training time and better hyperparameters, the flow matching models could reach or even surpass the diffusion models.

Table B.3: Comparison between Flow-Matching and Diffusion models, using the CAD-I architecture at $256\times 256$ resolution.

7. Why are there no experiments for Crop on CAD-I? Does it achieve unsatisfactory performance or fail to converge?

We originally just did not train with the crop strategy on CAD-I because the CutMix strategy seemed to be sufficient to get good performances.

For the rebuttal, we have trained a CAD-I model with the Crop augmentation strategy to complete the ablation. The training followed the same setup as our other CAD-I experiments.

Table B.4 contains the results that are in line with the ones obtained for DiT-I.

Table B.4: Comparison between Image Augmentation strategies with the CAD-I architecture.

8. Training only on ImageNet may result in an overfitting problem during the longer training process, which could be found in Figure 3. Thus, in many T2V tasks, a larger training dataset is required to provide more diverse visual patterns and image-text alignment datapoints.

We agree that overfitting is a common concern when training on small datasets like ImageNet. However, as shown in Figure 3, our proposed strategy, particularly the use of image augmentations such as CutMix, significantly mitigates overfitting over the course of training. The FID remains stable for longer, and the GenEval score degrades more slowly.

Our approach does not require prolonged training to reach competitive performance. As demonstrated in Tables B.1 and B.2 from reviewer tjpV, our model already matches SDXL on GenEval while being trained with $1000\times$ fewer images, $10\times$ fewer parameters and a fraction of the compute budget.

This supports the central goal of our paper: to challenge the "bigger is better" paradigm by showing that with thoughtful data augmentation and captioning strategies, reproducible, resource-efficient training on smaller datasets can still yield strong performance. Exploring how to extend training duration or scale model capacity further while maintaining data efficiency is indeed an interesting direction, and we leave this for future work.

TL;DR: ImageNet Outperforms LAION-POP & High-Res Capabilities Demonstrated

New experiments confirm:

• ImageNet (trained from scratch): Consistently outperforms LAION-POP on compositional understanding (GenEval) and even aesthetic metrics. ImageNet offers better generalization and reproducibility.
• LAION-POP: While fine-tuning on it improves aesthetics, training from scratch yields significantly worse compositionality. Issues with dead URLs and inappropriate content also exist.
• High-Resolution: Successfully fine-tuned our model to $1024\times 1024$ on high-res ImageNet, achieving comparable aesthetic quality.
• Competitiveness: Our ImageNet-trained models (smaller, less data) show competitive performance on benchmarks (GenEval, DPGBench) against recent SOTA models like PixArt-$\Sigma$.

Conclusion: ImageNet remains a strong, reproducible dataset for competitive diffusion model training, especially for compositional understanding.

Addressing Weaknesses:

Lack of Comparison with High-quality (high aesthetic & high resolution & enriched caption) Dataset - LAION-POP [1]:

We appreciate the reviewer’s detailed feedback and the reference to LAION-POP and the KOALA study.

We conducted additional experiments:

• We trained a model from scratch on LAION-POP (only ~$470$k images because of url decay over time) at $256\times 256$ resolution in Tables A.1 and A.2
• We fine-tuned our ImageNet trained model at $1024\times 1024$ resolution using $1024$ resolution LAION-POP samples (~$375$k images) in Table A.3
• We fine-tuned a $1024\times 1024$ resolution model on high-resolution ImageNet samples (only ~$26$k images) in Table A.3

Table A.1: Comparison on GenEval Benchmark between ImageNet and LAION-POP for our DiT-I model trained from scratch at $256\times 256$ resolution.

Table A.2: Comparison on Aesthetic metrics between ImageNet and LAION-POP for our DiT-I model trained from scratch at $256\times 256$ resolution. Scores are calculated using PartiPrompts.

Table A.3: Comparison on Aesthetic metrics between ImageNet and LAION-POP for finetuning at $1024\times 1024$, starting from the DiT-I Crop $512\times 512$ resolution checkpoint. PickScore and Aesthetic Scores are using Partiprompts whereas HPSv2.1 and ImageReward are calculated using the respective benchmark prompts.

Results

Fine-tuning on LAION-POP improves aesthetic metrics, aligning with the KOALA findings. However, our main motivation is to train a model from scratch. As such when training from scratch on LAION-POP, we observe significantly worse compositionality. In contrast, ImageNet-based models exhibited better generalization and compositional understanding under the same training conditions.

ImageNet vs LAION-POP:

We intentionally chose ImageNet for its accessibility, reproducibility, and stable long-term hosting. While LAION-POP offers high-quality images, it suffers from reproducibility issues. Over 20% of the image urls are already unavailable on HuggingFace. We could only retrieve ~$470$k out of $600$k images. Additionally, LAION-POP contains inappropriate content (e.g., NSFW material appears within the first few hundred samples from HuggingFace).

As shown in Table A.1, ImageNet-trained models outperformed those trained on LAION-POP in both aesthetic metrics quality and GenEval.

Comparison with KOALA:

We believe there may have been a misunderstanding: KOALA uses LAION-POP in a fine-tuning/distillation setup starting from an SDXL checkpoint that has already been trained on over 1B images. Our work is different: we train models from scratch using only ~1M images. Models trained with KOALA benefit from all the images seen during the SDXL pretraining, which is crucial as shown in Table A.1.

Limited Evaluation of Aesthetic Quality:

We have added the ImageReward and HPSv2 as metrics in the comparison tables.

Limited Practical Quality of Generated Images and qualitative comparison with SoTA open-source models:

We acknowledge that for certain applications, particularly in creative industries or commercial design; high-resolution, fine-grained visual quality is essential.

To address this, we have fine-tuned our model at $1024\times 1024$ resolution using high-resolution ImageNet samples. As shown in Table A.3, the finetuned model achieves comparable performance in terms of aesthetics. We will include updated qualitative comparisons with recent open-source models such as KOALA, PixArt-$\Sigma$, and SANA-series in the final version of the paper to better highlight relative strengths and trade-offs (see Tables B.1 and B.2 below).

We would also like to emphasize that our primary goal is not to replace large-scale commercial models in creative generation tasks, but to demonstrate that smaller, more reproducible datasets like ImageNet can still produce competitive models. There are domains where prompt fidelity and compositional alignment are the hard constraint and aesthetic quality is irrelevant, examples include medical imaging, rare event generation, or simulation (e.g., autonomous driving). In these contexts, our method matches or outperforms much larger models, despite using $10\times$ fewer parameters and more than $1000\times$ less data.

Addressing questions:

What resolution is used for SDXL?

Thank you for pointing this out.

The quantitative results reported in Tables 5 (GenEval Benchmark) and 6 (DPG Benchmark) are taken directly from the original papers (e.g., SDXL, SD1.5, PixArt). However, we acknowledge that in the qualitative comparison in Figure 5 SDXL images are generated at $512\times 512$ resolution.

Using a $1024\times 1024$ model fine-tuned using only the high-resolution subset of ImageNet, shown in Table A.3, we obtain comparable visual quality compared to its $512\times 512$ counterpart.

In the camera ready version, we will include updated qualitative comparisons at $1024\times 1024$ resolution between our model and SDXL.

Could the Authors provide more up-to-date baselines

Thank you for the helpful suggestion.

We have updated Tables 5 (GenEval) and 6 (DPGBench) to include comparisons with more recent models that leverage enriched or multi-level captions, including PixArt-$\Sigma$, Janus, FLUX, SANA, and SD3. (See Table B.1 and B.2 below).

At $512\times 512$ resolution, our models outperform PixArt-$\Sigma$ on GenEval, particularly in the Two Objects, Position, and Color Attribution sub-tasks. However, we acknowledge that our models still fall short of the strongest commercial baselines such as SANA and SD3.

On DPGBench, our best model again achieves comparable performances to PixArt-$\Sigma$ at $512\times 512$ resolution and Janus, but lag behind SANA, SD3 and FLUX, which use larger-scale training and commercial infrastructure.

Table B.1: Comparison with SOTA models on GenEval Benchmark

Table B.2: Comparison with SOTA models on DPG Benchmark

Thank you for the clarifications provided in the rebuttal. I appreciate the authors' effort to address the raised concerns. That said, I would like to elaborate on a few remaining points regarding the evaluation and interpretation of your results.

Concerns on the KOALA Comparison in Table A.2b:

The KOALA model was trained at a resolution of 1024×1024 in Table A.4, When it is evaluated at significantly lower resolutions (e.g., 256×256 or 512×512), severe artifacts and distortions may occur. As a result, comparisons made at such low resolutions could unfairly disadvantage KOALA. Since your proposed method is trained only at 256² or 512² resolutions, it naturally performs better under these settings. Therefore, the comparison in Table A.2b does not appear to be a fair evaluation of model quality across architectures.

Limited Upper Bound of Performance for ImageNet-trained Models:

Although the proposed model was fine-tuned(w/o inference scaling) at 1024×1024 resolution, its aesthetic scores still show a noticeable performance gap compared to SDXL in high-resolution evaluations. This suggests that models trained from scratch on ImageNet may face fundamental limitations in terms of scalability and upper-bound quality. As such, I remain uncertain about the broader applicability of the proposed approach in scenarios that demand high visual fidelity.

Fair Compositionality Evaluation at High Resolution Against LAION-POP:

In the rebuttal, you mentioned that “ImageNet-trained models outperformed those trained on LAION-POP in both aesthetic metrics and GenEval” (Table A.1). However, to more convincingly argue for ImageNet’s utility, it would be important to show competitive performance against LAION-POP under the high-resolution fine-tuning setting (e.g., 1024²) in compositionality benchmarks such as Table A.2. Without this, it is difficult to attribute the observed improvements solely to the ImageNet dataset. Moreover, since ImageNet lacks native text captions, your strategy to augment captions using an external captioner is not novel—it has already been explored in prior works such as Pixart-Sigma and KOALA. Therefore, it’s hard to conclude that ImageNet alone contributes meaningfully to improving compositionality. Given that captions can be synthetically augmented for any dataset, I believe that the quality of the images themselves becomes the most critical factor in T2I research. From this standpoint, the limitations of ImageNet remain quite evident.

From my perspective, the statement that “models designed for artistic content creation are unable to generate realistic pictures of dangerous scenarios for autonomous driving” is not well-supported. The LAION-POP and LAION-Aesthetic datasets are not limited to images optimized purely for visual beauty—they also contain a substantial number of highly photo-realistic images, arguably far more than ImageNet. Moreover, the aesthetic scores I emphasize (e.g., HPSv2, ImageReward) are not mere measures of visual beauty; rather, they reflect human preferences for overall visual quality, which includes both aesthetic appeal and realism. Therefore, it is not accurate to claim that models trained on LAION-based datasets are “unable to generate realistic pictures,” especially when compared to models trained from scratch solely on ImageNet. In fact, as evidenced in Figure 1, the realism of the results does not appear clearly superior.

Second, as I have repeatedly noted, the improvement in compositionality observed with ImageNet is not a result of the raw ImageNet dataset itself, but rather of text augmentation through an image captioner. It is this augmented text that drives the improvement in compositionality metrics. This directly challenges the claim that the proposed method outperforms SDXL in terms of compositionality. SDXL is an older model from 2023 that was not trained with enriched text captions, whereas the proposed approach benefits from augmented text generated by an image captioner. In this context, the observed compositionality improvement cannot be attributed to the underlying ImageNet dataset itself, but rather to the additional caption augmentation. Furthermore, as shown in Table B.2, the compositionality metrics of the proposed method are lower than those of PixArt-Sigma, a more recent text-augmented model. This suggests that the claimed advantage over SDXL does not necessarily extend to stronger, more comparable baselines.

Considering these points, I am not fully convinced that ImageNet with augmented text captions will have a substantial impact on future text-to-image research, particularly when high-quality, large-scale datasets such as LAION already offer higher resolution, greater sample diversity, and enriched captions. While the proposed approach demonstrates interesting results, its broader impact may be more limited than suggested.