UltraHR-100K: Enhancing UHR Image Synthesis with A Large-Scale High-Quality Dataset

Q1: The vast majority of this paper's contributions lie in the proposal of a large dataset, with limited technical innovations and more engineering-oriented content.

Answer: We respectfully disagree with the reviewer’s assessment that our work is primarily a dataset paper with limited technical contributions. While the construction of the UltraHR dataset is indeed a key component of our work, our paper also proposes a novel and effective frequency-aware post-training strategy tailored for ultra-high-resolution (UHR) text-to-image generation. This algorithmic contribution is clearly distinct from the dataset itself and has been positively acknowledged by other reviewers.

For instance, Reviewer 4Jyg noted:

“Helping models to capture the details in generation also makes sense. A new post-training strategy is proposed to compile the new dataset.”

Similarly, Reviewer WG7N highlighted:

“The proposed post-training method is well-motivated and technically sound, with strong rationale provided for the introduced components.”

and further emphasized that

“The post-training method introduced offers a relatively new approach to improving text-to-image generation in the ultra-high-resolution regime.”

Reviewer 1nz3 also affirmed the value of our method:

“Leveraging the proposed frequency-aware post-training strategy, the model achieves strong performance in UHR image synthesis.”

Although our method is relatively simple, we believe that simplicity does not undermine novelty or impact. On the contrary, simple yet effective approaches often demonstrate greater potential for scalability, reproducibility, and broader adoption within the community.

We therefore believe our work presents both a significant dataset and a technically novel contribution, which together offer substantial value to the main NeurIPS conference.

---

Q2: As a dataset-focused paper, the comparison between the proposed dataset and existing ones is rather superficial.

Answer: We thank the reviewer for the valuable feedback and apologize for the missing citations in Table 2. We will add all relevant references in the revised version to ensure proper attribution and clarity.

Regarding the comparison with existing datasets, we respectfully clarify that our UltraHR-100K dataset presents significant improvements over previous datasets in terms of ** resolution, scale, and diversity** . Beyond basic comparison, we also introduce a novel, task-specific pipeline for curating high-quality ultra-high-resolution (UHR) data, which is a key contribution of our work. This pipeline incorporates multi-dimensional filtering and tailored captioning strategies that are specifically designed to address the unique challenges of UHR image-text data curation.

The motivation and significance of our dataset have been well recognized by other reviewers. For example, Reviewer 4Jyg noted:

"The motivation of proposing UHR100K is straight-forward and intuitive. Instead of using complex training-free methods, which still struggle generating detail-rich images, the authors curate more and better data as a simple solution."

Reviewer WG7N emphasized:

"The methodology for constructing the dataset is carefully considered and well-documented. The filtering and captioning strategies demonstrate a rigorous approach to dataset curation."
"The proposed UHR dataset has the potential to serve as a valuable resource for the community. High-quality, caption-rich ultra-high-resolution image datasets remain scarce, and this contribution could help advance research in text-to-image generation at scale."

Reviewer 1nz3 also acknowledged the strength of our dataset:

"The paper presents a rigorously curated UHR dataset (UltraHR-100K) with significantly larger scale and higher quality than previous datasets such as Aesthetic-4K. The multi-dimensional filtering process is clearly motivated and ensures both diversity and fidelity of the data."

Given the above, we believe that UltraHR-100K will make a substantial contribution to the field of ultra-high-resolution image synthesis and serve as a valuable open resource for the community.

---

Q3: Regarding the DOTS module proposed in Section 4, the paper lacks clarity on the specific strategies for selecting α and β, as well as the parameter sensitivity analysis.

Answer: We thank the reviewer for pointing this out. We apologize for the lack of detailed explanation in the main paper and have included a more comprehensive description in the supplementary materials. Specifically, the DOTS module utilizes a Beta(α, β) distribution to guide the timestep sampling process. The parameters α and β are used to control the sampling bias toward different phases of the denoising trajectory:

• When α < β, the distribution favors later denoising steps (closer to t = 0), which tend to focus on high-frequency detail refinement.
• When α > β, the sampling shifts toward early denoising steps (closer to t = 1), which contributes more to global semantic structure.

In our main experiments, we set α = 2 and β = 4, intentionally biasing the sampling toward later timesteps to better capture fine details, which are critical for ultra-high-resolution image generation.

To further investigate the sensitivity of α and β, we conducted an ablation study, as shown in Table 1. Below is a summary of our findings:

• (α = 3, β =4): Increasing α shifts the distribution rightward, leading to more early-step sampling. This negatively impacts high-frequency learning, as the model receives less supervision in later steps.
• (α =1, β =4): Decreasing α moves the distribution leftward, favoring later-step sampling. This can reduce semantic fidelity, as the model may not learn sufficient global structure.
• (α =2, β =5): Increases the distribution’s peak near the center, causing sampling to concentrate too narrowly, limiting diversity in supervision.
• (α =2, β =3): Lowers the distribution peak, increasing sampling across both early and late steps. While more balanced, this can dilute the benefits of our targeted sampling strategy.

These results support our choice of (α = 2, β = 4) as a balanced and effective configuration that encourages detail learning without sacrificing semantic structure. We will include these discussions and the corresponding table in the revised paper to improve clarity.

---

Q4: Is Google Imagen a comparable approach to the proposed method? If so, it is not discussed in the paper.

Answer: We appreciate the reviewer’s suggestion regarding Google Imagen. While Google Imagen is indeed a powerful closed-source text-to-image generation model, it is not directly comparable to our proposed method for several reasons:
1）Imagen, as described in the original paper [1], focuses on generating 256×256 or 1024×1024 images. However, it does not natively support ultra-high-resolution generation (e.g., ≥ 4K resolution), which is the primary focus of our work. 2）Imagen is not publicly released, making it infeasible to evaluate on our UltraHR benchmarks or to integrate into our framework for a fair comparison. As a result, a quantitative or qualitative comparison is currently inaccessible to the research community. We will clarify this distinction in the updated version and will consider mentioning Imagen explicitly to avoid confusion.

[1] Saharia, Chitwan, et al. "Photorealistic text-to-image diffusion models with deep language understanding." Advances in neural information processing systems 35 (2022): 36479-36494.

---

Q5: It is suggested to use vector graphics for figures, such as Figure 1.

Answer: We thank the reviewer for pointing this out and sincerely apologize for the current rasterized version of Figure 1. We fully agree that vector graphics would improve visual clarity and readability. We will update all relevant figures, including Figure 1, to vector format (e.g., PDF or SVG) in the revised version to ensure better scalability and presentation quality.

---

Table 1. Ablation study for DOTS.

Dear Reviewer Reviewer UXyJ,

As the discussion deadline approaches [Aug 06 '25 (AoE)], we fully understand that you may have other pressing commitments. We are genuinely grateful for the time and effort you have dedicated to reviewing our submission, as well as for your insightful and constructive feedback. The discussion process not only provides an invaluable opportunity to improve the quality of the paper for the authors, but also deepens the reviewers' understanding of the work, contributing to the overall advancement of the academic community. We sincerely hope to continue addressing any unresolved concerns through our responses. Thank you once again for your invaluable time and for your role in helping us refine and strengthen our work!

Firstly, it is a great honor to receive such high praise. We appreciate this helpful suggestion.

For the comparisons on established public datasets

We conduct a quantitative comparison on the publicly available Aesthetic-4K benchmark, specifically the Aesthetic-Eval@4096 subset, as reported in Table 1. This evaluation set contains 195 image-text pairs, where all images have a short side greater than 4096 pixels. Due to the limited number of samples, the reported FID scores are relatively high. Nonetheless, the results clearly demonstrate the superior performance of our method, supporting its robustness and generalizability beyond our proposed benchmark.

---

For the qualitative ablation

We apologize for not including qualitative results in the ablation study. Due to NeurIPS rebuttal guidelines, we are unable to provide visualizations in the response. However, we will include intuitive qualitative comparisons of different ablation variants in the revised version of the paper to better illustrate the contribution of each component.

---

Table 1. Quantitative comparison on Aesthetic-Eval@4096.
We report FID, FID_patch, CLIP, and FG-CLIP metrics. The results demonstrate the superior performance of our model on this public UHD benchmark.

Quality

1. We appreciate the reviewer’s valuable comments on aesthetic filtering. 1) Existing datasets, including LAION, OpenVid-1M, OpenHumanVid, and OmniStyle-1M, commonly employ aesthetic filtering to enhance visual quality by filtering out low-quality samples. Our filtering strategy aligns with these practices, aiming to ensure a cleaner training corpus. Notably, to prevent overly aggressive filtering, we removed only the bottom 50% of images based on aesthetic scores. This approach serves as a lightweight filter, primarily targeting severely degraded or visually corrupted content. 2) We recognize that aesthetic filtering may introduce biases by excluding images with monochrome or low contrast, text-heavy or document-like images, and non-photographic compositions. Although these may seem less “aesthetic,” they often contain important semantic information for text-to-image tasks, so filtering them reduces content diversity. 3) The LAION Aesthetics Predictor is trained on subjective aesthetic scores and may inherently reflect cultural or stylistic biases, favoring certain photographic conventions. It tends to exhibit reduced tolerance for minimalistic, abstract, or utilitarian designs (e.g., charts, educational diagrams, UI mockups). Such biases can inadvertently lead to the exclusion of valuable non-photographic content.

2. The UltraHR-eval4K evaluation set was randomly sampled from the entire dataset under the constraint that both height and width exceed 4096 pixels. This ensures that the evaluation set remains unbiased and reflects the natural data distribution within the full dataset. As shown in Table 1, we provide statistical insights for our UltraHR-eval4K benchmark. We will include these statistics in the revised version of the paper

3. FID measures the distance between two data distributions. Our model achieves strong performance on FID_patch, indicating that the generated image patches exhibit high structural and textural similarity to real high-quality images, reflecting the effectiveness of our approach in preserving fine-grained local details. However, the IS_patch score is relatively low, which warrants further analysis. We attribute this disparity primarily to the characteristics and inherent limitations of the IS. IS evaluates the quality and diversity of generated images via a pre-trained classifier. In the context of UHR image generation, evaluating local patches using IS presents two key challenges: (1) UHR image patches often lack clear semantic category attributes, which reduces the classifier's confidence and thus lowers IS scores; and (2) the classifier used in IS is typically trained on low-resolution data, which introduces a significant distribution gap when applied to high-resolution content—further undermining its reliability. Moreover, since IS is based on classification confidence, it is not well-suited to assess the local detail quality of UHR images. As further evidence, as shown in Table 2, Model A—trained solely on our UHR dataset—already yields a low IS_patch score. This strongly suggests that the discrepancy stems more from the characteristics of the data than from the design of our method.

4. We sincerely apologize for not providing a more detailed discussion of this issue in the paper. In the ablation study, the observed drop in CLIP score when comparing Model D to Model A reflects a limitation of our frequency-aware post-training strategy, which incorporates both DOTS and SWFR. This strategy primarily focuses on enhancing detail learning by allocating more training capacity to the denoising stages in the later sampling steps. However, this reallocation inevitably reduces the emphasis on the earlier denoising steps, which are more critical for modeling global semantics. As a result, the model’s ability to generate semantically rich and globally coherent content may be compromised, thereby affecting the CLIP score.

5. As shown in Table 2, we have already included Model A in the main comparison. Model A refers to the variant trained solely on our proposed dataset using full fine-tuning. The experimental results demonstrate that our curated dataset alone can substantially enhance the generative capability of the baseline model. To ensure a fair comparison, the training setup, hardware environment, and number of training iterations are kept consistent with those of Model D.

---

Clarity

1. We will replace the current table with a pipeline figure to present the data processing steps in a clearer and more objective manner.

2. In the revised paper, we will explicitly state the three key contributions to clarify the scope and focus of the paper.

3. We will correct “Aesthetics” to “Aesthetic” in Line 144 in the revised version.

4. We will add an appropriate citation or footnote for Shannon entropy [1] in Line 141 to ensure clarity and proper attribution for readers unfamiliar with the concept.

5. We agree that Figure 3 would be better placed in Section 3. We will revise the paper accordingly to improve the alignment between figures and corresponding content.

6. We will review Lines 165–170 and either incorporate key points into earlier sections or remove repetition to improve clarity and flow.

7. Model A shares the same architecture as our full model D, both based on SANA. Specifically, Model A refers to the variant trained solely on our proposed dataset using full fine-tuning. To ensure a fair comparison, the training setup, hardware environment, and number of training iterations are kept consistent with those of Model D.

[1] C. E. Shannon, "A mathematical theory of communication," in The Bell System Technical Journal.

---

Significance

1. We conduct a quantitative comparison on the publicly available Aesthetic-Eval@4096 benchmark, as reported in Table 3. This evaluation set contains 195 image-text pairs, where all images have a short side greater than 4096 pixels. The results clearly demonstrate the superior performance of our method, supporting its robustness and generalizability beyond our proposed benchmark.

2. We acknowledge that advances in photorealistic high-resolution image generation raise important ethical and societal concerns, including risks related to misinformation, privacy, and misuse of synthetic media. We emphasize the need for increased awareness of these issues and encourage the community to collaboratively develop ethical guidelines and mitigation strategies. We will include a brief discussion of these ethical challenges in the revised manuscript.

3. The main limitation of our dataset currently lies in the relatively small amount of portrait data, which results in limited improvements over prior methods in UHR portrait generation, as also illustrated in Figure 8 of the paper. In future work, we plan to collect more high-quality UHR portrait images to enhance our model’s performance specifically in UHR portrait synthesis.

---

Originality

1. We have added a comparison experiment with LoRA, as shown in Table 4, to further highlight the effectiveness of our proposed frequency-aware post-training strategy. Compared to existing post-training approach, our method demonstrates particularly notable improvements in detail enhancement, as reflected by the FID_patch metric.

---

Questions

1. We confirm that our dataset will be released under the Creative Commons Attribution-NonCommercial (CC BY-NC) license, which permits redistribution and use for non-commercial purposes, such as academic research. This license ensures a permissive usage framework while maintaining responsible use boundaries. We are fully committed to open science—upon acceptance of the paper, we will release all data, code, and trained models to the community. The primary goal of our dataset is to support and enrich open research in the field of ultra-high-resolution text-to-image generation.

2. For aesthetic, please refer to our response to Comment 1 in the Quality section.

3. For limitations, please refer to our response to Comment 3 in the Significance section. Furthermore, we have discussed the broader social implications and ethical considerations in our response to Comment 2 in the Significance section.

4. Please refer to Comments 3 and 4 in the Quality section for a detailed discussion and clarification regarding Model A.

---

Table 1. (a) Distribution statistics of image resolutions in the UltraHR-eval4K benchmark.

(b) Caption length distribution in our UltraHR-eval4K benchmark.

---

Table 2. Quantitative comparison on our UltraHR-eval4K (4096 × 4096) benchmark.

---

Table 3. Quantitative comparison on Aesthetic-Eval@4096.

---

Table 4. Comparison experiment with LoRA.

•	Regarding Quality Point 1:

Dear Reviewer Reviewer WG7N,

As the discussion deadline approaches [Aug 06 '25 (AoE)], we fully understand that you may have other pressing commitments. We are genuinely grateful for the time and effort you have dedicated to reviewing our submission, as well as for your insightful and constructive feedback. The discussion process not only provides an invaluable opportunity to improve the quality of the paper for the authors, but also deepens the reviewers' understanding of the work, contributing to the overall advancement of the academic community. We sincerely hope to continue addressing any unresolved concerns through our responses. Thank you once again for your invaluable time and for your role in helping us refine and strengthen our work!

We appreciate this valuable suggestion. However, due to this year’s NeurIPS rebuttal guidelines, we are unable to include visual results in the response. We will include the requested visualizations in the revised version of the paper, including qualitative examples for minimalism prompts and ablation study results.

---

For the Idea / Motivation

We thank the reviewer for raising the concern about potential bias toward rich visual content, which may not align with aesthetics such as minimalism. To investigate this, we conducted preliminary qualitative experiments using minimalist prompts before and after post-training. The results suggest that our method, while enhancing detail, does not significantly degrade the model’s ability to generate stylistically minimalist images. This robustness may stem from the strong generalization ability of the pretrained model. Our primary motivation is to improve model performance in Ultra-High-Resolution (UHR) scenarios, especially under complex content and fine-grained detail—areas where existing models often struggle. For simple or minimalist generation tasks, pretrained models such as FLUX or SD3 are often already sufficient without further UHR fine-tuning. We plan to include further discussion and illustrative minimalist examples in the revised version.

---

For the Method

To assess whether the long captions generated by Gemini 2.0 align with human perception, we conducted a user study, as shown in Table 1. Ten volunteers each evaluated 50 caption pairs (our original long captions vs. Gemini-generated short captions) based on semantic richness (the extent to which the caption captures semantic content of the image) and semantic accuracy (how accurately the caption reflects the image content), using a 1–5 scale (1 = very poor, 2 = poor, 3 = fair, 4 = good, 5 = excellent). We report the average scores across all participants. Additionally, we collected overall preference judgments between the two captions. Results show that the long captions are consistently rated higher in both semantic richness and accuracy, and are more frequently preferred, demonstrating their stronger semantic expressiveness. Additionally, in Table 2, we evaluated models using only the first sentence (short version) of the long captions. Results confirm that longer, richer captions lead to higher-quality image generation. While other LLMs like GPT-4o can also generate captions, we focused on Gemini due to current resource constraints. We plan to explore more LLMs in future work.

---

Experiments

1. Unrealistic Styles

We acknowledge that in some cases, our method produces over-smoothed textures or unnatural color tones. For skin smoothness, We believe this may be due to the scarcity of portrait data in our dataset, resulting in suboptimal portrait generation performance. This is a current limitation we plan to address by collecting more high-quality UHR portrait data. Unnatural color tones may also stem from base model limitations (SANA) and the difficulty of balancing detail enhancement with natural appearance. We will refine our post-training strategy to improve visual realism.

2. User Study

As shown in Table 3, we conducted a user study with 5 volunteers evaluating 50 randomly selected cases. Images were rated on overall quality, detail quality, text-image alignment and preference. The results demonstrate the superiority of our method across all aspects.

3. Ablation Study

We will include qualitative comparisons of ablation variants in the revised version.

4. Generalization to Other Models

Due to limited computational resources, we currently focus on SANA. We plan to extend our training strategy and dataset to other models such as FLUX and SD3 in future work.

5. Image Compression

All images were compressed using the same method due to Overleaf compilation limits. We apologize for not including original-resolution images and will release all data and outputs publicly after the paper accepted.

---

Clarifications & Minor Issues

• We apologize for the inconvenience caused by the grammatical and formatting issues. We will correct the grammatical error (“a open-source” → “an open-source”) and improve the color distinguishability in Figure 5.
• The term "4K" in UltraHR-eval4K refers to image resolution, not the number of samples.
• The frequency loss weight $\lambda_{\text{freq}}$ is set to 1 in our experiments.

---

Limitations & Future Work

As mentioned, the lack of high-quality portrait data limits performance in portrait generation scenarios. Additionally, while our frequency-aware post-training strategy enhances detail synthesis, it slightly compromises text-image alignment, as evidenced by the CLIP score: Model A outperforms Model D in Table 4 of the paper. In future work, we will:

• Collect more high-quality UHR portrait data.
• Explore more effective training strategies to address this trade-off and improve robustness.

---

Table 1. Caption User Study. Long_cap refers to our original long captions containing both global and fine-grained descriptions, while Short_cap refers to newly generated short captions that include only global descriptions.

---

Table 2. Quantitative Comparison on UltraHR-eval4K Using Short Captions. SANA-s and Ours-s denote results evaluated on the UltraHR-eval4K benchmark using short captions (the first sentence of the original long captions), while SANA and Ours indicate evaluation using the full original long captions.

---

Table 3. User study results conducted on our UltraHR-eval4K benchmark, evaluating overall quality, detail fidelity, text-image alignment and user preference across different methods.