acceptodds
Under review as a conference paper at ICLR 2027

Fine-T2I: An Open, Large-Scale, and Diverse Dataset for High-Quality T2I Fine-Tuning

Abstract

High-quality and open datasets remain a major bottleneck for text-to-image (T2I) fine-tuning. Despite rapid progress in model architectures and training pipelines, most publicly available fine-tuning datasets suffer from low resolution, poor text–image alignment, or limited diversity, resulting in a clear data gap for open text-to-image research. In this work, we present Fine-T2I, a large-scale, high-quality, and fully open dataset for open T2I fine-tuning. Fine-T2I spans 10 task combinations, 32 prompt categories, 11 visual styles, and 5 prompt templates, and combines synthetic images generated by strong models with carefully curated real images from professional photographers. Unlike previous datasets that simply present a set of samples, Fine-T2I is built with: 1) systematical dataset processing pipeline, 2) principled distribution design, 3) multi-granularity prompt design, 4) dynamic aspect-ratio & resolution strategy, and 5) advanced reasoning-VLM-as-a-judge protocol. Furthermore, all samples are rigorously filtered for text–image alignment, visual fidelity, and prompt quality, with over 95% of initial candidates removed. The final dataset contains over 6 million text–image pairs, around 2 TB on disk, approaching the scale of pretraining datasets while maintaining fine-tuning-level quality. Across the open diffusion and autoregressive models we study, fine-tuning on Fine-T2I improves generation quality and instruction adherence, as validated by human evaluations and visual comparisons, with gains that depend on the base model's starting capability. Our analysis further shows that data aligned with a benchmark's distribution can inflate automatic scores without improving human-judged quality, underscoring the value of preference-oriented curation. We release Fine-T2I under an open license to help close the data gap in the open community.

Primary area:
datasets and benchmarks
Keywords:
FineT2i, Image generation, Dataset

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_aWUfOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Fine-T2I is a large-scale dataset for text-to-image fine-tuning, combining 6,145,693 synthetic samples with 168,424 curated real images. Its construction pipeline integrates controlled prompt generation, semantic deduplication, prompt enhancement, multiple aspect ratios and resolutions, aesthetic filtering, and reasoning-VLM-based verification. The authors evaluate the dataset by fine-tuning LlamaGen, SD-XL, and DeCo, reporting improvements in human preference, particularly in visual quality, alongside mixed results on GenEval. The main contribution is a substantial data resource and an integrated curation pipeline intended to support open text-to-image research.

Critical strengths
  1. A substantial data contribution addressing a practical research need. Fine-T2I brings together millions of synthetic samples and curated real photographs, with coverage across content categories, visual styles, prompt granularities, and aspect ratios. This combination of scale and structured coverage makes it a potentially useful resource for studying text-to-image fine-tuning. Its release could reduce the substantial cost of assembling comparable training data and enable further research on data selection and curation.

  2. Encouraging results across different model families. Experiments on LlamaGen, SD-XL, and DeCo demonstrate the dataset’s applicability to different generation frameworks. Human evaluation indicates improved visual quality after fine-tuning, particularly for LlamaGen and SD-XL, supporting the dataset’s practical utility.

Critical weaknesses
  1. Insufficient validation of the filtering strategy. The pipeline combines Aesthetic Predictor V2.5 with reasoning-VLM verification, but provides no controlled comparison with alternatives such as HPSv3 or HPSv3++. The limitations of scalar reward models do not, by themselves, establish the effectiveness of the proposed approach. Given the reported preference-prediction and downstream optimization results of these alternatives, a comparison on the same candidate pool, with matched retained-data sizes and independent quality assessment, would better justify the filtering choices.

  2. Limited evaluation coverage. The evaluation relies primarily on human preferences over 500 prompts, while GenEval is the only automatic benchmark reported. Human evaluation is useful for assessing visual quality, but the lack of complementary benchmarks leaves complex instruction following, text rendering, and other claimed capabilities insufficiently evaluated. Results on a representative selection of DPG-Bench, TIIF-Bench testmini, OneIG-EN, or Qwen-Image-Bench would strengthen the evidence. Confidence intervals and inter-annotator agreement would also help assess the reliability of the human evaluation.

  3. Declines in basic instruction-following capabilities. GenEval evaluates basic capabilities such as counting, spatial relations, and attribute binding, which remain essential to text-to-image generation regardless of aesthetic quality. DeCo’s overall score decreases from 0.82 to 0.77, with declines in counting (0.66 to 0.60), position (0.72 to 0.54), and attribute binding (0.77 to 0.68). Correctly satisfying these constraints is compatible with producing aesthetically appealing images. I therefore do not consider improvements in visual preference a sufficient justification for these declines, nor should they be dismissed as merely a benchmark-specific effect.

  4. Confounded and incomplete dataset comparisons. Fine-T2I, T2I-2M, and BLIP3o-60k are each used for one epoch despite substantial differences in dataset size, making it difficult to separate data quality from training scale. BLIP3o-60k’s lower training loss does not resolve this issue, since losses on different data distributions are not directly comparable measures of downstream utility. Comparisons with matched training budgets and matched sample counts are needed. Including representative text-to-image subsets from OpenGPT-4o-Image, ShareGPT-4o-Image, or Echo-4o-Image would also provide a more complete assessment against relevant synthetic-data resources.

  5. Limited validation on contemporary latent-space flow-matching models. Although DeCo incorporates a DiT and a flow-matching objective, the study lacks fine-tuning experiments on representative latent-space flow-based transformers, such as FLUX.1-dev from the FLUX family. Evaluating such a backbone would help establish whether the dataset’s benefits extend to widely used models with stronger initial generation capabilities, strengthening the evidence for its generalizability.

Questions for authors
  • What controlled experiments show that the proposed filtering pipeline produces better training data than HPSv3 or HPSv3++?

  • Do the reported gains hold on complementary benchmarks such as DPG-Bench, TIIF-Bench testmini, OneIG-EN, or Qwen-Image-Bench?

  • Does Fine-T2I retain its advantage under matched training budgets and data sizes, including against recent datasets such as OpenGPT-4o-Image, ShareGPT-4o-Image, or Echo-4o-Image?

  • Have the authors tested whether the gains extend to latent-space flow-based models such as FLUX.1-dev?

6 comments
AuthorsOpenReview ↗

Rebuttal (part 1 / 3)

Dear Reviewer aWUf,

Thank you for your valuable questions and suggestions. Here is our rebuttal. To address your concerns and questions, we ran 15 new fine-tuning experiments on three different backbones (LlamaGen-XL, DeCo, FLUX.1-dev), generated and evaluated ~100k new images, added four benchmarks and five reward models, and directly validated our filtering strategy.


W1 & Q1: Validation of the filtering strategy. We will revise the reward-hacking paragraph accordingly. To validate the filtering strategy, we have compared filters on 5,530 images generated using the GenEval prompts by three models (LlamaGen-XL, DeCo, FLUX.1-dev). We labelled correct or incorrect by GenEval's official detector. Our verifier (Qwen3-VL-8B-Thinking with exactly the Fig. 11 system prompt) is compared with top-k selection by five scalar rewards, all keeping the same 59.0% of the images:

Filter (59% kept)Correct among kept (%)Incorrect removed (%)
Random56.139.2
Aesthetic v2.558.142.2
CLIPScore63.950.2
PickScore68.155.9
ImageReward69.557.9
HPSv375.566.1
Ours (reasoning VLM)80.573.1

Clearly, our method with our system prompt significantly outperforms the strongest method HPSv3 by 5.0 points. Besides, these only care about alignment, but our verifier also rejects artifacts that they don't care.


W2 & Q2: Evaluation coverage, CIs and agreement. We added additional benchmarks, including DPG-Bench, TIIF-Bench testmini, LongText-Bench and paraphrased GenEval (the latter is reported under W3):

BenchmarkBackboneBase → + Fine-T2I
DPG-BenchLlamaGen-XL (28k)64.8 → 67.6 (BLIP3o-60k: 65.8, T2I-2M: 65.9)
DPG-BenchDeCo81.7 → 82.5
DPG-BenchFLUX.1-dev83.3 → 84.5
TIIF-Bench testmini (short / long)FLUX.1-dev73.2 / 75.3 → 75.7 / 79.3
LongText-BenchFLUX.1-dev0.566 → 0.653

On TIIF-Bench, the biggest improvements are from long prompts and on the text dimension (+15.8 / +14.0 for short / long prompts). Fine-T2I improves LongText-Bench in all eight text categories.

CIs for the original human study (win − loss with 95% CI, visual quality / alignment): LlamaGen +73.5 [68.4, 78.6] / +52.9 [46.7, 59.1]; SD-XL +48.9 [42.4, 55.4] / +33.1 [26.3, 39.9]; DeCo +42.3 [36.1, 48.5] / +6.8 [0.5, 13.1]. All of them are significant.

Agreement for the new results: we additionally evaluated 4,500 pairs with an independent SOTA model GPT-5.5. For all comparisons, they agree on 74.1% for alignment and 68.7% for quality, showing great agreement.

AuthorsOpenReview ↗

Rebuttal (part 2 / 3)

W3: Declines in basic instruction following. We agree these obersvations should not be ignored, and we have identified the reason behind.

BLIP3o-60k includes geneval_train set (7.1k samples, 11.8%) from the GenEval prompt, and DeCo's checkpoint was finetuned on BLIP3o-60k. We added this set in and out at a matched budget of 28k samples, including continued fine-tuning of DeCo itself:

BackboneFine-tuning data (28k)GenEvalCountingPositionGenEval (paraphrased)DPG-Bench
DeConone (public checkpoint)79.1617075.581.7
DeCoFine-T2I73.5525573.082.5
DeCoFine-T2I + 11.8% compositional80.4647275.982.9
LlamaGen-XLFine-T2I40.730838.267.6
LlamaGen-XLFine-T2I + 11.8% compositional49.4401546.667.7
LlamaGen-XLBLIP3o-60k48.1391246.365.8
LlamaGen-XLBLIP3o-60k w/o geneval_train39.828739.5–
  • Cause. Without its geneval_train part, BLIP3o-60k falls below Fine-T2I on GenEval (39.8 vs. 40.7).
  • Fix. Mixing these samples into Fine-T2I removes the decline on DeCo (80.4 vs. 79.1, position 72 vs. 70). We employed GPT-4.1 and GPT-5.5 to further evaluate, and we see the preference gains are real. GPT-4.1 prefers our counterpart over the public DeCo in visual quality (+17.2, p < 0.001, confirmed by GPT-5.5: +15.2) and in alignment (+8.8, p = 0.04). We also improve HPSv3, ImageReward, PickScore, Aesthetic and CLIPScore metrics.
  • Not a general loss. Even without mixing geneval_train, Fine-T2I improves DPG-Bench from 81.7 to 82.5, and the gains persist on our paraphrased GenEval.

We will remove the "benchmark-specific" word and add the new mixing experiments.


W4 & Q3: Matched comparisons, including GPT-4o-distilled datasets. Thanks for pointing out this and we will remove the presentation of training-loss. Here, we fine-tuned LlamaGen-XL on exactly 28k samples from each of seven datasets, with identical settings (580 steps, batch 192, lr 3e-5, 512², ≤120-token captions). We report Fine-T2I win rate (win-loss) evaluated by GPT-4.1 comapred with other datasets in the first two columns:

Data (28k each)We Win AlignmentWe Win Visual qualityImageRewardPickScoreAesthetic
none (base)+33.6**+40.0**0.1020.7885.24
BLIP3o-60k−0.6+15.0**0.5310.8095.38
T2I-2M+17.0**+7.80.4640.8085.40
ShareGPT-4o-Image+3.4+14.8**0.5260.8065.32
OpenGPT-4o-Image−2.2+5.40.5720.8095.28
Echo-4o-Image+25.8**+16.4**0.4560.8015.33
Fine-T2I——0.6120.8125.44

Answer to Q3. According to above table, our Fine-T2I significantly better than all other datasets across different evaluations and benchmarks.

  • It is significantly better in visual quality over BLIP3o-60k, ShareGPT-4o-Image and Echo-4o-Image, and in alignment over T2I-2M and Echo-4o-Image.
  • It is compared with OpenGPT-4o-Image in pairwise judgment, while higher on DPG-Bench (67.6 vs. 67.4) and on all three reward models.
  • It obtains the highest ImageReward, PickScore and Aesthetic scores of all seven datasets.
AuthorsOpenReview ↗

Rebuttal (part 3 / 3)

W5 & Q4: Latent flow-matching models. We also finetuned FLUX.1-dev with LoRA (rank 32, 750 steps, batch 16) on 22k Fine-T2I images at native aspect ratios. Both models are sampled with identical settings (1024², 28 steps, guidance 3.5, same seeds). The last two columns compare the fine-tuned model with the base model in alignment (win − loss):

ModelGenEvalCountingPositionDPG-BenchTIIF (short / long)LongTextImageRewardGPT-4.1GPT-5.5
FLUX.1-dev63.1642083.373.2 / 75.30.5661.031——
+ Fine-T2I67.9742784.575.7 / 79.30.6531.107+19.8+24.8

Both judge results have p < 0.001.

Even on the strong model FLUX.1-dev, we can see Fine-T2I improves instruction following and text rendering abilities significantly.

  • Non-square generation. A LoRA trained on square and random-aspect-ratio Fine-T2I images is preferred over the base model in alignment at 16:9 (+20.5, p = 0.003) and at 9:16 (+13.0).
  • Half-strength LoRA. At scale 0.5, the LoRA keeps the base model's aesthetic score (6.15 vs. 6.12) while retaining most of the gains (GenEval 67.5, DPG-Bench 84.1).
  • Not self-distillation. FLUX.1-dev is not one of the generators used to build Fine-T2I.

We will add all of these results to the revision. Thank you again for the insightful review and questions, which makes our work stronger. We hope our responses address your concerns and would be grateful if you would consider raising your score. We are happy to discuss any further questions.

Best,

The Authors

Reviewer_NfTeOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Fine-T2I, a large-scale dataset for high-quality text-to-image fine-tuning, comprising over six million synthetic text–image pairs and approximately 168,000 curated real images. Its construction pipeline combines controlled prompt generation, semantic deduplication, original and enhanced prompts, diverse resolutions and aspect ratios, and aesthetic and reasoning-VLM-based filtering. The authors evaluate the dataset by fine-tuning LlamaGen, SD-XL, and DeCo, using human preference studies, qualitative comparisons, and GenEval. The results show improvements primarily in visual quality, with more variable gains in text–image alignment and automatic metrics.

Critical strengths
  • A valuable resource for the community. High-quality, publicly accessible fine-tuning data remains an important bottleneck for open text-to-image research. Fine-T2I addresses this need at a substantial scale, with the potential to support further work on generation quality and instruction following.

  • A carefully designed construction pipeline. The dataset goes beyond collecting existing image-text pairs. Its explicit distribution design, multiple prompt granularities, diverse image formats, and detailed filtering provide a practical recipe for constructing high-quality fine-tuning data.

  • Validation across different model families. Experiments cover autoregressive, latent-diffusion, and pixel-space diffusion models. Human evaluations show consistent preferences for the fine-tuned models in visual quality, supporting the dataset's practical usefulness across these architectures.

Critical weaknesses
  • Cross-dataset comparisons do not control for training budget. Each dataset is used for approximately one epoch despite substantial differences in dataset size. Consequently, the comparisons vary both the amount of training data and the training computation, making it difficult to isolate the contribution of data quality. The results demonstrate the usefulness of the complete resource, but provide weaker evidence for its superiority under a fixed budget. Lower training loss on BLIP3o-60k does not resolve this confound.

  • Benefits for stronger models remain unverified. The evaluation is limited to LlamaGen, SD-XL, and DeCo. Since the observed gains depend on the base model's initial capability, it remains unclear whether Fine-T2I offers substantial improvements for stronger models such as Qwen-Image. This limits the evidence for its usefulness in further improving models that already have strong generation quality and instruction following.

Questions for authors
  1. Under a matched training budget, such as the same number of optimization steps and comparable image resolution, how does Fine-T2I compare with BLIP3o-60k and T2I-2M? A controlled comparison would help establish whether the advantage persists independently of additional training computation.

  2. Can the authors provide evidence of the dataset's benefit for a stronger base model, particularly one already trained with high-quality fine-tuning data? This would help clarify whether Fine-T2I primarily improves less capable models or also provides additional value at a higher capability level.

1 comment
AuthorsOpenReview ↗

Response to Reviewer NfTe

Dear Reviewer NfTe,

Thank you for your positive assessment and focused questions. Here is our rebuttal. To address your concerns and questions, we ran 15 new fine-tuning experiments on three different backbones (LlamaGen-XL, DeCo, FLUX.1-dev), and generated and evaluated ~100k new images.


W1 & Q1: Matched training budget. We agree. We fine-tuned LlamaGen-XL on the same number of samples (28k), with the same number of steps (580) and the same resolution (512²), for all seven datasets; all other settings were also identical. The first two columns report Fine-T2I vs. each model, judged by GPT-4.1 (win − loss; ** p < 0.001):

Data (28k each)AlignmentVisual qualityImageRewardPickScoreAesthetic
none (base)+33.6**+40.0**0.1020.7885.24
BLIP3o-60k−0.6+15.0**0.5310.8095.38
T2I-2M+17.0**+7.80.4640.8085.40
ShareGPT-4o-Image+3.4+14.8**0.5260.8065.32
OpenGPT-4o-Image−2.2+5.40.5720.8095.28
Echo-4o-Image+25.8**+16.4**0.4560.8015.33
Fine-T2I——0.6120.8125.44

Answer to Q1. Under this matched budget, Fine-T2I:

  • is significantly preferred over BLIP3o-60k in visual quality (+15.0) and over T2I-2M in alignment (+17.0);
  • is never significantly worse than any alternative;
  • scores highest of all seven datasets on ImageReward, PickScore and Aesthetic;
  • scores higher on DPG-Bench than BLIP3o-60k and T2I-2M (67.6 vs. 65.8 / 65.9).

Besides, the GenEval gap to BLIP3o-60k comes from its GenEval-style geneval_train subset. Without it, BLIP3o-60k scores 39.8 vs. 40.7 for Fine-T2I. Including this set to our Fine-T2I, our results show better performance than the base model. Please also see our rebuttal to Reviewer i6cn and aWUf for detailed results on this. We will remove the discussion on training-loss.


W2 & Q2: A stronger, already-tuned base model. To address this, We fine-tuned FLUX.1-dev with LoRA on 22k Fine-T2I images at native aspect ratios, with rank of 32, 750 steps, a batch size of 16. We also sample both models under same settings (1MP, 28 steps, guidance 3.5, same seeds). We report the alignment win rate (win − loss) evaluated by GPT-4.1 and GPT-5.5 in the last two columns:

ModelGenEvalCountingPositionDPG-BenchTIIF (short / long)LongTextImageRewardGPT-4.1GPT-5.5
FLUX.1-dev63.1642083.373.2 / 75.30.5661.031——
+ Fine-T2I67.9742784.575.7 / 79.30.6531.107+19.8+24.8

Clearly, finetuned FLUX.1-dev on our Fine-T2I significantly improve all metrics, for difference evaluation aspects.

Both judge results have p < 0.001. Beside, The word accuracy in CVTG-2K also improves from 50.5 to 65.7.

Answer to Q2: yes, mainly in instruction following and text rendering. For a base model that is already strongly aesthetically tuned, the benefit are more about compositional ability and text, and litter improvements for aesthetics. At LoRA scale 0.5, We slightly improve the base model's aesthetic score (6.15 vs. 6.12) and keep most of the gains (GenEval 67.5, DPG-Bench 84.1). Since FLUX.1-dev is not one of the image generation models used to build Fine-T2I, so this is not self-distillation and results are fair. Qwen-Image (20B) was beyond our compute budget during the rebuttal period.


For Ethics flag: data sources. The synthetic images were generated by us with publicly released models (Z-Image and FLUX.2 dev). And the prompts are generated by public LLM. the 168k real images come from Pexels, Pixabay and Unsplash and are used under their free-use licenses. Any image that raises a concern will be removed.


Thank you again for your support and helpful review. We hope our rebuttal addresses your concerns and answers your questions. We also would be grateful if you would consider raising your score. Please feel free to let us know if you have any additional quetions or concerns.

Best,

The Authors

Reviewer_i6cnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Fine-T2I is a dataset for text-to-image fine-tuning containing 6,145,693 synthetic text–image pairs and 168,424 curated real images. Its construction combines controlled prompt generation, semantic deduplication, original and enhanced prompts, varied resolutions and aspect ratios, aesthetic filtering, and reasoning VLM verification. The paper evaluates fine-tuning on LlamaGen, SD-XL, and DeCo using human preferences on 500 prompts and GenEval, and compares against two existing fine-tuning datasets using LlamaGen. The strongest evidence is improved human-rated visual quality across all three model families. The instruction-following results are more mixed.

Critical strengths
  • A potentially valuable training resource at substantial scale. The combination of approximately 6.15 million synthetic pairs, curated photography, multiple prompt granularities, and varied image formats addresses a practical need for accessible fine-tuning data. The documented category, style, and resolution coverage could make the resource useful across different training pipelines.

  • Consistent improvements in human-perceived visual quality across architectures. Figure 6 shows substantial visual-quality gains for LlamaGen, SD-XL, and DeCo after fine-tuning on Fine-T2I. These improvements are supported by a reasonably rigorous human evaluation (Appendix C), providing evidence of the dataset's effectiveness.

Critical weaknesses
  • Dataset comparisons do not isolate the effects of data quality, scale, or curation. The authors train each dataset for one epoch at its original size, giving Fine-T2I (~6M samples) substantially more training exposure than BLIP3o-60k or T2I-2M. BLIP3o-60k's lower training loss does not rule out a scale effect, since losses across different training distributions are not directly comparable. Without matched-budget comparisons or curation ablations, the results cannot establish that data quality, rather than scale, drives the improvements.

  • The explanation for DeCo's GenEval regression is not sufficiently supported. DeCo's GenEval score decreases from 0.82 to 0.77, with declines in counting (0.66→0.60), position (0.72→0.54), and attribute binding (0.77→0.68). Human alignment preferences are also relatively modest (29.1% wins, 22.3% losses, 48.6% ties). The authors attribute the GenEval drop to benchmark-distribution effects, but provide no controlled evidence distinguishing this explanation from degraded compositional alignment.

  • Several proposed dataset capabilities lack targeted validation. While human evaluations show improvements in visual quality, both studies use the same 500-prompt suite without separately assessing complex instruction following, prompt granularity, or text rendering. The benefits of diverse resolutions and aspect ratios are also not evaluated, with no mixed-resolution versus square-only comparison. Thus, it remains unclear whether the improvements extend to these capabilities.

Questions for authors
  • Does Fine-T2I retain its advantage over BLIP3o-60k and T2I-2M under matched training budgets (updates or samples seen)?

  • How do the authors distinguish DeCo's proposed benchmark-distribution effect from a genuine loss of compositional alignment? Evaluation on complementary benchmarks, such as GenEval 2 or DPG-Bench, would help clarify this. Results on TIIF-Bench for instruction following and CVTG-2K or LongText-Bench for text rendering would also strengthen the broader capability claims.

  • How are the varied resolutions and aspect ratios handled during fine-tuning for each backbone (e.g., resizing, cropping, or resolution bucketing)? Does mixed-resolution training outperform square-only training?

5 comments
AuthorsOpenReview ↗

Response to Reviewer i6cn (1/2)

Dear Reviewer i6cn,

Thanks a lot for your detailed comments and helpful questions, which help us a lot to improve our work. Here is our rebuttal.

To answer your concerns and questions, we further added 15 sft experiments on three different models (LlamaGen-XL, DeCo, FLUX.1-dev), generated and evaluated around 100k new images, and tested DPG-Bench, TIIF-Bench, LongText-Bench and CVTG-2K.


W1 & Q1: Quality vs. scale under a matched budget.

Even the dataset size is a big contribution, we do agree that losses on different distributions are not comparable. To this end, we finetuned LlamaGen-XL on exactly 28k samples from each of seven datasets with identical settings (580 steps, batch 192, lr 3e-5, 512², ≤120-token captions). Besides the auto-eval scores like ImageReward, PickScore, and Aesthetic, we also employ GPT-4.1 to report the win rate (win − loss) of Fine-T2I over other datasets in the first two columns:

Data (28k each)AlignmentVisual qualityImageRewardPickScoreAesthetic
none (base)+33.6**+40.0**0.1020.7885.24
BLIP3o-60k−0.6+15.0**0.5310.8095.38
T2I-2M+17.0**+7.80.4640.8085.40
ShareGPT-4o-Image+3.4+14.8**0.5260.8065.32
OpenGPT-4o-Image−2.2+5.40.5720.8095.28
Echo-4o-Image+25.8**+16.4**0.4560.8015.33
Fine-T2I——0.6120.8125.44

Answer to Q1: yes. Under the same number of samples and updates, we can see our Fine-T2I is:

  • significantly better than BLIP3o-60k in visual quality (+15.0) and over T2I-2M in alignment (+17.0);
  • clearly better than all other SFT datasets;
  • highest of all seven datasets on ImageReward, PickScore and Aesthetic;
  • higher on DPG-Bench than both BLIP3o-60k and T2I-2M (67.6 vs. 65.8 / 65.9).

W2 & Q2 (first part): DeCo's GenEval decline.

We agree that our original claim lacked evidences and experimental supports. Here, we present controlled experiments on DeCo.

Note that BLIP3o-60k includes geneval_train part (7.1k samples, 11.8%) from the GenEval prompt distribution, and DeCo's checkpoint was fine-tuned on BLIP3o-60k. Hence, we included this set in and out at a matched budget of 28k Fine-T2I samples. We re-generate all samples, so the reported results may be slightly different from our paper.

BackboneFine-tuning data (28k)GenEvalCountingPositionGenEval (paraphrased)DPG-Bench
DeConone (public checkpoint)79.1617075.581.7
DeCoFine-T2I73.5525573.082.5
DeCoFine-T2I + 11.8% compositional80.4647275.982.9
LlamaGen-XLFine-T2I40.730838.267.6
LlamaGen-XLFine-T2I + 11.8% compositional49.4401546.667.7
LlamaGen-XLBLIP3o-60k48.1391246.365.8
LlamaGen-XLBLIP3o-60k w/o geneval_train39.828739.5–
  • Cause. Without geneval_train set, BLIP3o-60k even shows worse results than our Fine-T2I on GenEval (39.8 vs. 40.7).
  • Fix. Adding this set into Fine-T2I removes the decline on DeCo (80.4 vs. 79.1, position 72 vs. 70). And we can see GPT-4.1 prefers our results over the public DeCo in both visual quality (+17.2, p < 0.001; GPT-5.5: +15.2) and alignment (+8.8, p = 0.04).
  • Not a general loss. Without mixing, DeCo trained on Fine-T2I still improves DPG-Bench (81.7 → 82.5). We also see the improvements on paraphrased GenEval, where we rewrite the GenEval prompts to natural sentence to shift the GenEval-style prompt distribution.

Besides, the human alignment result for DeCo (29.1% wins vs. 22.3% losses) presents a win−loss of +6.8 [0.5, 13.1] (95% CI), which is a clear gain. We will remove the "benchmark-specific" explanation and add these analysis. Thanks for pointing out these.

AuthorsOpenReview ↗

Response to Reviewer i6cn (2/2)

W3 & Q2 (second part): Targeted validation of capabilities. Thank you for the suggested benchmarks. We evaluated additional evlaution metrics as below:

Benchmark (capability)BackboneBase → + Fine-T2I
DPG-Bench (complex instructions)LlamaGen-XL / DeCo / FLUX.1-dev64.8 → 67.6 / 81.7 → 82.5 / 83.3 → 84.5
TIIF-Bench testmini, short / long prompts (instruction following)FLUX.1-dev73.2 / 75.3 → 75.7 / 79.3
LongText-Bench (text rendering)FLUX.1-dev0.566 → 0.653
CVTG-2K, word accuracy (multi-region text)FLUX.1-dev50.5 → 65.7
  • Prompt granularity. From results on TIIF-Bench, we see our dataset improve the long prompts by +4.0, and improve the short prompts by +2.5. On LongText-Bench, we also improve long and short prompts clearly, 0.538 → 0.609 and 0.631 → 0.756, respectively.

  • Text rendering. TIIF-Bench's text is improved by +15.8 / +14.0 (short / long). On CVTG-2K (400 prompts, 100 per number of text regions; PaddleOCR word accuracy), word accuracy improves for every region count: 2 regions 64.1 → 73.5, 3 regions 54.0 → 69.5, 4 regions 45.5 → 65.4, 5 regions 38.5 → 54.5.

  • Why FLUX.1-dev for text. We evaluate text rendering on FLUX.1-dev because we notice that LlamaGen, SD-XL and DeCo shows clear performance gap when compared with FLUX.1-dev.


Q3: Resolution handling; mixed vs. square. In the paper, all three backbones were fine-tuned at a single square resolution: images were resized and cropped to each backbone's default resolution (512² for LlamaGen and DeCo, 1024² for SD-XL), with no bucketing. We will state this in Sec. 5.

The random-aspect-ratio subsets are designed for backbones that train at native aspect ratios. Evaluated by GPT-4.1, a LoRA for FLUX.1-dev trained on square and random-aspect-ratio Fine-T2I images is preferred over the base model in alignment at 16:9 (+20.5, p = 0.003) and at 9:16 (+13.0), showing great improvements.


We will add all of these results to the revision. Thank you again for the insightful review. We hope our responses address your concerns and would be grateful if you would consider raising your score. We are happy to discuss any further questions.

Best,

The Authors

AuthorsOpenReview ↗

Thanks for your time and helpful comments

Dear Reviewer i6cn,

Thanks a lot for your efforts and time, and we understand this is a busy time.

Considering the discussion period will end in less than two days, we would like to gently follow up on our response.

In summary:


  • W1 & Q1: with the same 28k samples and training steps for all seven datasets, Fine-T2I is significantly better over BLIP3o-60k and over T2I-2M, and has the highest reward scores even compared with other SFT datasets.

  • W2 & Q2: Our experiments further validate the DeCo decline comes from the GenEval-style subset in BLIP3o-60k. Adding it to Fine-T2I brings DeCo to 80.4 (vs. 79.1).

  • W3 & Q2: Fine-T2I improves other benchmarks, including DPG-Bench, TIIF-Bench, LongText-Bench and CVTG-2K, as supported by our experimental results.

  • Q3: we clarified the resolution settings, and FLUX.1-dev LoRA trained with mixed aspect ratios improves alignment at 16:9 and 9:16.


If our rebuttal has addressed your concerns and answered your questions, we would sincerely appreciate it if you would consider updating your score.

Thank you again for your time and thoughtful review.

Best,

The Authors

Comments on the paper

5 comments
Area_Chair_Jj2NOpenReview ↗

Reviewer-author discussion period is ending. Please feedback.

Dear Reviewers,

Thanks for your time and effort in reviewing ICLR2027 submissions.

As the reviewer-author discussion period will end very soon, please take time to read the author’s rebuttal and give your feedback. Your continuous engagement is highly appreciated.

Best regards,

Your AC

Dear AC and Reviewers,

Thanks for AC's reminder and all reviewers' comments.

As the discussion period is closing, we would like to briefly summarize our rebuttal.

To answer the questions, we ran 15 new fine-tuning experiments on three backbones (LlamaGen-XL, DeCo, FLUX.1-dev) and generated and evaluated about 100k new images.

All experimental results show the clear advantages of our Fine-T2I.


Concerns raised by several reviewers

1. Matched training budget (i6cn, NfTe, aWUf).

Same 28k samples and 580 steps on LlamaGen-XL for all seven datasets; first two columns are Fine-T2I's win − loss by GPT-4.1:

Data (28k each)AlignmentVisual qualityImageRewardPickScoreAesthetic
BLIP3o-60k−0.6+15.0**0.5310.8095.38
T2I-2M+17.0**+7.80.4640.8085.40
ShareGPT-4o-Image+3.4+14.8**0.5260.8065.32
OpenGPT-4o-Image−2.2+5.40.5720.8095.28
Echo-4o-Image+25.8**+16.4**0.4560.8015.33
Fine-T2I——0.6120.8125.44

Conclusion: Fine-T2I is clearly better than other SFT datasets, evaluated by both SOTA models and auto benchmarks.

2. DeCo's GenEval decline (i6cn, aWUf).

DeCo was fine-tuned on BLIP3o-60k, which contains a GenEval-style subset (11.8%):

DeCo fine-tuning dataGenEvalCountingPositionDPG-Bench
none (public checkpoint)79.1617081.7
Fine-T2I73.5525582.5
Fine-T2I + 11.8% GenEval-style80.4647282.9

Conclusion: Experiments validate that such decline is from GenEval-style subset in BLIP3o-60k (train & test from same distribution), not Fine-T2I. With it, DeCo + FineT2I outperforms public checkpoint on GenEval and is also preferred by GPT-4.1 (+17.2 quality, +8.8 alignment).

3. Evaluation beyond GenEval and the 500-prompt suite (i6cn, aWUf).

BenchmarkBackboneBase → + Fine-T2I
DPG-BenchLlamaGen-XL / DeCo / FLUX.1-dev64.8 → 67.6 / 81.7 → 82.5 / 83.3 → 84.5
TIIF-Bench (short / long)FLUX.1-dev73.2 / 75.3 → 75.7 / 79.3
LongText-BenchFLUX.1-dev0.566 → 0.653
CVTG-2K (word accuracy)FLUX.1-dev50.5 → 65.7

Conclusion: With Fine-T2I, we see clear gains on every benchmark and backbone, and on different models.

4. Stronger, latent flow-matching backbone (NfTe, aWUf).

LoRA fine-tuning of FLUX.1-dev on 22k Fine-T2I images:

ModelGenEvalCountingPositionDPG-BenchImageRewardGPT-4.1GPT-5.5
FLUX.1-dev63.1642083.31.031——
+ Fine-T2I67.9742784.51.107+19.8+24.8

Conclusion: Fine-T2I still helps a strong and aesthetically tuned model across various evaluations.


Other points

  • Filtering (aWUf):

    • Results: 80.5% of images kept by our verifier are correct, vs. 75.5% for HPSv3 and 56.1% for random (+5.0, 95% CI [3.4, 6.7]).
    • Conclusion: Our verifier is the best filter amony tested.
  • Human study CIs (aWUf):

    • Results: win − loss in quality / alignment is +73.5 / +52.9 (LlamaGen), +48.9 / +33.1 (SD-XL), +42.3 / +6.8 (DeCo).
    • Conclusion: All improvements are significant.
  • Second judge (aWUf):

    • Results: GPT-5.5 agrees with GPT-4.1 on 74.1% (alignment) and 68.7% (quality) on 4,500 pairs.
    • Conclusion: Our good performance are not judge-specific, holds for different judgers.
  • Paraphrased GenEval (i6cn, aWUf):

    • Results: DeCo + Fine-T2I 75.9 vs. 75.5 for the public checkpoint.
    • Conclusion: The improvements from Fine-T2I are hold under paraphrased prompts.
  • Prompt granularity (i6cn):

    • Results: TIIF-Bench +4.0 on long prompts; LongText-Bench 0.538 → 0.609 (long) and 0.631 → 0.756 (short).
    • Conclusion: Both short and long prompts improve a lot.
  • Text rendering (i6cn, aWUf):

    • Results: TIIF-Bench text +15.8 / +14.0; CVTG-2K with 4 text regions 45.5 → 65.4.
    • Conclusion: Largest gains on harder text.
  • Aspect ratios (i6cn):

    • Results: alignment vs. base +20.5 at 16:9 aspect ratio and +13.0 at 9:16 aspect ratio.
    • Conclusion: Fine-T2I helps non-square generation a lot.
  • Aesthetics (NfTe):

    • Results: FLUX.1-dev keeps its aesthetic score (6.15 vs. 6.12) with GenEval 67.5 and DPG-Bench 84.1.
    • Conclusion: Fine-T2I can improve strong models without aesthetic cost, or even slightly improve.

We will include these in the revision.

We are grateful for the reviewers' comments and suggestions, which helped us improve the paper a lot, and we are happy to answer any further questions.

Best,

Authors

AuthorsOpenReview ↗

Dear Reviewers, thanks for your time!

Dear Reviewers,

Thank you again for your time and effort in reviewing our paper and providing valuable suggestions!

We understand that this is a busy time. We would like to kindly ask whether our detailed explanations and extensive experiments have addressed your concerns.

Through the additional experiments for your questions / concerns, we have further validated the advantages of our Fine-T2I dataset and demonstrated that our dataset can benefit models much more than other SFT datasets.

We are happy to address any further questions or suggestions, even on the last day of the discussion period.

Thank you again, and we look forward to your further comments.

Best,

Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.