The success of modern text-to-image generation is largely attributed to massive, high-quality datasets. Currently, these datasets are curated through a filter-first paradigm that aggressively discards low-quality raw data based on the assumption that it is detrimental to model performance. Is the discarded bad data truly useless, or does it hold untapped potential? In this work, we critically re-examine this question. We propose LACON (Labeling-and-Conditioning), a novel training framework that exploits the underlying uncurated data distribution. Instead of filtering, LACON re-purposes quality signals, such as aesthetic scores and watermark probabilities as explicit, quantitative condition labels. The generative model is then trained to learn the full spectrum of data quality, from bad to good. By learning the explicit boundary between high- and low-quality content, LACON achieves superior generation quality compared to baselines trained only on filtered data using the same compute budget, proving the significant value of uncurated data.
Figures & tables
Figure 2: Comparison of generations produced by the baseline trained on full raw data and by LACON under different conditioning settings . From left to right: (1) model trained on the full raw dataset without quality conditioning; (2) LACON conditioned on a low aesthetic score saes ; (3) LACON conditioned on a high aesthetic score saes ; (4) LACON jointly conditioned on high aesthetic score saes and high clarity score scla .
Figure 3: Overview of LACON . High-level training pipeline that repurposes quality signals as explicit attribute conditioning during training.
Metric ( smetric )
Meaning
Value Range
Clipping
Tokens ( N )
Centroid Anchors ( pimetric )
Aesthetic ( saes )
Aesthetic quality
0–10
None
10
{0.5, 1.5, …, 9.5}
Watermark ( swat )
Watermark probability
0–1
None
10
{0.05, 0.15, …, 0.95}
Clarity ( scla )
Perceptual sharpness
0–1M+
>3000
10
{150, 450, …, 2850}
Entropy ( sent )
Information density
0–8
None
8
{0.5, 1.5, …, 7.5}
Luminance ( sluma )
Visual brightness
0–1
None
10
{0.05, 0.15, …, 0.95}
Table 1: Configuration of cluster-centroid anchors for each attribute embedding. The number and placement of anchors ( pi ) are selected based on the value range of the corresponding quality signals.
Figure 4: Visualization comparison of LACON-S and LACON-A against Baselines, which demonstrate LACON can still achieve superior visual generation quality even when training on the full set images without filtering.
Figure 5: (a) Comparison of images jointly conditioned on aesthetic score saes and HSV-luma score sluma . From left to right, the columns correspond to saes=3 , 5 , 7 . From top to bottom, the rows correspond to sluma=0.3 , 0.4 , 0.5 . (b) Comparison of images jointly conditioned on clarity lapvar score scla and entropy score sent . From left to right, the columns correspond to scla=200 , 1000 , 2500 . From top to bottom, the rows correspond to sent=5 , 6 , 7 .
Ratio
saes-min
swat-min
scla-min
sent-min
sluma-min
sluma-max
≈5%
5.0
0.3
800.0
6.0
0.1
0.9
≈30%
4.0
0.5
600.0
4.0
0.1
0.9
≈50%
3.5
0.6
500.0
3.0
0.1
0.9
≈65%
3.0
0.7
400.0
2.0
0.1
0.9
≈80%
3.0
0.8
200.0
2.0
0.1
0.9
Table 2: The filtering thresholds for different filtering ratios.
Ratio
GenEval ↑
DPG ↑
FID ↓
≈5%
58.3
73.4
14.0
≈30%
62.1
73.6
13.7
≈50%
64.5
74.5
13.4
≈65%
68.0
76.1
13.1
≈80%
67.5
75.8
13.5
100%
67.0
75.4
13.5
Table 3: Quantitative comparisons on Sana-1.6B are conducted between models trained on subsets obtained under different filtering thresholds, evaluated on GenEval, DPG and FID. The results reveal a clear trade-off between data quantity and data quality in terms of model performance.
512 × 512 resolution
Model
Sana-0.6B
Sana-1.6B
Metric
GenEval ↑
DPG ↑
FID ↓
GenEval ↑
DPG ↑
FID ↓
Baseline-A
55.0
68.1
14.9
58.3
73.4
14.0
Baseline-B
62.8
70.4
13.4
68.0
76.1
13.1
Baseline-C
60.8
69.7
15.4
67.0
75.4
13.5
Baseline-D
61.0
69.2
15.3
67.1
75.5
13.5
Table 4: Quantitative comparisons on Sana-0.6B and Sana-1.6B, evaluated using GenEval, DPG, and FID, demonstrate that training on the full unfiltered dataset with explicit conditioning enables LACON to consistently outperform the strongest baseline at both 512 × 512 and 1024 × 1024 resolutions.
Model
Qwen3-0.6B
Qwen3-1.7B
Metric
GenEval ↑
DPG ↑
FID ↓
GenEval ↑
DPG ↑
FID ↓
Baseline-B
58.9
73.1
12.4
66.4
78.2
12.1
LACON-S (Ours)
61.3
74.9
11.4
69.7
79.4
11.0
LACON-A (Ours)
61.3
75.4
11.2
70.3
80.1
10.9
Table 5: Quantitative comparisons on Qwen3-0.6B and Qwen3-1.7B, evaluated using GenEval, DPG, and FID, demonstrate that training on the full unfiltered dataset with explicit conditioning enables LACON to consistently outperform the strongest baseline.
Figure 6: The evidence of knowledge gap between Baseline-B (the first row) and LACON (the second row). From left to right, the concepts are about morpho butterfly, pastel portrait of elderly woman, mineral terraces, wolf and camel.
Method
GenEval ↑
DPG ↑
FID ↓
Linear Interpolation
69.5
76.1
12.7
Discrete Binning
69.8
77.2
12.4
Fourier Feature
70.4
77.0
12.5
GCC (Ours)
71.6
78.1
11.2
Table 6: Comparison of different token-injection strategies on Sana-1.6B, including linear interpolation, discrete binning, Fourier-feature embedding, and Gaussian-weighted Cluster Centroids (GCC), evaluated using GenEval, DPG, and FID.
Appendix figures & tables27 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Visualization comparison of images conditioned on watermark guidance scale ωwat . From left to right, the columns correspond to ωwat=0 , 3.5 , 7 with swat-low=0.05 and swat-high=0.95 .
Figure 8: Visualization comparison of images conditioned on aesthetic score saes and aesthetic guidance scale ωaes . From left to right, the columns correspond to saes=3 , 5 , 7 , and ωaes=1.5 with saes-low=5 and saes-high=7 .
Figure 9: Visualization comparison of images conditioned on clarity score scla . From left to right, the columns correspond to scla=200 , 500 , 1500 , 2500 , respectively.
Figure 10: Visualization comparison of images conditioned on entropy score sent . From left to right, the columns correspond to sent=5 , 6 , 7 , respectively.
Figure 11: Visualization comparison of images conditioned on luminance score sluma . From left to right, the columns correspond to sluma=0.3 , 0.4 , 0.5 , respectively.
Figure 12: More visualization comparison of LACON-S and LACON-A against Baselines, demonstrating our LACON can still achieve superior visual generation quality even when trained on the full set of unfiltered raw images.
Figure 13: More visualization comparison of LACON-S and LACON-A against Baselines, demonstrating our LACON can still achieve superior visual generation quality even when trained on the full set of unfiltered raw images.
Figure 14: More visualization comparison of LACON-S and LACON-A against Baselines, demonstrating our LACON can still achieve superior visual generation quality even when trained on the full set of unfiltered raw images.
Figure 15: Aesthetic score distribution of the raw data. Each bar indicates the proportion of samples whose aesthetic score saes falls within the corresponding score bin.
Figure 16: Watermark score distribution of the raw data. Each bar indicates the proportion of samples whose watermark score swat falls within the corresponding score bin.
Figure 17: Clarity score distribution of the raw data. Each bar indicates the proportion of samples whose clarity score scla falls within the corresponding score bin.
Figure 18: Entropy score distribution of the raw data. Each bar indicates the proportion of samples whose entropy score sent falls within the corresponding score bin.
Figure 19: Luminance score distribution of the raw data. Each bar indicates the proportion of samples whose luminance score sluma falls within the corresponding score bin.
Model
Overall ↑
Objects
Counting
Position
Colors
Color Attribution
Single
Two
Ratio-5
58.3
92.2
55.8
47.8
43.8
75.5
34.8
Ratio-30
62.1
95.6
65.2
50.3
42.5
78.2
41.3
Ratio-50
64.5
97.8
69.7
44.1
50.3
77.7
47.5
Ratio-65
68.0
97.8
72.7
48.1
57.0
82.2
50.0
Ratio-80
67.5
98.1
73.2
53.4
53.0
80.9
46.5
Appendix
Table 7: Quantitative comparison on GenEval with details for Sana-1.6B trained on subsets obtained under different quality-signal filtering thresholds . The results reveal a clear trade-off between data quantity and data quality in terms of model performance. We highlight the best and second-best entries.
Model
Overall ↑
Entity
Relation
Global
Attribute
Other
Ratio-5
73.4
81.9
91.1
81.8
85.3
69.2
Ratio-30
73.6
82.5
91.0
81.5
85.6
67.2
Ratio-50
74.5
82.9
91.7
81.8
85.9
64.4
Ratio-65
76.1
84.7
91.7
82.1
85.7
70.8
Appendix
Table 8: Quantitative comparison on DPG with details for Sana-1.6B trained on subsets obtained under different quality-signal filtering thresholds . The results reveal a clear trade-off between data quantity and data quality in terms of model performance. We highlight the best and second-best entries.
Model
Overall ↑
Objects
Counting
Position
Colors
Color Attribution
Single
Two
Baseline-A
55.0
92.8
53.3
42.8
37.0
70.7
33.3
Baseline-B
62.8
98.8
70.5
44.4
42.3
82.2
38.8
Baseline-C
60.8
96.6
64.7
45.0
43.8
80.3
34.3
Appendix
Table 9: Quantitative comparison between LACON and baseline methods for Sana-0.6B on GenEval with details . Training on the full unfiltered dataset with explicit conditioning enables LACON to outperform both filter-first baselines and unconditioned baselines. We highlight the best and second-best entries.
Model
Overall ↑
Entity
Relation
Global
Attribute
Other
Baseline-A
68.1
79.0
89.5
77.2
84.8
59.6
Baseline-B
70.4
80.0
90.0
83.6
85.2
62.8
Baseline-C
69.7
79.8
90.4
80.9
84.9
62.4
Baseline-D
69.2
79.4
89.7
81.8
84.2
60.8
Appendix
Table 10: Quantitative comparison between LACON and baseline methods for Sana-0.6B on DPG with details . Training on the full unfiltered dataset with explicit conditioning enables LACON to outperform both filter-first baselines and unconditioned baselines. We highlight the best and second-best entries.
Model
Overall ↑
Objects
Counting
Position
Colors
Color Attribution
Single
Two
Baseline-A
58.3
92.2
55.8
47.8
43.8
75.5
34.8
Baseline-B
68.0
97.8
72.7
48.1
57.0
82.2
50.0
Baseline-C
67.0
96.9
71.2
50.3
55.3
81.4
46.8
Appendix
Table 11: Quantitative comparison between LACON and baseline methods for Sana-1.6B on GenEval with details . Training on the full unfiltered dataset with explicit conditioning enables LACON to outperform both filter-first baselines and unconditioned baselines. We highlight the best and second-best entries.
Model
Overall ↑
Entity
Relation
Global
Attribute
Other
Baseline-A
73.4
81.9
91.1
81.8
85.3
69.2
Baseline-B
76.1
84.7
91.7
82.1
85.7
70.8
Baseline-C
75.4
83.8
92.0
82.1
85.6
68.0
Baseline-D
75.5
84.1
91.2
80.9
85.3
72.8
Appendix
Table 12: Quantitative comparison between LACON and baseline methods for Sana-1.6B on DPG with details . Training on the full unfiltered dataset with explicit conditioning enables LACON to outperform both filter-first baselines and unconditioned baselines. We highlight the best and second-best entries.
Model
Overall ↑
Objects
Counting
Position
Colors
Color Attribution
Single
Two
Baseline-B
68.3
97.5
77.5
54.1
54.5
82.5
43.5
LACON-S (Ours)
70.9
98.8
76.3
54.1
59.5
85.1
52.0
Appendix
Table 13: Quantitative comparison between LACON and the strongest baseline for Sana-1.6B at 1024 × 1024 resolution on GenEval with details . We highlight the best and second-best entries.
Model
Overall ↑
Entity
Relation
Global
Attribute
Other
Baseline-B
76.3
84.9
91.5
80.9
86.6
70.0
LACON-S (Ours)
77.6
86.2
92.4
82.1
88.0
74.8
LACON-A (Ours)
78.8
85.9
93.0
82.1
86.7
80.0
Appendix
Table 14: Quantitative comparison between LACON and the strongest baseline for Sana-1.6B at 1024 × 1024 resolution on DPG with details . We highlight the best and second-best entries.
Model
Overall ↑
Objects
Counting
Position
Colors
Color Attribution
Single
Two
Baseline-B
58.9
99.4
73.5
26.3
46.5
77.9
30.0
LACON-S (Ours)
61.3
98.8
70.7
39.1
51.5
75.0
32.5
Appendix
Table 15: Quantitative comparison between LACON and the strongest baseline for Qwen3-0.6B on GenEval with details , indicating that LACON generalizes well to autoregressive architecture. We highlight the best and second-best entries.
Model
Overall ↑
Entity
Relation
Global
Attribute
Other
Baseline-B
73.1
82.7
91.4
79.3
84.0
69.6
LACON-S (Ours)
74.9
84.1
92.6
80.9
84.6
71.6
LACON-A (Ours)
75.4
84.3
91.0
79.3
86.0
73.2
Appendix
Table 16: Quantitative comparison between LACON and the strongest baseline for Qwen3-0.6B on DPG with details , indicating that LACON generalizes well to autoregressive architecture. We highlight the best and second-best entries.
Model
Overall ↑
Objects
Counting
Position
Colors
Color Attribution
Single
Two
Baseline-B
66.4
98.1
82.8
36.3
59.8
76.9
44.5
LACON-S (Ours)
69.7
99.1
88.9
33.1
60.8
86.2
50.5
Appendix
Table 17: Quantitative comparison between LACON and the strongest baseline for Qwen3-1.7B on GenEval with details , indicating that LACON generalizes well to autoregressive architecture. We highlight the best and second-best entries.
Model
Overall ↑
Entity
Relation
Global
Attribute
Other
Baseline-B
78.2
86.3
92.6
81.5
85.1
72.0
LACON-S (Ours)
79.4
86.9
92.1
81.4
86.7
74.8
LACON-A (Ours)
80.1
87.0
92.4
80.2
86.7
72.8
Appendix
Table 18: Quantitative comparison between LACON and the strongest baseline for Qwen3-1.7B on DPG with details , indicating that LACON generalizes well to autoregressive architecture. We highlight the best and second-best entries.
Model
Overall ↑
Objects
Counting
Position
Colors
Color Attribution
Single
Two
Linear Interpolation
69.5
97.8
73.5
57.2
54.3
84.3
50.3
Discrete Binning
69.8
98.1
74.5
56.3
59.8
83.2
50.3
Fourier Feature
70.4
99.1
77.3
57.8
53.3
85.9
47.0
Appendix
Table 19: Quantitative comparison of different token-injection strategies for Sana-1.6B on GenEval with details . The results indicate that our GCC token-injection strategy can achieve better performance on most subcategories. We highlight the best and second-best entries.
Model
Overall ↑
Entity
Relation
Global
Attribute
Other
Linear Interpolation
76.1
85.0
92.5
80.2
86.4
71.6
Discrete Binning
77.2
84.8
92.6
81.2
86.2
72.4
Fourier Feature
77.0
84.1
91.7
82.1
86.3
74.0
GCC (Ours)
78.1
86.1
92.6
82.1
87.9
76.0
Appendix
Table 20: Quantitative comparison of different token-injection strategies for Sana-1.6B on DPG with details . The results indicate that our GCC token-injection strategy can achieve better performance on most subcategories. We highlight the best and second-best entries.
Text-to-image models trained on large-scale data often inevitably ingest unsafe content. While some people observe input-output amplifications, it remains unclear whether and how training data composition directly drives model output safety or by other factors. We shed light on this question by isolating this variable: we train the same text-to-image model on datasets that differ only in their fraction of unsafe images (0% to 9.6%), across several dataset scales (100K to 8M). Then we generate images with the resulting models, and evaluate them with four independent safety classifiers. Output unsafety rises monotonically from 16.6% at 0% contamination to 25.5% at 5%. A factorial design reveals that the proportion, not the absolute count, of unsafe training images is the operative variable. The 16.6% irreducible baseline at zero contamination implicates the other components, e.g. frozen text encoder, as a residual safety risk -- confirmed by a text encoder ablation showing that SafeCLIP reduces this floor to 9.6%, while the dose-response effect persists across all three encoders tested. Critically, no quality degradation in terms of FID, CLIPscore and ImageReward accompanies safety filtering. These results establish that data curation and text encoder safety are complementary and independently effective interventions. At the same time, the remaining level of unsafety poses questions for future research about emerging capabilities and compositionality.
Recent text-to-image generation models have achieved remarkable visual quality, but improving them through post-training remains challenging because no single reward signal captures the full range of human preference. In this work, we develop a simple and effective post-training recipe for open-domain text-to-image generation based on the composition of complementary reward signals. Our reward system consists of two main components: a preference reward, trained on large-scale human preference data using a Bradley-Terry objective to capture overall human aesthetic and perceptual preferences, and rubric-based rewards, which explicitly evaluate prompt faithfulness and other desirable properties while providing safeguards against reward hacking. A key challenge is how to combine these heterogeneous reward signals. We show that a naive weighted average leads to suboptimal optimization behavior, and propose a simple reward composition strategy that more effectively balances preference optimization with rubric satisfaction. In the Arena text-to-image leaderboard (https://arena.ai/), our RL-trained Flux2dev achieves an Elo rating 69 points above the base model, and our post-trained Ideogram-4 surpasses every open-source model on the leaderboard, reaching an Elo of 1223.5. (Claims of state-of-the-art performance are based on the Arena leaderboard snapshot as of September 4, 2026.) Our results suggest that effective rewards for frontier generative-model training require broad coverage of user intent and robustness to exploitation under optimization. To support reproducible research, we release Arena-T2I-Training, a 1K subset of training data that recovers some gains of full-scale training, providing a resource that we hope will facilitate future work on post-training for text-to-image models.
Post-training is crucial for improving text-to-image models, yet existing pipelines typically rely on manually curated training data and fixed training distributions. We introduce APT (Agentic Post-Training), a fully automated framework that adaptively post-trains a base model according to its evolving capability profile. APT coordinates three agents to construct evaluation sets from user requirements, diagnose model-specific strengths and failure modes, and generate reinforcement learning data with an adaptive prompt distribution that emphasizes underperforming capabilities. By repeatedly evaluating, diagnosing, and updating the generator, APT lets the current model's behavior shape its next training tasks, forming a feedback loop inspired by recursive self-improvement. For general-purpose settings without user-specified evaluation scenarios, we further develop an evaluation harness that enables broad and diverse evaluation set generation. Experiments across multiple base models, benchmarks, and generation architectures demonstrate that APT is model-agnostic, benchmark-agnostic, architecture-agnostic, and fully automated.
Sirui Zhang, Fangyun Wei, Dong Chen +5
University of Science and Technology of China · Microsoft Research Asia · University of Sydney +2