LACON: Training Text-to-Image Model from Uncurated Data

Zhiyang Liang; Ziyu Wan; Hongyu Liu; Dong Chen; Qiu Shen; Hao Zhu; Dongdong Chen

cs.CVMar 27, 2026

LACON: Training Text-to-Image Model from Uncurated Data

Authors: Zhiyang Liang, Ziyu Wan, Hongyu Liu, Dong Chen, Qiu Shen, Hao Zhu, Dongdong Chen

Organizations: Nanjing University · Microsoft Research · HKUST

Abstract

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 $s_{\text{aes}}$ ; (3) LACON conditioned on a high aesthetic score $s_{\text{aes}}$ ; (4) LACON jointly conditioned on high aesthetic score $s_{\text{aes}}$ and high clarity score $s_{\text{cla}}$ .

Figure 3: Overview of LACON . High-level training pipeline that repurposes quality signals as explicit attribute conditioning during training.

Metric ( $s_{\text{metric}}$ )	Meaning	Value Range	Clipping	Tokens ( $N$ )	Centroid Anchors ( $p_{i}^{\text{metric}}$ )
Aesthetic ( $s_{\text{aes}}$ )	Aesthetic quality	0–10	None	10	{0.5, 1.5, …, 9.5}
Watermark ( $s_{\text{wat}}$ )	Watermark probability	0–1	None	10	{0.05, 0.15, …, 0.95}
Clarity ( $s_{\text{cla}}$ )	Perceptual sharpness	0–1M+	$>3000$	10	{150, 450, …, 2850}
Entropy ( $s_{\text{ent}}$ )	Information density	0–8	None	8	{0.5, 1.5, …, 7.5}
Luminance ( $s_{\text{luma}}$ )	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 (

p_{i}

) 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 $s_{\mathrm{aes}}$ and HSV-luma score $s_{\mathrm{luma}}$ . From left to right, the columns correspond to $s_{\text{aes}}=3$ , $5$ , $7$ . From top to bottom, the rows correspond to $s_{\text{luma}}=0.3$ , $0.4$ , $0.5$ . (b) Comparison of images jointly conditioned on clarity lapvar score $s_{\mathrm{cla}}$ and entropy score $s_{\mathrm{ent}}$ . From left to right, the columns correspond to $s_{\text{cla}}=200$ , $1000$ , $2500$ . From top to bottom, the rows correspond to $s_{\text{ent}}=5$ , $6$ , $7$ .

Ratio	$s_{\texttt{aes-min}}$	$s_{\texttt{wat-min}}$	$s_{\texttt{cla-min}}$	$s_{\texttt{ent-min}}$	$s_{\texttt{luma-min}}$	$s_{\texttt{luma-max}}$
$\approx 5\%$	5.0	0.3	800.0	6.0	0.1	0.9
$\approx 30\%$	4.0	0.5	600.0	4.0	0.1	0.9
$\approx 50\%$	3.5	0.6	500.0	3.0	0.1	0.9
$\approx 65\%$	3.0	0.7	400.0	2.0	0.1	0.9
$\approx 80\%$	3.0	0.8	200.0	2.0	0.1	0.9

Table 2: The filtering thresholds for different filtering ratios.

Ratio	GenEval $\uparrow$	DPG $\uparrow$	FID $\downarrow$
$\approx 5\%$	58.3	73.4	14.0
$\approx 30\%$	62.1	73.6	13.7
$\approx 50\%$	64.5	74.5	13.4
$\approx 65\%$	68.0	76.1	13.1
$\approx 80\%$	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 $\times$ 512 resolution
Model	Sana-0.6B			Sana-1.6B
Metric	GenEval $\uparrow$	DPG $\uparrow$	FID $\downarrow$	GenEval $\uparrow$	DPG $\uparrow$	FID $\downarrow$
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

\times

512 and 1024

\times

1024 resolutions.

Model	Qwen3-0.6B			Qwen3-1.7B
Metric	GenEval $\uparrow$	DPG $\uparrow$	FID $\downarrow$	GenEval $\uparrow$	DPG $\uparrow$	FID $\downarrow$
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 $\uparrow$	DPG $\uparrow$	FID $\downarrow$
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 $\omega_{\text{wat}}$ . From left to right, the columns correspond to $\omega_{\text{wat}}=0$ , $3.5$ , $7$ with $s_{\text{wat-low}}=0.05$ and $s_{\text{wat-high}}=0.95$ .

Figure 8: Visualization comparison of images conditioned on aesthetic score $s_{\text{aes}}$ and aesthetic guidance scale $\omega_{\text{aes}}$ . From left to right, the columns correspond to $s_{\text{aes}}=3$ , $5$ , $7$ , and $\omega_{\text{aes}}=1.5$ with $s_{\text{aes-low}}=5$ and $s_{\text{aes-high}}=7$ .

Figure 9: Visualization comparison of images conditioned on clarity score $s_{\text{cla}}$ . From left to right, the columns correspond to $s_{\text{cla}}=200$ , $500$ , $1500$ , $2500$ , respectively.

Figure 10: Visualization comparison of images conditioned on entropy score $s_{\text{ent}}$ . From left to right, the columns correspond to $s_{\text{ent}}=5$ , $6$ , $7$ , respectively.

Figure 11: Visualization comparison of images conditioned on luminance score $s_{\text{luma}}$ . From left to right, the columns correspond to $s_{\text{luma}}=0.3$ , $0.4$ , $0.5$ , 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

s_{\text{aes}}