Research map / Image generation / Text-to-image diffusion
LACON: Training Text-to-Image Model from Uncurated Data
Zhiyang Liang, Ziyu Wan, Hongyu Liu, DONG CHEN, Qiu Shen, Hao Zhu, Dongdong Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add quality-attribute conditioning and train on the full 110M unfiltered data (LACON-A) vs filtered Baseline-B | Sana-1.6B · GenEval 512px | GenEval: 68.0 -> 71.6 | Improved |
| add quality-attribute conditioning (LACON-A) vs Baseline-B | Sana-0.6B · GenEval 512px | GenEval: 62.8 -> 65.6 | Improved |
| add quality conditioning vs unconditioned full-data training (Baseline-C) | Sana-1.6B · FID 512px | FID: 13.5 -> 11.2 | Improved |
| use negative-prompt guidance on full unfiltered data (Baseline-D vs Baseline-C) | Sana-1.6B · GenEval 512px | GenEval: 67.0 -> 67.1 | No clear effect |
| train on full data without quality conditioning (Baseline-C) vs 65% filtered (Baseline-B) | Sana-0.6B · FID 512px | FID: 13.4 -> 15.4 | Hurt |
| raise retained data ratio from 5% to 65% (ablation) | Sana-1.6B · GenEval 512px | GenEval: 58.3 -> 68.0 | Improved |
| raise retained data ratio from 65% to 100% without conditioning (ablation) | Sana-1.6B · GenEval 512px | GenEval: 68.0 -> 67.0 | Hurt |
| add aggressive per-attribute guidance (LACON-A over LACON-S) (ablation) | Sana-1.6B · GenEval 512px | GenEval: 70.9 -> 71.6 | Improved |
| LACON-A over LACON-S at 1024 resolution (ablation) | Sana-1.6B · DPG 1024px | DPG: 77.6 -> 78.8 | Improved |
| apply LACON to autoregressive Qwen3-1.7B | Qwen3-1.7B · GenEval | GenEval: 66.4 -> 70.3 | Improved |
| apply LACON to autoregressive Qwen3-0.6B | Qwen3-0.6B · FID | FID: 12.4 -> 11.2 | Improved |
| use GCC instead of linear interpolation tokens (ablation) | LACON-A · GenEval, Sana-1.6B | GenEval: 69.5 -> 71.6 | Improved |
Closest papers by meaning
- Data Attribution for Text-to-Image Models by Unlearning Synthesized Images (NeurIPS 2024 · similarity 0.86)
- Boosting Alignment for Post-Unlearning Text-to-Image Generative Models (NeurIPS 2024 · similarity 0.86)
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.86)
- SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025 · similarity 0.85)
- Return of Unconditional Generation: A Self-supervised Representation Generation Method (NeurIPS 2024 · similarity 0.85)
- Rethinking Training for De-biasing Text-to-Image Generation: Unlocking the Potential of Stable Diffusion (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
