Research map / Image generation / Text-to-image diffusion
Imbalance in Balance: Online Concept Balancing in Generation Models
Yukai Shi, Jiarong Ou, Rui Chen, Haotian Yang, Jiahao Wang, Xin Tao, Pengfei Wan, Di Zhang, Kun Gai
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 |
|---|---|---|---|
| balance the head and tail data (1:1) (ablation) | diffusion model · piano-submarine | success rate: 20% -> 56% | Improved |
| balance the head and tail data (1:1) (ablation) | diffusion model · volcano-twins | success rate: 20% -> 64% | Improved |
| replace diffusion loss with IMBA loss (from scratch) | 1B DiT · LC-Mis | VQA: 46.21% -> 62.89% | Improved |
| replace diffusion loss with IMBA loss (from scratch) | 1B DiT · Inert-CompBench | VQA: 44% -> 57% | Improved |
| replace diffusion loss with IMBA loss (from scratch) | 1B DiT · T2I-CompBench Spatial | score: 0.1966 -> 0.2518 | Improved |
| finetune pretrained model with IMBA loss | 1B DiT · LC-Mis | VQA: 46.21% -> 51.82% | Improved |
| train IMBA from scratch rather than finetune (ablation) | 1B DiT · LC-Mis | VQA: 51.82% -> 62.89% | Improved |
| use IMBA loss instead of A&E | 1B DiT · LC-Mis | CLIP score: 0.3198 -> 0.3121 | Hurt |
| use IMBA loss instead of frequency-based weight (ablation) | diffusion model · Head/tail subset | success rate: 49.3% -> 65.7% | Improved |
| use token-wise instead of sample-wise weight (ablation) | IMBA loss · piano-submarine | success rate: 64% -> 72% | Improved |
| add sample-wise weighting (ablation) | diffusion model · piano-submarine | success rate: 32% -> 64% | Improved |
| increase model size beyond 200M (ablation) | diffusion model · LC-Mis | success rate | No clear effect |
Closest papers by meaning
- Handling Imbalanced Pseudolabels for Vision-Language Models with Concept Alignment and Confusion-Aware Calibrated Margin (ICML 2025 · similarity 0.83)
- TTS-VAR: A Test-Time Scaling Framework for Visual Auto-Regressive Generation (NeurIPS 2025 · similarity 0.82)
- MetaMetrics: Calibrating Metrics for Generation Tasks Using Human Preferences (ICLR 2025 · similarity 0.82)
- IDEA-Bench: How Far are Generative Models from Professional Designing? (CVPR 2025 · similarity 0.82)
- Turning Internal Gap into Self-Improvement: Promoting the Generation-Understanding Unification in MLLMs (ICLR 2026 · similarity 0.81)
- WeGen: A Unified Model for Interactive Multimodal Generation as We Chat (CVPR 2025 · similarity 0.81)
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.
