Research map / Image generation / Autoregressive and latent generation
A Hidden Semantic Bottleneck in Conditional Embeddings of Diffusion Transformers
Trung X. Pham, Kang Zhang, Ji Woo Hong, Chang D. Yoo
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 |
|---|---|---|---|
| prune low-magnitude tail dimensions (tau=0.01, ~39%) of the conditional embedding only at the first step (ablation) | REPA · ImageNet-1K class-conditional | FID: 7.1694 -> 7.1690 | No clear effect |
| prune tail dimensions (tau=0.01) during the last steps (ablation) | REPA · ImageNet-1K class-conditional | FID: 7.1694 -> 7.1598 | Improved |
| prune tail dimensions (tau=0.01) at every step (ablation) | REPA · ImageNet-1K class-conditional | FID: 7.1694 -> 7.2143 | Hurt |
| prune 66% of tail dimensions (tau=0.02) (ablation) | REPA · ImageNet-1K class-conditional | FID: 7.1694 -> 9.2202 | Hurt |
| prune 96% of tail dimensions (tau=0.05) (ablation) | REPA · ImageNet-1K class-conditional | FID: 7.1694 -> 56.2308 | Hurt |
| remove 8 high-magnitude head dimensions (ablation) | REPA · ImageNet-1K class-conditional | FID: 7.1694 -> 523.7637 | Hurt |
| prune tail dimensions (tau=0.01, every step) on LightningDiT (ablation) | LightningDiT · ImageNet-1K class-conditional | FID: 7.0802 -> 7.0130 FID; IS 169.8574 -> 166.0569 | Mixed |
| prune tail dimensions (tau=0.01) at t0 on Model-Guide (ablation) | Model-Guide · ImageNet-1K class-conditional | FID: 7.2478 -> 7.2466 | No clear effect |
| prune ~40% dimensions (tau=0.1) of conditional embedding (ablation) | X-MDPT · DeepFashion pose-guided generation | FID: 18.6372 -> 18.6692 | No clear effect |
Closest papers by meaning
- Rare Text Semantics Were Always There in Your Diffusion Transformer (NeurIPS 2025 · similarity 0.90)
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.89)
- Scaling Diffusion Transformers Efficiently via $\mu$P (NeurIPS 2025 · similarity 0.89)
- On improved Conditioning Mechanisms and Pre-training Strategies for Diffusion Models (NeurIPS 2024 · similarity 0.89)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.88)
- Dual Diffusion for Unified Image Generation and Understanding (CVPR 2025 · similarity 0.88)
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.
