Research map / Image generation / Autoregressive and latent generation
UniGP: Taming Diffusion Transformer for Prior-Preserved Unified Generation and Perception
Qin Guo, Hao Luo, Dongxu Yue, Weixuan Jin, Xiao Fu, Fan Wang, Dan Xu
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 perception training to controllable generation (ablation) | UniGP · Depth-to-image | RMSE to condition: 10.73 -> 6.89 | Improved |
| add generation training to perception (ablation) | UniGP · NYUv2 depth | AbsRel: 5.5 -> 5.2 | Improved |
| add generation training to perception (ablation) | UniGP · NYUv2 normal | mean angular error: 17.1 -> 16.4 | Improved |
| use binary-weighted unified training strategy (ablation) | UniGP · COCO-5K text-to-image | FID: 25.59 -> 21.05 | Improved |
| use full data instead of 50% generation data (ablation) | UniGP · Depth-to-image | RMSE to condition: 8.40 -> 6.89 | Improved |
| use full perception data instead of 50% (ablation) | UniGP · NYUv2 depth | AbsRel: 5.9 -> 5.2 | Improved |
| copy only the image branch (DUGP) instead of duplicating backbone (ablation) | UniGP · Depth-to-image | RMSE to condition: 8.99 -> 6.89 | Improved |
| freeze backbone and copy branch instead of fine-tuning backbone (Marigold-style) (ablation) | UniGP · COCO-5K text-to-image | FID: 25.71 -> 21.05 | Improved |
| balance control and perception layers at (12,12) instead of (6,18) (ablation) | UniGP · Depth-to-image | RMSE to condition: 9.21 -> 6.89 | Improved |
| balance layers at (12,12) instead of (18,6) (ablation) | UniGP · NYUv2 depth | AbsRel: 5.9 -> 5.2 | Improved |
| duplicate image branch | UniGP · Inference cost | 20-step inference time: 5.07s -> 9.54s | Hurt |
| use UniGP versus DepthAnything on NYUv2 depth | UniGP · NYUv2 depth | AbsRel: 4.3 -> 5.2 | Hurt |
Closest papers by meaning
- A Simple Approach to Unifying Diffusion-based Conditional Generation (ICLR 2025 · similarity 0.89)
- UniCon: Unidirectional Information Flow for Effective Control of Large-Scale Diffusion Models (ICLR 2025 · similarity 0.88)
- What Matters When Repurposing Diffusion Models for General Dense Perception Tasks? (ICLR 2025 · similarity 0.88)
- Scaling Properties of Diffusion Models For Perceptual Tasks (CVPR 2025 · similarity 0.88)
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.88)
- Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models (ICLR 2025 · similarity 0.87)
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.
