Research map / Image generation / Autoregressive and latent generation
Dual-Path Condition Alignment for Diffusion Transformers
Changhao Peng, Yuqi Ye, Shuangjun Du, Wenxu Gao, Wei Gao
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 dual-path sampling to DDT (ablation) | DDT-L/2 · ImageNet 256x256 (L/2, 400K iters, no CFG) | FID: 14.9 -> 12.5 | Improved |
| add condition alignment on top of dual-path sampling (ablation) | DDT-L/2 · ImageNet 256x256 (L/2, 400K iters, no CFG) | FID: 12.5 -> 11.1 | Improved |
| resample both t and noise instead of only t (ablation) | DUPA-L/2 · ImageNet 256x256 (L/2, 400K iters, no CFG) | FID: 13.2 -> 11.1 | Improved |
| resample both t and noise instead of only noise (ablation) | DUPA-L/2 · ImageNet 256x256 (L/2, 400K iters, no CFG) | FID: 12.4 -> 11.1 | Improved |
| align at depth 8 rather than 4 or 16 (ablation) | DUPA-L/2 · ImageNet 256x256 (L/2, 400K iters, no CFG) | FID: 11.8 (depth 4), 12.1 (depth 16) vs 11.1 (depth 8) | Improved |
| cosine similarity instead of NT-Xent (ablation) | DUPA-L/2 · ImageNet 256x256 (L/2, 400K iters, no CFG) | FID: 11.6 -> 11.1 | Improved |
| vary lambda from 0.25 to 1 (ablation) | DUPA-L/2 · ImageNet 256x256 (L/2, 400K iters, no CFG) | FID: 11.2, 11.1, 11.1, 11.1 | No clear effect |
| increase noising times K from 2 to 4 (ablation) | DUPA-L/2 · ImageNet 256x256 (L/2, 400K iters, no CFG) | FID: 11.1 -> 10.7 | No clear effect |
| double batch size of DDT instead of using dual-path (ablation) | DDT-L/2 · ImageNet 256x256 (L/2, 400K iters, no CFG) | FID: 13.9 vs DUPA 11.1 | Hurt |
| replace SiT with DUPA across model sizes | SiT/DDT · ImageNet 256x256 (400K steps, no CFG) | FID: B/2 33.0 -> 25.2; L/2 18.8 -> 11.1; XL/2 17.2 -> 8.71 | Improved |
| self-alignment without external encoder vs REPA | DUPA-XL/2 · ImageNet 256x256 (with CFG) | FID: 1.46 (400 ep) vs REPA 1.42 (800 ep) | Mixed |
Closest papers by meaning
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.93)
- Representation Entanglement for Generation: Training Diffusion Transformers Is Much Easier Than You Think (NeurIPS 2025 · similarity 0.92)
- AHPA: Adaptive Hierarchical Prior Alignment for Diffusion Transformers (NeurIPS 2026 · similarity 0.91)
- REPA Works Until It Doesn’t: Early-Stopped, Holistic Alignment Supercharges Diffusion Training (NeurIPS 2025 · similarity 0.91)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.89)
- U-REPA: Aligning Diffusion U-Nets to ViTs (NeurIPS 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.
