Research map / Image generation / Text-to-image diffusion
Wasserstein-Aware Transfer: Class-Level Alignment for Robust Diffusion Model Adaptation
Zixian Huang, Chuan-Xian Ren
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 |
|---|---|---|---|
| use WAT (class matching + class-level retention sampling) instead of CFG after fine-tuning | DiT-XL/2-256 fine-tuning · Seven downstream datasets average | FID: 13.69 -> 10.73 | Improved |
| use WAT instead of Domain guidance | DiT-XL/2-256 fine-tuning · Seven downstream datasets average | FID: 11.55 -> 10.73 | Improved |
| use WAT instead of Domain guidance | DiT-XL/2-256 fine-tuning · Seven downstream datasets average | FDDINOv2: 292.62 -> 257.71 | Improved |
| use WAT instead of Domain guidance | DiT-XL/2-256 fine-tuning · ArtBench-10 (large domain shift) | FID: 16.76 -> 13.43 | Improved |
| use WAT instead of Domain guidance | DiT-XL/2-256 fine-tuning · CUB, Stanford Cars, Caltech-101 | FID: 3.52 -> 3.55 (CUB); 4.38 -> 4.43 (Cars); 23.05 -> 23.40 (Caltech) | Mixed |
| restrict fine-tuning timestep sampling to [0, 600] (ablation) | DiT fine-tuning with CFG · Seven downstream datasets | FID: Food -2.84%, SUN +5.52%, Caltech -4.82%, CUB +6.70%, Cars +5.38%, DF20 +13.54%, ArtBench +1.15% | Mixed |
| use pretrained model for noisy steps [600,1000] and fine-tuned model for the rest at sampling (ablation) | Hybrid sampling · Stanford Cars and Food-101 samples | qualitative generation | Improved |
| use WAT at fewer sampling steps (25) instead of CFG (ablation) | WAT sampling · SUN397 | FID: 24.34 -> 15.66 | Improved |
| use WAT at 100 sampling steps instead of DoG (ablation) | WAT sampling · CUB-200-2011 | FID: 3.35 -> 3.66 | Hurt |
| vary CLR scale w away from 0.5 (ablation) | WAT · Food, CUB, Car | FID | Hurt |
Closest papers by meaning
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 2025 · similarity 0.91)
- Transfer Learning for Diffusion Models (NeurIPS 2024 · similarity 0.90)
- Domain Guidance: A Simple Transfer Approach for a Pre-trained Diffusion Model (ICLR 2025 · similarity 0.89)
- The Emergence of Reproducibility and Consistency in Diffusion Models (ICML 2024 · similarity 0.89)
- Tracing the Roots: Leveraging Temporal Dynamics in Diffusion Trajectories for Origin Attribution (NeurIPS 2025 · similarity 0.89)
- Wasserstein Transfer Learning (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.
