Research map / Image generation / Fast diffusion sampling
DogFit: Domain-guided Fine-tuning for Efficient Transfer Learning of Diffusion Models
Yara Bahram, Mohammadhadi Shateri, Eric Granger
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 domain-aware guidance offset to the fine-tuning loss (DogFit) | DiT-XL/2 fine-tuning · Six target domains avg | FID: 19.14 -> 12.34 | Improved |
| use DogFit instead of CFG sampling guidance | DiT-XL/2 · Six target domains avg | FID: 14.46 -> 12.34, 732.28 -> 366.14 TFLOPS | Improved |
| use DogFit instead of MG training-time guidance on SiT | SiT-XL/2 · Six target domains avg | FID: 10.36 -> 10.91 | Hurt |
| use DogFit instead of MG on SiT | SiT-XL/2 · Six target domains avg | FDDINOV2: 288.70 -> 252.29 | Improved |
| add guidance-strength conditioning (DogFit + Control) (ablation) | DogFit DiT-XL/2 · Six target domains avg | FID: 12.34 -> 12.63 | No clear effect |
| add DogFit to unlabeled FFHQ instead of CFG or MG | DiT-XL/2 · FFHQ | FID: 15.94 -> 10.48 | Improved |
| use DogFit instead of MG with fewer sampling steps (10) (ablation) | DiT-XL/2 · ArtBench | FID: 108.26 -> 88.84 | Improved |
| use DogFit instead of MG at 100 sampling steps (ablation) | DiT-XL/2 · Food | FID: 7.60 -> 8.09 | Hurt |
| combine DogFit with DiffFit PEFT (ablation) | DogFit DiT-XL/2 · Food and ArtBench | FID: Food 10.64 -> 11.86; Art 16.32 -> 15.80 | Mixed |
| sample guidance w with excessive exposure to high values (large training exposure) (ablation) | DogFit controllable guidance · Food and ArtBench | FID | Hurt |
| delay guidance start (late-start) moderately (ablation) | DogFit · DiT-XL/2 target domains | FID | Improved |
| set late-start threshold too high (ablation) | DogFit · DiT-XL/2 target domains | Precision | Hurt |
Closest papers by meaning
- Domain Guidance: A Simple Transfer Approach for a Pre-trained Diffusion Model (ICLR 2025 · similarity 0.91)
- Transfer Learning for Diffusion Models (NeurIPS 2024 · similarity 0.91)
- Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration (NeurIPS 2025 · similarity 0.89)
- Plug-and-Play Diffusion Distillation (CVPR 2024 · similarity 0.89)
- Diffusion Tuning: Transferring Diffusion Models via Chain of Forgetting (NeurIPS 2024 · similarity 0.88)
- Learning Diffusion Models with Flexible Representation Guidance (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.
