Research map / Image generation / Text-to-image diffusion
DomainStudio: Fine-Tuning Diffusion Models for Domain-Driven Image Generation Using Limited Data
Jingyuan Zhu, Huimin Ma, Jiansheng Chen, Jian Yuan
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 pairwise similarity loss L_img (lambda2 0 -> 0.5) (ablation) | DDPM fine-tuning · FFHQ->Babies (10-shot) | Intra-LPIPS: 0.520 -> 0.572 | Improved |
| add pairwise similarity loss L_img (lambda2 0 -> 0.5) (ablation) | DDPM fine-tuning · FFHQ->Babies (10-shot) | FID: 89.94 -> 55.97 | Improved |
| raise lambda2 from 0.5 to 5.0 (ablation) | DomainStudio L_img · FFHQ->Babies (10-shot) | Intra-LPIPS: 0.572 -> 0.537 | Hurt |
| add high-frequency pairwise loss L_hf (lambda3 0 -> 0.5) (ablation) | DomainStudio · FFHQ->Babies (10-shot) | Intra-LPIPS: 0.572 -> 0.592 | Improved |
| add high-frequency pairwise loss L_hf (lambda3 0 -> 0.5) (ablation) | DomainStudio · FFHQ->Babies (10-shot) | FID: 55.97 -> 53.11 | Improved |
| add high-frequency MSE loss L_hfmse (lambda4 0 -> 0.05) (ablation) | DomainStudio · FFHQ->Babies (10-shot) | FID: 53.11 -> 33.26 | Improved |
| add high-frequency MSE loss L_hfmse (lambda4 0 -> 0.05) (ablation) | DomainStudio · FFHQ->Babies (10-shot) | Intra-LPIPS: 0.592 -> 0.599 | Improved |
| raise lambda4 from 0.05 to 0.5 (ablation) | DomainStudio L_hfmse · FFHQ->Babies (10-shot) | FID: 33.26 -> 53.69 | Hurt |
| use image-level pairwise similarity instead of diffusion features (Design A) (ablation) | pairwise similarity loss · FFHQ->Sunglasses (10-shot) | FID: 40.30 -> 37.92 | Improved |
| use image-level pairwise similarity instead of multi-step images (Design B) (ablation) | pairwise similarity loss · FFHQ->Sunglasses (10-shot) | FID: 58.28 -> 37.92 | Improved |
| DomainStudio instead of direct fine-tuning | DDPM · FFHQ->Babies (10-shot) | Intra-LPIPS: 0.513 -> 0.599 | Improved |
| DomainStudio instead of CDC | pre-trained generator · FFHQ->Babies (10-shot) | FID: 74.39 -> 33.26 | Improved |
Closest papers by meaning
- Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation (CVPR 2025 · similarity 0.90)
- Bridging Data Gaps in Diffusion Models with Adversarial Noise-Based Transfer Learning (ICML 2024 · similarity 0.89)
- When Model Knowledge meets Diffusion Model: Diffusion-assisted Data-free Image Synthesis with Alignment of Domain and Class (ICML 2025 · similarity 0.89)
- You Only Sample Once: Taming One-Step Text-to-Image Synthesis by Self-Cooperative Diffusion GANs (ICLR 2025 · similarity 0.89)
- ZoomLDM: Latent Diffusion Model for Multi-scale Image Generation (CVPR 2025 · similarity 0.89)
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 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.
