Research map / Image generation / Text-to-image diffusion
Att-Adapter: A Robust and Precise Domain-Specific Multi-Attributes T2I Diffusion Adapter via Conditional Variational Autoencoder
Wonwoong Cho, Yan-Ying Chen, Matthew Klenk, David I. Inouye, Yanxia Zhang
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 |
|---|---|---|---|
| train Att-Adapter on real unpaired multi-attribute data instead of using GAN-derived paired baselines | Att-Adapter · FFHQ latent control, single attribute | control range (CR): S-I-GANs 49.45 -> 67.87 | Improved |
| joint multi-attribute training for disentanglement | Att-Adapter · FFHQ latent control, multi attribute | DIS: 20.1% -> 32.3% | Improved |
| use continuous attribute embedding instead of discretized LoRA tokens | Att-Adapter · FFHQ absolute control | facial composition L1: 17.48 -> 15.57 | Improved |
| use continuous attribute embedding instead of discretized LoRA tokens on EVOX | Att-Adapter · EVOX size control | size accuracy: 67% -> 95% | Improved |
| add CVAE to the adapter (ablation) | Att-Adapter · FFHQ fixed attribute condition | ID similarity: 28.5% -> 6.2% | Improved |
| add CVAE (ablation) | Att-Adapter · FFHQ fixed attribute condition | control range (CR): 78.27 -> 78.59 | No clear effect |
| train naive adapter longer (20k to 200k) (ablation) | Att-Adapter w/o CVAE · FFHQ | CR and ID similarity: CR 47.04 -> 78.27; ID Sim 6.0% -> 28.5% | Mixed |
| use dropout 0.25 instead of CVAE for regularization (ablation) | naive Att-Adapter · FFHQ | ID similarity: CVAE 6.2% vs dropout 23.6% | Hurt |
| increase attribute strength lambda from 0.0 to 0.9 (ablation) | Att-Adapter · FFHQ | facial composition L1 and ChatGPT alignment: L1 69.43 -> 15.57; ChatGPT 97% -> 78% | Mixed |
| use non-isotropic instead of isotropic Gaussian prior (ablation) | Att-Adapter · FFHQ | facial composition L1 | No clear effect |
| increase latent dimension Z from 128 to 512 (ablation) | Att-Adapter · FFHQ | facial composition L1: 20.54 -> 17.42 | Improved |
| compare ITI-GEN (naive) against Att-Adapter for 20 attributes | ITI-GEN · FFHQ absolute control | facial composition L1: 15.57 -> 52.16 | Hurt |
Closest papers by meaning
- Continuous, Subject-Specific Attribute Control in T2I Models by Identifying Semantic Directions (CVPR 2025 · similarity 0.89)
- T2I-Adapter: Learning Adapters to Dig Out More Controllable Ability for Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.88)
- Orthogonal Adaptation for Modular Customization of Diffusion Models (CVPR 2024 · similarity 0.88)
- $\text{I}^2\text{AM}$: Interpreting Image-to-Image Latent Diffusion Models via Bi-Attribution Maps (ICLR 2025 · similarity 0.87)
- EasyRef: Omni-Generalized Group Image Reference for Diffusion Models via Multimodal LLM (ICML 2025 · similarity 0.87)
- Localizing and Editing Knowledge In Text-to-Image Generative Models (ICLR 2024 · 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.
