Time-Varying LoRA: Towards Effective Cross-Domain Fine-Tuning of Diffusion Models
Authors: Zhan Zhuang, Yulong Zhang, Xuehao Wang, Jiangang Lu, Ying Wei, Yu Zhang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on various benchmark datasets, empirically demonstrate the effectiveness of Terra. |
| Researcher Affiliation | Academia | Zhan Zhuang1,2, Yulong Zhang3, Xuehao Wang1 Jiangang Lu3 Ying Wei3, Yu Zhang1, 1Southern University of Science and Technology 2City University of Hong Kong 3Zhejiang University |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our source code is publicly available on https://github.com/zwebzone/terra. |
| Open Datasets | Yes | For the UDA experiments, we utilize three benchmark datasets, including Office31 [51], which consists of 4,110 images from 31 categories across three domains: Amazon (A), Webcam (W), and Dslr (D); Office-Home [59], containing 15,588 images from 65 categories across four domains: Art (Ar), Clipart (Cl), Product (Pr), and Real-World (Rw); and Vis DA [43], featuring 207,785 images from 12 categories across two domains: Synthetic and Real. For the DG experiments, we employ the PACS [33], Office-Home, and VLCS [9] datasets. |
| Dataset Splits | No | No explicit statement of dataset split percentages or sample counts for training, validation, and test sets is provided within the paper for their specific experiments, rather it references external protocols or discusses data generation. |
| Hardware Specification | Yes | All experiments are conducted on an NVIDIA A100 GPU with three random trials. |
| Software Dependencies | No | The text-to-image diffusion model used in this paper is the Stable Diffusion XL (SDXL) model [45]. |
| Experiment Setup | Yes | The default resolutions of the generated images are 1024 x 1024. The rank of Lo RA is set as 16 for generative interpolation tasks and 32 for generation-based UDA and DG tasks. |