PaRa: Personalizing Text-to-Image Diffusion via Parameter Rank Reduction
Authors: Shangyu Chen, Zizheng Pan, Jianfei Cai, Dinh Phung
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | With comprehensive experiments, we show that Pa Ra achieves great advantages over existing finetuning approaches on single/multisubject generation as well as single-image editing. Notably, compared to the prevailing fine-tuning technique Lo RA, Pa Ra achieves better parameter efficiency (2 fewer learnable parameters) and much better target image alignment. |
| Researcher Affiliation | Academia | Shangyu Chen & Zizheng Pan & Jianfei Cai & Dinh Phung Monash University Melbourne, Australia EMAIL |
| Pseudocode | No | The paper describes methods using mathematical formulations and descriptive text, but no explicit pseudocode or algorithm blocks are provided. |
| Open Source Code | No | The paper refers to third-party open-source projects like LoRA (cloneofsimo, 2022) and mentions public pre-trained Lo RA models from civitai.com, but does not provide specific access to the source code for the proposed Pa Ra methodology. |
| Open Datasets | Yes | We evaluated the effects of Pa Ra on customized single-subject generation, based on the Dreambooth dataset, where each label consists of five to six images. |
| Dataset Splits | No | The paper mentions using the Dreambooth dataset and generating images for evaluation, but it does not specify explicit training, validation, or test dataset splits in terms of percentages, counts, or predefined partitions for reproducibility. |
| Hardware Specification | Yes | All experiments are conducted on a single A100 GPU with 40GB of VRAM. |
| Software Dependencies | No | The paper mentions using SDXL1.0 and the DDIM sampler, but it does not specify version numbers for any underlying software libraries or programming languages (e.g., Python, PyTorch, CUDA). |
| Experiment Setup | Yes | All baselines were trained for 1000 steps with a batch size of 1, and Lo RA chose a rank of r = 16 and a scale of α = 2.2 as the best model for a fair comparison. In our experiments, we found that Pa Ra already achieved ideal results at around 200 steps. Therefore, we compare the results of Pa Ra at 200 steps with the baselines at 1000 steps. Also, we employ a rank boundary γ = 1/40 for Pa Ra. |