Generating Images of Rare Concepts Using Pre-trained Diffusion Models
Authors: Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, Gal Chechik
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We assess the faithfulness, quality and diversity of Seed Select in creating rare objects and generating complex formations like hand images, and find it consistently achieves superior performance. We further show the advantage of Seed Select in semantic data augmentation. Generating semantically appropriate images can successfully improve performance in few-shot recognition benchmarks, for classes from the head and from the tail of the training data of diffusion models. |
| Researcher Affiliation | Collaboration | 1Bar-Ilan University, Ramat-Gan, Israel 2Origin AI, Tel-Aviv, Israel 3NVIDIA Research, Tel-Aviv, Israel |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper states: 'To ensure a fair comparison, we replicated the methods mentioned above with Stable Diffusion v2.1 using the code published by the respective authors.' This refers to code from other authors/papers, not explicit release of their own methodology's code. |
| Open Datasets | Yes | We evaluated Seed Select on three common benchmarks: (1) Image Net (Deng et al. 2009):, the cannonical dataset with 1000 classes. ... (2) CUB (Wah et al. 2011): A fine-grained dataset... (3) i Naturalist (Van Horn et al. 2018): A large-scale, fine-grained dataset... |
| Dataset Splits | Yes | For each class, the set of reference images for all methods was taken from the trainset. Specifically, for Image Net, we used (Tu et al. 2022), achieving 88.2% accuracy on the corresponding test set. For CUB, we used (Chou, Kao, and Lin 2023), which attains 93.1% test accuracy. For i Naturalist, we used (Ryali et al. 2023), which has 83.8% test accuracy. |
| Hardware Specification | Yes | Typically, finding an optimal z G T takes between 1-4 minutes on an NVIDIA A100 GPU. |
| Software Dependencies | No | The paper mentions 'Stable Diffusion v2.1 with a guidance scale of 7.5 and 7 denoising steps using Euler Discrete Scheduler', but does not list specific versions for broader software dependencies like Python, PyTorch/TensorFlow, or CUDA. |
| Experiment Setup | Yes | Implementation details: We use Stable Diffusion v2.1 with a guidance scale of 7.5 and 7 denoising steps using Euler Discrete Scheduler (Karras et al. 2022). Stopping criteria: We stop optimizing z G T when LT otal plateaus or its value increases for more than 3 iterations. |