Butterfly GAN Model
| Property | Value |
|---|---|
| Base Paper | Towards Faster and Stabilized GAN Training for High-fidelity Few-shot Image Synthesis |
| Training Data | 1000 unique butterfly images |
| Output Resolution | 512x512 |
| Training Hardware | 2x A4000 GPUs |
| Training Duration | ~1 day (good results in 7-12h) |
What is butterfly_cropped_uniq1K_512?
This is a specialized GAN model designed to generate high-quality butterfly images, trained on a carefully curated dataset of 1000 unique butterfly species. The model implements the Light-GAN architecture, known for its efficiency in few-shot image synthesis, and can generate detailed 512x512 pixel images.
Implementation Details
The model utilizes a sophisticated training pipeline with specific image transformations including resizing, random horizontal flips, and normalization. It was trained using mixed precision (fp16) with a batch size of 64 and gradient accumulation of 4.
- Trained on carefully filtered dataset using CLIP scores
- Implements custom transforms from the official Light-GAN repository
- Optimized for generating diverse butterfly species
Core Capabilities
- Generates high-fidelity butterfly images at 512x512 resolution
- Supports batch generation with configurable latent dimensions
- Produces diverse outputs across different butterfly species
- Achieves stable training with limited data samples
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its ability to generate high-quality butterfly images using a relatively small dataset of 1000 images, while maintaining species diversity through careful dataset curation and CLIP-based filtering.
Q: What are the recommended use cases?
The model is intended for fun and learning purposes, particularly useful for generating diverse butterfly images for artistic or educational applications. It's especially valuable for scenarios requiring high-quality butterfly image synthesis without access to large datasets.
