Scholarlink

Few-shot learning is a promising solution to the longstanding problem of limited data in the emerging fields of Natural Language Processing (NLP) and computer vision tasks. Our proposed research utilizes generative models to augment datasets with very few samples to produce synthetic data. Using the generative pre-trained transformer 2 (GPT-2) language model, we established domain-specific prompts to generate synthetic text data about aircraft scenes. We then employed those prompts to fine-tune the Stable Diffusion model with LoRA (Low-Rank Adaptation) to generate realistic images. We assessed the quantitative evaluation of the images using CLIP (Contrastive Language–Image Pretraining) scores for textual–image alignment, Fréchet Inception Distance (FID) for distribution similarity, and Learned Perceptual Image Patch Similarity (LPIPS) for perceptual similarity. The best score for CLIP was 0.312 ± 0.021, the FID was 73.54, and the best LPIPS distance was 0.231, reflecting strong semantic alignment, moderate realism, and high perceptual similarity, notwithstanding the lightweight nature of our pipeline that produces highly good quality image synthesis and semantic validity. Our system is interpretable, modular, and fundamentally deployable in low resource operational environments, while synthesizing novel and domain-relevant data. This study takes initial steps to connect modern generative approaches to deployment.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Few-Shot Learning with Generative Augmentation: A Multimodal Pipeline Using GPT-2 and Diffusion Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Rohit Beeravalli,
  • Sujal Naduvinamani,
  • Anup Kolabal,
  • Manikantha Shivalli,
  • Sumaiya Pathan

摘要

Few-shot learning is a promising solution to the longstanding problem of limited data in the emerging fields of Natural Language Processing (NLP) and computer vision tasks. Our proposed research utilizes generative models to augment datasets with very few samples to produce synthetic data. Using the generative pre-trained transformer 2 (GPT-2) language model, we established domain-specific prompts to generate synthetic text data about aircraft scenes. We then employed those prompts to fine-tune the Stable Diffusion model with LoRA (Low-Rank Adaptation) to generate realistic images. We assessed the quantitative evaluation of the images using CLIP (Contrastive Language–Image Pretraining) scores for textual–image alignment, Fréchet Inception Distance (FID) for distribution similarity, and Learned Perceptual Image Patch Similarity (LPIPS) for perceptual similarity. The best score for CLIP was 0.312 ± 0.021, the FID was 73.54, and the best LPIPS distance was 0.231, reflecting strong semantic alignment, moderate realism, and high perceptual similarity, notwithstanding the lightweight nature of our pipeline that produces highly good quality image synthesis and semantic validity. Our system is interpretable, modular, and fundamentally deployable in low resource operational environments, while synthesizing novel and domain-relevant data. This study takes initial steps to connect modern generative approaches to deployment.