FLORA: Efficient Synthetic Data Generation for Object Detection in Low-Data Regimes via finetuning Flux LoRA

Alvaro Patricio, Atabak Dehban, Rodrigo M. M. VenturaarXiv 2025 · DOI · arXiv

Abstract

FLORA (Flux LoRA Augmentation) is a lightweight pipeline designed to generate high-quality synthetic data for object detection tasks, particularly in low-data regimes. By leveraging the Flux 1.1 diffusion model and Low-Rank Adaptation (LoRA), FLORA prioritizes generation quality over volume. Experimental results demonstrate that FLORA outperforms state-of-the-art methods using 10x less synthetic data while remaining computationally accessible, running on consumer-grade hardware.

Introduction

Modern object detection models, such as YOLOv7, require extensive labeled datasets to achieve high performance. However, acquiring such data is often prohibitively expensive or difficult in specialized domains like medical imaging, underwater robotics, or precision agriculture. While generative AI has been proposed as a solution, existing "brute-force" approaches like ODGEN require massive computational resources and complex post-generation filtering to handle low-quality or "hallucinated" images. FLORA addresses these challenges by providing a high-fidelity, efficient synthetic data generation framework accessible to teams with limited budgets.

Methodology

FLORA utilizes a streamlined two-stage pipeline to produce domain-specific synthetic data.

Object-Centric LoRA Fine-Tuning

Instead of full-parameter fine-tuning, FLORA employs Low-Rank Adaptation (LoRA) to update only the attention layers of the Flux 1.1 Dev model. The model is trained on a minimal dataset—approximately 30 object crops per category—associated with unique "trigger words" (e.g., Aquarium-Shark). This approach allows the model to learn niche visual concepts with extreme efficiency.

LoRA-Conditioned Inpainting

The generation process involves "surgically" modifying objects within their original context to preserve background integrity and label accuracy.

The FLORA generation pipeline: from original image and bounding boxes to a new synthetic image via masked inpainting and a LoRA-tuned diffusion sampler.

As shown in the pipeline, a real image and its corresponding bounding box are used to create a mask. A Gaussian blur is applied to the mask edges to ensure seamless blending. The LoRA-tuned model then inpaints a new variation of the object into the masked region, ensuring that the ground-truth labels remain perfectly aligned with the synthetic content.

Experimental Results

The authors evaluated FLORA across six diverse datasets from the Roboflow 100 benchmark, including MRI scans, agricultural textures (Cotton), and underwater environments.

Detection Performance and Realism

FLORA consistently outperformed the ODGEN baseline on 5 out of 6 datasets. Notably, in the "Road Traffic" category, FLORA achieved a 62.3% mAP, representing a 21.3-point improvement over the no-augmentation baseline.

Comparison between real images and synthetic images generated by FLORA across various domains (Jellyfish, Cyclist, Cotton, UI).

Visual inspections and Fréchet Inception Distance (FID) metrics confirm the high realism of the generated images. In the MRI domain, FLORA achieved an FID of 27.60, significantly better than ODGEN’s 93.82, indicating superior perceptual similarity to real medical data.

Efficiency and Accessibility

A key advantage of FLORA is its efficiency. By focusing on quality over quantity, the model achieves superior results with only 500 synthetic images—10% of the volume required by previous methods. Furthermore, the transition to LoRA reduces hardware requirements from enterprise clusters to a single NVIDIA RTX 4090, with training completed in 9–12 hours.

Discussion and Conclusion

FLORA demonstrates that high-quality, contextually relevant synthetic data is more effective than large-scale, lower-precision generation. By leveraging Flux 1.1’s Flow Matching objective, the pipeline achieves high stylistic consistency without the need for resource-intensive post-generation filtering.

Future work may involve exploring more complex prompt engineering and a broader range of LoRA hyperparameters to further enhance the diversity and robustness of the generated data across even more specialized domains.

Summary updated

BibTeX
@misc{patricio2025flora,
  author = {Alvaro Patricio and Atabak Dehban and Rodrigo M. M. Ventura},
  title = {{FLORA: Efficient Synthetic Data Generation for Object Detection in Low-Data Regimes via finetuning Flux LoRA}},
  year = {2025},
  eprint = {2508.21712},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.21712},
}
Figure