Home

FlashFusion Documentation

Welcome to the FlashFusion documentation. FlashFusion is a multi-model vision fusion framework that enables ensemble, cascade, and multi-task pipelines for detection, classification, segmentation, and OCR models.

Overview

FlashFusion provides:

• Ensemble Fusion — Combine predictions from multiple detection models using Weighted Box Fusion, Voting, or NMS strategies
• Cascade Pipelines — Chain models sequentially for progressive refinement
• Multi-Task Fusion — Fuse detection + classification + segmentation in a unified pipeline
• Strategy Registry — Plug-and-play fusion strategies via the registry pattern
• LoRA/QLoRA — Efficient fine-tuning of fusion layers with frozen base models
• ONNX Export — Deploy fused models on edge devices

Quick Links

Page	Description
Installation	Install FlashFusion and dependencies
Quick Start	Get running in 5 minutes
Models	FlashFusion model architecture
Fusion Strategies	WBF, Voting, Cascade, Stacking, NMS
Pipelines	Pre-built multi-task pipelines
Training	Train fusion layers
FAQ	Frequently asked questions

Architecture

FlashFusion
├── models/         # FlashFusion model, LoRA adapters
├── strategies/     # Fusion strategies (WBF, Voting, Cascade, etc.)
├── pipelines/      # Pre-built multi-model pipelines
├── engine/         # Training, validation, prediction, export
├── solutions/      # High-level APIs (EnsembleDetector, MultiModelAnalyzer)
├── analytics/      # Benchmarking, profiling, plotting
├── data/           # Dataset, DataLoader, transforms
├── losses/         # Multi-task and consistency losses
├── utils/          # Checkpoints, logging, metrics, visualization
├── nn/             # Neural network building blocks
└── cfg/            # Configuration management

Citation

If you use FlashFusion in your research, please cite:

@software{flashfusion2024,
  title={FlashFusion: Multi-Model Vision Fusion},
  author={FlashVision},
  year={2024},
  url={https://github.com/FlashVision/FlashFusion}
}