FlashFusion

Multi-Model Vision Fusion — Ensemble, Cascade, and Fuse Detection, Classification, Segmentation, and OCR Models

Documentation · Quick Start · Fusion Strategies · Pipelines

---

Overview

FlashFusion is a unified framework for combining multiple vision models into powerful multi-model pipelines. Leverage Weighted Box Fusion, Voting Ensembles, Cascading, Stacking, and Feature-Level Fusion to boost accuracy and robustness beyond any single model.

Key Features

• Multi-Model Ensembling — Combine detection, classification, segmentation, and OCR models
• Flexible Strategies — WBF, voting, cascade, stacking, NMS fusion out-of-the-box
• Pre-Built Pipelines — Det→Cls, Det+Seg, multi-task parallel execution
• FlashVision Integration — Native support for FlashDet, FlashCls, FlashSeg, FlashOCR
• ONNX Export — Export fused pipelines for edge deployment
• LoRA Fine-Tuning — Efficiently adapt fusion layers with minimal parameters
• Extensible Registry — Plug in custom strategies, backbones, and pipelines

Note: Fusion strategies (WBF, voting, cascade, stacking, NMS) are fully implemented and tested. Pipelines, training engine, and export require trained model weights (FlashVision checkpoints or custom .pt files). See CHANGELOG.md for details on implemented vs. planned features.

---

Installation

pip install flashfusion

With all FlashVision models:

pip install flashfusion[flash]

Full install (export + analytics + flash models):

pip install flashfusion[all]

From source:

git clone https://github.com/FlashVision/FlashFusion.git
cd FlashFusion
pip install -e ".[dev]"

---

Quick Start

Ensemble Detection Models

from flashfusion import FlashFusion
from flashfusion.strategies import WeightedBoxFusion

model = FlashFusion(
    models=["flashdet-m", "flashdet-l"],
    strategy=WeightedBoxFusion(weights=[0.4, 0.6]),
    input_size=(320, 320),
)

results = model.predict("image.jpg")

Detection → Classification Pipeline

from flashfusion.pipelines import DetClsPipeline

pipeline = DetClsPipeline(
    detector="flashdet-m",
    classifier="flashcls-m",
    det_threshold=0.5,
)

results = pipeline("image.jpg")
for obj in results:
    print(f"{obj.label}: {obj.confidence:.2f} at {obj.bbox}")

Cascade Pipeline with Early Exit

from flashfusion.strategies import CascadeFusion

cascade = CascadeFusion(
    models=["flashdet-s", "flashdet-m", "flashdet-l"],
    confidence_thresholds=[0.8, 0.6, 0.0],
)

results = cascade.predict("image.jpg")

---

Fusion Strategies

Strategy	Use Case	Description
Weighted Box Fusion	Detection	Merges overlapping boxes with learned weights
Voting	Classification	Majority/soft voting across models
Cascade	Any	Sequential models with early exit on high confidence
Stacking	Any	Meta-learner on top of model outputs
NMS Fusion	Detection	NMS-based deduplication across model outputs

---

Training

Train fusion layers on top of frozen base models:

flashfusion train --config configs/flashfusion_det_cls_320.yaml

---

Export

Export fused pipeline to ONNX:

flashfusion export --config configs/flashfusion_ensemble_320.yaml --format onnx

---

Project Structure

FlashFusion/
├── flashfusion/          # Core library
│   ├── models/           # Fusion model definitions
│   ├── strategies/       # Fusion strategies (WBF, voting, cascade, etc.)
│   ├── pipelines/        # Pre-built multi-model pipelines
│   ├── engine/           # Training, validation, prediction, export
│   ├── solutions/        # High-level solution APIs
│   └── analytics/        # Benchmarking and profiling
├── configs/              # YAML configuration files
├── examples/             # Runnable example scripts
├── tests/                # Unit tests
└── docs/                 # Documentation

---

Benchmarks

Pipeline	Models	mAP	Latency (ms)	vs Single Model
WBF Ensemble	FlashDet-M × 3	52.1	28.4	+3.2 mAP
Det→Cls	FlashDet-M + FlashCls-M	48.7	18.2	+2.1 mAP
Cascade	FlashDet-S/M/L	50.8	15.6	+2.8 mAP, -40% latency

---

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

---

License

FlashFusion is released under the MIT License.

---

Part of the FlashVision ecosystem