LFNet: Liquid Fusion of Heterogeneous Representations Towards General Salient Object Detection

---

📢 News

• [2026.6.25] Training weights and prediction maps are now available on Baidu Netdisk.
• [2026.6.25] Code and paper will be released soon. Stay tuned!

---

📖 Introduction

The proposed LFNet is designed for General Salient Object Detection, covering five main downstream tasks:

• 🟢 RGB SOD
• 🔵 RGB-D SOD
• 🟠 RGB-T SOD
• 🟣 Video SOD
• 🟤 VDT SOD

The implementation smartly integrates VMamba and ConvNeXt features through a novel Liquid Fusion mechanism, and employs saliency-guided upsampling to achieve high-quality and robust saliency prediction across various heterogeneous scenarios.

---

⚙️ Framework

Figure 1: Overall architecture of the proposed LFNet.

---

👁️ Visual Examples

We provide qualitative comparisons of LFNet across all five SOD tasks.

1. RGB SOD Results (Click to expand)

2. RGB-D SOD Results (Click to expand)

3. RGB-T SOD Results (Click to expand)

4. Video SOD Results (Click to expand)

5. VDT SOD Results (Click to expand)

---

📥 Model Weights & Prediction Results

The trained checkpoints and prediction maps across all evaluated datasets are provided via Baidu Netdisk.

📦 Resource	📝 Description	🔗 Link	🔑 Extraction Code
ckpt	Trained model checkpoints	Baidu Disk	lsod
results.zip	Prediction maps / saliency results	Baidu Disk	lsod

---

🛠️ Environment / Configuration

For environment setup and configuration, please refer to Samba: https://github.com/Jia-hao999/Samba.

---