DiTFuse: Towards Unified Semantic and Controllable Image Fusion: A Diffusion Transformer Approach (Official Weights)

This repository provides the official pretrained weights for DiTFuse. The project code is available on GitHub:

👉 GitHub: https://github.com/Henry-Lee-real/DiTFuse

DiTFuse supports multiple fusion tasks—including infrared–visible fusion, multi-focus fusion, multi-exposure fusion, and instruction-driven controllable fusion / segmentation—all within a single unified model.

---

📌 Available Model Versions

🔹 V1 — Stronger Zero-Shot Generalization

• Designed with better zero-shot fusion capability.
• Performs robustly on unseen fusion scenarios.
• Recommended if your use case emphasizes cross-dataset generalization.

🔹 V2 — Full Capability Version (Paper Model)

• This is the main model used in the DiTFuse paper.

• Provides the most comprehensive capabilities:

  • Full instruction-following control
  • Joint fusion + segmentation
  • Better fidelity and controllability
  • Stronger alignment with text prompts

• Recommended for research reproduction, benchmarking, and controllable image fusion tasks.