OpenDFM/Qwen2.5-VL-7B-IEA

Overview

OpenDFM/Qwen2.5-VL-7B-IEA is a 7 billion parameter conversational Image Editing Agent (IEA) built on the Qwen2.5-VL architecture. It addresses the gap between amateur users' intent and outcomes in image editing by providing an interpretable, tool-centric approach. Unlike generative models that may produce artifacts or lack transparency, IEA learns to operate 16 parameterized editing tools step-by-step, generating transparent edit traces that can be inspected and debugged.

Key Capabilities

• Conversational Image Editing: Interacts with users to understand and execute image editing instructions.
• Interpretable Action Space: Manipulates images using 16 explicit, parameterized tools, allowing for transparent and debuggable edit traces.
• Multitask Alignment: Trained via a three-stage pipeline involving supervised fine-tuning (SFT) on expert edits, Generative Reinforcement Learning from Human Feedback (GRPO) for likeness improvement and tool usefulness, and large-scale synthetic fine-tuning for joint mastery of editing, refinement, and intent summarization.
• High Instruction Following: Ranks best among tool-calling methods in user studies for instruction following.
• Superior Perceptual Quality: Surpasses generative methods in overall perceptual quality in user studies.
• Quantitative Performance: Achieves lower pixel distance on edit tasks and higher ROUGE-L on summary tasks compared to strong baselines.

Good for

• Amateur users who need precise, human instruction-guided image retouching.
• Applications requiring transparent and debuggable image editing processes.
• Developing conversational agents for visual content manipulation.
• Research into interpretable AI and tool-calling vision-language models.