JarvisArt/JarvisArt-1208
JarvisArt-1208 is a 7 billion parameter multi-modal large language model (MLLM) developed by researchers from Xiamen University, HKUST (Guangzhou), CUHK, Bytedance, NUS, and Tsinghua University. It functions as an intelligent agent for photo retouching, designed to understand user intent and coordinate over 200 tools in Adobe Lightroom. The model utilizes a two-stage training framework, including Chain-of-Thought supervised fine-tuning and Group Relative Policy Optimization for Retouching (GRPO-R), and is supported by the MMArt dataset (55K samples) and MMArt-Bench. It excels in pixel-level content fidelity for photo retouching, outperforming GPT-4o by 60% in this metric.
Loading preview...
Overview of JarvisArt-1208
JarvisArt-1208 is a 7 billion parameter multi-modal large language model (MLLM) specifically engineered as an intelligent agent for photo retouching. Developed by a collaborative team from Xiamen University, HKUST (Guangzhou), CUHK, Bytedance, NUS, and Tsinghua University, its primary goal is to enhance human artistic creativity by interpreting user intentions and orchestrating over 200 tools within Adobe Lightroom.
Key Capabilities
- Intelligent Photo Retouching: Understands complex user instructions for image manipulation.
- Tool Coordination: Capable of coordinating a wide array of tools in professional software like Adobe Lightroom.
- Advanced Training: Employs a novel two-stage training framework, combining Chain-of-Thought supervised fine-tuning for foundational reasoning with Group Relative Policy Optimization for Retouching (GRPO-R) to refine decision-making and tool proficiency.
- Superior Pixel-Level Fidelity: Demonstrates a 60% improvement over GPT-4o in pixel-level metrics for content fidelity, while maintaining comparable instruction-following capabilities.
- Multi-Granularity Retouching: Supports both global, scene-level adjustments and precise, region-specific refinements, allowing for intuitive edits via text prompts and bounding boxes.
Good for
- Automated Photo Editing: Ideal for tasks requiring intelligent, agent-driven photo retouching.
- Creative Workflows: Liberating artists from manual adjustments by automating complex editing processes.
- Research in MLLM Agents: Provides a strong baseline and framework for further research into multi-modal agents for creative applications.
- Applications requiring high pixel-level accuracy: Particularly useful where precise image manipulation and content fidelity are critical.