An MLLM Agent That Controls 200+ Lightroom Tools via Natural Language

What it does JarvisArt is a multimodal large language model agent that interprets free-form user instructions—text, bounding boxes, or high-level scene goals—and translates them into sequences of edits inside Adobe Lightroom. It was built to mimic the reasoning of professional retouchers rather than blindly applying filters, using a two-stage training pipeline: Chain-of-Thought supervised fine-tuning followed by GRPO-R, a reinforcement-learning phase tailored specifically for retouching decisions. The project ships with the MMArt dataset (55K samples derived from PPR10K and other licensed sources), MMArt-Bench for evaluation, and an Agent-to-Lightroom Protocol that lets the model actually drive the software.

The interesting bit The unusual part is the granularity: the model handles everything from global color grading to local spot adjustments, and it reportedly outperforms GPT-4o on pixel-level content fidelity by 60% while keeping instruction-following competitive. That specificity—turning natural language into Lightroom Lua/XMP actions—is what separates it from a chatbot with a vignette slider.

Key highlights

• Accepted at NeurIPS 2025; training and evaluation code, model weights, and data construction scripts are all released.
• Claims a 60% improvement over GPT-4o in pixel-level fidelity metrics on its benchmark, with comparable instruction adherence.
• Supports multi-granularity editing: scene-level down to region-specific refinements using text and bounding-box inputs.
• Provides an Agent-to-Lightroom Protocol for direct integration with Adobe Lightroom, plus a Hugging Face demo for quick browser trials.
• The MMArt-PPR10K dataset is Apache 2.0 licensed and includes paired original/edited images with Lightroom Lua/XMP files.

Caveats

• The README notes that a full per-source breakdown for every subset in Figure 9 is not yet published in the repository, so full data provenance remains partially opaque.
• Performance claims are self-reported against the authors’ MMArt-Bench; independent verification on other retouching benchmarks is not discussed in the provided sources.

Verdict Worth a look if you research vision-language agents, computational photography, or Lightroom automation. Skip it if you need a generic image generator—this is strictly about parametric, tool-based photo retouching.