Empowering a Multimodal Aesthetic Agent for Photographic Reference Grading

Reference-guided photographic color grading helps users realize their aesthetic intent by selectively drawing on the color and tonal characteristics of a reference photograph. Doing so requires understanding which characteristics users wish to adopt and translating these preferences into controllable edits. However, image-only reference methods align overall appearance without instruction-grounded selection, while pixel-space generative editors typically produce flattened images without editable parameters, limiting subsequent refinement. To address these challenges, we introduce JarvisAura, a multimodal aesthetic agent that formulates photographic color grading as structured decision-making in an editable parameter space. JarvisAura offers three key advantages: (1) an aesthetic agent paradigm that integrates visual attribute perception, strategy-level reasoning, and parameterized execution for interpretable and controllable photographic color grading; (2) a unified reasoning-to-execution pipeline that connects aesthetic planning with backend rendering and editable parameter refinement; and (3) a three-stage training strategy comprising Parameter-Grounded Continual Pre-Training for parameter–appearance grounding, Reference-Grounded Supervised Fine-Tuning for reference-guided reasoning and parameter decisions, and Reference-Aligned Group Relative Policy Optimization for rendered-outcome refinement with multidimensional rewards and dynamic data derivation. On Ref-Bench, JarvisAura achieves the best scores among methods with available results across all reported automatic metrics and a 74% human-rated task-success rate, exceeding GPT-Image-2.5 by 9 percentage points.