🧵10/10 Lastly, huge thanks to my co-advisors Niloy and Duygu!
For more details check out our paper below-
🌐 Project Website: monetgpt.github.io
📄 Arxiv: arxiv.org/abs/2505.06176
🧵4/10 🧩 Puzzle A builds an understanding of individual operations. The MLLM learns to map visual changes in before/after images to a specific tool and its precise parameter value, effectively learning the semantics of our procedural library.
🧵5/10 🧩 Puzzle B imparts aesthetic judgement. By ranking professionally edited photos against altered versions, the MLLM learns to recognize the visual characteristics of an optimally adjusted image for any given operation, building an internal aesthetic model.
🧵9/10 We quantitaively evaluate on the Adobe5k dataset as well as conduct user studies by expert and novice users. Our evaluations show that MonetGPT outperforms open-source alternatives and performs comparably to Google Photos AutoEnhance (closed-source).
🧵7/10 Our puzzle-based training with a 'reasoning as a pathway' approach allows MonetGPT to generate detailed justifications for each edit, delivering truly explainable image retouching
🧵3/10 Our key recipe: MLLMs struggle to predict edit values directly. We solve this by generating rich textual reasoning for each puzzle ✍️. We then fine-tune MonetGPT on this data, creating a 'reasoning pathway' that enables it to regress final adjustment values.
