Apple MGIE released as open-source AI image-editing tool
AI can be used for image editing as well as generation; however, this application of machine learning can struggle to match human instructions, which might be subject to technical, intentional or idealized cosmetic variation, with the actual goal and the corresponding output. However, Apple's new MGIE model is already regarded as capable of "revolutionizing" that technology.
It is credited with the improved 'interpretation' of instructions, complete with context such as 'realizing' that a prompt such as "change the background and add a Star Wars background" could entail the addition of "a lightsaber or a spaceship" thanks to the integration of MLLMs, thereby giving potentially superior results in qualtitative analysis and human evaluation compared to its rival InsPix2Pix or predecessor LLM-Guided Image Editing (LGIE).
It can also leverage its MLLM to 'reason' that a call to make a food picture "healthier" could involve its augmentation of some vegetables. MGIE is rated to do so at a "Photoshop" level, and also to carry out either spot or general "photo optimization" with pixel-level accuracy and precision.
This latest foray by Apple into AI research has been presented at the International Conference on Learning Representations 2024 (ICLR 2024) in collaboration with a team at the University of California Santa Barbara (UCSB), which has also published a paper based on the same work currently available on arXiv.
Buy an Apple MacBook Air M2 as a Renewed Premium model in Starlight on Amazon
Source(s)
arXiv via VentureBeat