The Apple research team has made a big move and officially launched the latest multimodal AI model UniGen1.5. This is not an ordinary model, it marks a big step forward in image processing technology. In the past, processing images relied on several different tools, but now UniGen1.5 can handle the three major functions of image understanding, generation, and editing with just one model. This efficiency improvement is not insignificant, isn't it still something to look forward to? Compared to traditional models that separately process images with different functions, UniGen1.5 is simply too convenient.
UniGen1.5 is not simple, as it adopts a unified framework that integrates the three functions of image understanding, generation, and editing together. Researchers say that this integrated design can have significant benefits. When generating images, the model can fully utilize its powerful image understanding ability, resulting in higher visual output quality. Just like a chef who understands the characteristics of ingredients and can cleverly mix them, the dishes they make will definitely be delicious.
In terms of image editing, UniGen1.5 has an innovative gameplay called "Edit Command Alignment" technology. This technology is quite interesting. It doesn't involve directly modifying images, but rather having the model generate detailed text descriptions based on the original image and instructions, capturing the user's editing intent first. It's like imagining the picture in your mind before drawing, and then starting to draw, so the picture you draw will definitely be more in line with your heart. This' think first, draw later 'approach greatly improves the model's understanding and execution accuracy for complex modification requests.
It is worth mentioning that the field of image editing has always been a challenge, and many models are prone to errors when processing complex editing requests. UniGen1.5's technology effectively solves this problem.
UniGen1.5 also has new breakthroughs in reinforcement learning. The research team has designed a unified reward system that can be used for both image generation and editing training simultaneously. Previously, inconsistent quality standards were a major issue in editing tasks, as different people had different requirements for editing effects, making it difficult to grasp the model. Now with this unified reward system, the model can maintain a high level of performance when dealing with various visual tasks.
In multiple industry standard tests, UniGen 1.5's performance is simply outstanding. In GenEval and DPG Bench tests, it achieved high scores of 0.89 and 86.83, respectively, leaving other popular models such as BAGEL and BLIP3o far behind.
In the specialized image editing test ImgEdit, its score is 4.31, not only surpassing the open-source model OminiGen2, but also performing similarly to the proprietary closed source model GPT-Image-1.
Although UniGen1.5 performs well, researchers are also aware that there is still room for improvement. For example, the model is prone to errors when generating text in images, sometimes resulting in either too many or too few words, or simply errors. In specific editing scenarios, the model may cause drift in the main features, such as deviations in animal hair texture and color. But it's okay, the Apple team is already planning to continue working hard to optimize these issues and make UniGen 1.5 more perfect.
The emergence of UniGen1.5 has brought new hope to the field of image processing. I believe that in the future, it will continue to evolve, bringing more surprises to developers and driving the entire industry forward.
