edit the tint, the temperature, color grading. You can change the saturation, the contrast, highlights, apply presets. Like 90 percent of everything you want to do in photography, you can do it with this neural network. Marcos is very much aware of the reasons why this was not done before. “I think in general,” he tells us, “communities focused on the upper bound on the complex models, exploring what is the best possible thing that we can get. But in our lab, we focus on the opposite, on the efficient models. So we start from the bottom and we try to add complexity. The rest of the world is from complex models, trying to distill them, to make them smaller and more efficient. So I guess this was only possible with that kind of mindset. And this is actually the second work in this direction. Our previous work InstructIR, which we presented at ECCV 2024, was also a feature in different media sources and was the initial step. It was the first model that allowed you to restore images using text.” Hence, this was the natural extension of the previous work, and both approaches are actually quite novel. This does not happen without challenges. Designing a very efficient neural network for small language models, diffusion models, is the key. “We don't tackle the problem using complex neural networks,” Marcos declares. “We design them tailor-made for these operations. It took some time, because when you try to have one model that can do all these things without increasing the complexity much, you need to run a lot of experiments and a lot of trial and error. But we finally got it! And we are very happy that at the first try, at ICCV, we got three strong accepts! And that is a very, very good indication of novelty!” Marcos thinks that the fundamental problems in computer vision, at least the ones that we tackle in lowlevel computer vision, are anything related to cameras and computational photography. PixTalk tackles the problem of deblurring. It tackles the problem of denoising, because we want to enhance the photos. But it also tackles well-known 16 DAILY ICCV Wednesday Poster Presentation
RkJQdWJsaXNoZXIy NTc3NzU=