To conclude, Gonzalo shares with us curious moment in this work that will make our readers smile. “At one point, I was training Marigold for different tasks, like infilling or segmentation. For segmentation, I was looking at the single step and multi-step predictions. And I was noticing how the single step predictions are generally blurrier. For example, the shelf was the color lilac. And with more steps, I was seeing the color lilac kind of slowly derailing into a different color. For me, that was like the realization that there is an accumulation of error. It's something weird that's happening here. I find it really funny because I didn't even know: is it my bug? Is it what is causing this drift? And then I went back to all other models. It turns out multi-step inference for these types of models does produce certain artifacts which do negatively impact the performance and with more iterations, they become more pronounced.” To find out how fine-tuning imageconditional diffusion models is easier than you think, visit Poster Session 1 today (Saturday) from 11:15 to 13:00 and Oral Session 4.1: 3D Computer Vision IV tomorrow (Sunday) from 10:15 - 11:15. 11 DAILY WACV Saturday Fine-Tuning Image-Conditional …
RkJQdWJsaXNoZXIy NTc3NzU=