خب میخوام یه چالش 90 روزه هر روز یه مقاله رو شروع کنم. مقاله ها سعی میشه از حوزه های مختلف یادگیری ماشین و همچنین معتبر باشن. فعلا قانون خاصی نداریم و مقاله ها هم در ادامه همین پست انتشار داده میشن.
خب برو که بریم...
#1 Imagic: Text-Based Real Image Editing With Diffusion Models

Imagic is a text-guided image editing framework that can operate on complex images. This method essentially has 3 consecutive steps:
1. Text embedding optimization: In this step the target text is passed through a text encoder to extract the target embedding e_tgt. Then, in order to match the the input image as closely as possible to the target embedding, generative diffusion model gets frozen and only target embeddings are optimized
2. Model fine-tuning: Next, text embeddings get frozen and the generative diffusion model is updated. This stage helps to close the gap between the image x and the optimized embedding e_opt.
3. Text embedding interpolation: Finally, Since the generative diffusion model was trained to fully recreate the image x at the optimized embedding e_opt, text embedding interpolation helps to apply the desired edit by advancing in the direction of the target text embedding e_tgt.
As a base network, they adopted Imagen and Stable Diffusion models.