Mehran Aghabozorgi

How can we train generative models on 100 examples without pretraining on large datasets? Presenting Adaptive IMLE, which outperforms both GANs and diffusion models. Details at mehranagh20.github.io/AdaIMLE/. Work w/ @mehranag and @PengShichong. Come by our #ICML2023 poster Wed AM 1/2

I stand with graduate students from @SFU_CompSci / @sfu in securing a fair wage for their (TA) work. As an associate faculty at #SFU and a SFUFA member, today I am cancelling my graduate class, and honour the @TSSU picket line. Best of luck

Excited to share our work on completing partial 3D shapes with IMLE. Whereas the prior leading method based on GANs generates similar completions due to mode collapse, our method uses IMLE to avoid mode collapse and can generate diverse completions. arxiv.org/abs/2106.16237 1/2

There are many ways to colourize a grayscale image. If the goal is just to colour it one way, does it matter if the model can colour one way (unimodal) or many ways (multimodal)? Joint work w/ @PengShichong & @Moazeni_Alireza Paper & code: niopeng.github.io/CAM-Net/ Thread👇 1/n