Thank you
@_akhaliq
for sharing!
At
#ICCV2023
, we will present a light-field-inspired method for viewpoint control over generative image models!
Come visit our poster on Friday!
Here's how I would have looked like as a Chinese migrant living in California from 1894.
It's not perfect, but I really enjoyed exploring my heritage with generative image models!
(1/n)
I visited the
@twosigma
robotics lab yesterday to help them prepare for the
@nycfirst
Corporate Robotics Challenge. It’s exciting to see adults get involved with high-school robotics for the first time.
@_akhaliq
By choosing a flexible and geometry-free representation, our generator is able to create photo-realistic images (but at the expense of view consistency)!
We hope that our project can inspire more work in geometry-free representations for view synthesis!
@SvitozarTaras
@apoorvkh
@DaniLischinski
@Jimantha
@ElorHadar
However, we've realized that historical photo datasets will always have biases which are difficult to control, no matter the method. This is why we think it is so important to have proper documentation of these biases.
@SvitozarTaras
@apoorvkh
@DaniLischinski
@Jimantha
@ElorHadar
I'm glad that you asked! One method we tried to reduce bias is to use an identity similarity network as a loss to guide the transformations. These models are trained on more diverse data than historical photos
Have you programmed with turtle graphics before 🐢? Turtles were how I was introduced to CS at Stuy. I wonder if turtles have practical uses in modern computer graphics other than as educational tools
@4000carbs
@MITEECS
@MIT
@Cornell
Thanks! The most interesting class I've taken is probably Numerical Analysis. It's so useful for understanding where machine learning algorithms come from
@fredodurand
I've noticed that Adobe Stock often contains hundreds of photos from the same photo shoot. This might explain the lack of diversity in Firefly samples