Week #6 #

Presentation: #

Unexpected bugs of web interface with unclear origin.
Solution: Unit tests to identify sources of problems.
Lack of suitable datasets for predictor of object position in the picture. This is necessary so that, for example, a light bulb does not end up on the floor.
Solution: Creating our own dataset using the transparent_backgrounds model. The model was used to get pictures of objects without background + their position from real pictures from the internet - just what we need to train our model.
Implementation of the gallery in the web interface.
Solution: Using a third-party library

In the final weeks of the course, our team focused on polishing the project and adding last features that could improve it.

Reworked the front-end with history (galleries), making the prompts more visual and intuitive, and giving it an overall more pleasing design
Fixed bugs found
Implemented a prototype of the object coordinates predicate model.
Research on models for sanity check was conducted. The model predicting the probability that the picture is generated by a neural network - AIDE - was selected.

For the following week, we intend to focus on: