Different Strokes for Robot Folks

The latest from the machine learning research community in this week’s newsletter is Paint Transformer, a novel transformer-based framework that can generate paintings from natural images by predicting parameters of multiple strokes with a feed-forward network.

The traditional approach for neural painting involves reinforcement learning but even if it produces satisfactory paintings, it is computationally expensive. The new approach that is fully based on feed-forward networks achieves better painting performance than the former, it is cheaper to train, and most notably, it learned to paint without seeing any existing paintings, eliminating the need for matched pairs of photos and paintings(a.k.a self-training).

For more about how Paint Transformer works and why it matters, check this.