Combining Semantic Guidance and Deep Reinforcement Learning For Generating Human Level Paintings

Intelli-Paint: Towards Developing Human-like Painting Agents

“…The final RL agent is trained for a total of 5M iterations with a batch size of 128. [36], (e) RL [12] and (f) Semantic-RL [26]. We observe that our approach results in more accurate depiction of the fine-grain features in the target image while using a low brushstroke count.…”

“…In recent years, there has been an increased focus on learning an unsupervised brushstroke decomposition without requiring access to dense human brushstroke annotations. For instance, recent works [6,12,13,22,26,31] use deep reinforcement learning and an adversarial training approach for learning an efficient brushstroke decomposition. Optimization-based methods [36] directly search for the optimal brushstroke parameters by performing gradient descent over a novel optimal-transport-based loss function.…”

“…where r gbp t is the guided backpropagation based focus reward from [26]. The hyperparameter µ is set to 10 for our experiments, while the human consistency penalty coefficients η, γ are raised from 10 ´4 to 0.1 in a linear schedule during the training process.…”

“…In this section, we provide extensive qualitative and quantitative results comparing our method with recent state of the art neural painting methods [12,21,26,36]. First, in Sec.…”

“…The quantitative values show that our method helps in significantly lowering the distance metrics between the painted canvas and the target image as compared to previous works. In particular, we note that for CUB-Birds dataset [30], our approach leads to a reduction of 47.1%, 25.6% 24.9% and 38.2% in the L pixel distance metric as compared to RL [12], Semantic-RL [26], Optim [36] and Paint Transformer [20], respectively. All painting sequences are generated using a different brushstroke count (indicated in the boxes), so as to ensure similar pixel-wise reconstruction loss with the target image.…”