Painting Many Pasts: Synthesizing Time Lapse Videos of Paintings

Intelli-Paint: Towards Developing Human-like Painting Agents

“…Human painters often exhibit a localized spatial attention span while focusing on distinct image areas [34]. This is in stark contrast with previous works which either compute stroke decomposition globally over the entire canvas or over a set of predefined grid regions [12,21,36].…”

“…However, these methods require access to dense human brushstroke annotations, which limits their applicability to most real world problems. In another work, Zhao et al [34] use a conditional variational autoencoder framework for synthesising timelapse videos depicting the recreation of a given target image. However, this requires access to painting time-lapse videos from real artists for training.…”

“…The human painting process is often progressive and multilayered [25,34]. That is, instead of painting everything on the canvas at once, humans often first paint a basic background layer before progressively adding each of the foreground objects on top of it (refer Fig.…”

“…Human artists inherently try to focus on spatially close image areas and try to avoid unnecessary spatial oscillations when painting a given image [34]. In this regard, while the Markovian adjustment procedure introduced in Sec.…”

“…Second, the human painting process is characterized by a localized spatial attention span. For instance, a potential artist would focus on different local image areas while painting distinct parts of the final canvas [34]. This is in sharp contrast with previous works, which either focus on the entire image or several predefined grid blocks [21,36].…”