Pictures don't have to be 2D for an RNN. Flatten an ASCII drawing row-by-row, treat newlines as "carriage return" tokens, and a sequence model learns to draw. Watch faces emerge from chaos.
Training Image
Hyperparameters
Training Loss
Top Character Transitions
After training, the most-frequent character bigrams in your data.
Train first.
Generated ASCII
How does this work?
An ASCII art image is just a string with newlines. The RNN reads characters one at a time and learns the pattern: "after a row of pound signs comes a newline, then dots and an O for an eye."
The newline character is the secret weapon — it acts as a structural marker that lets the network learn 2D shape from a 1D sequence. Without learning when to insert newlines, generated output collapses into nonsense.
Compare this approach to PixelRNN (van den Oord et al., 2016) which extends this idea to grayscale and color images.