Dataset expanded 3x to 300+ image caption pairs, trained using prodigy optimizer and a lightly modified version of the ai-toolkit script. Using wide aspect ratio's like 1408 x 768 that are similar to the film you may be trying to recreate images from is useful.