Controlling length
I'm having trouble reproducing your example outputs of These are two cats, one brown and one gray, lying on a pink blanket. sleep. brown and gray cat sleeping on a pink blanket.
I get: An image of two cats sleeping on a pink couch. The cat on the left is a tabby with black stripes, while the cat on the right is a tiger-striped cat. Both cats are curled up and appear to be sleeping. There is a remote control on the couch next to the cats. The image is taken from a high angle, looking down on the cats. The pink color of the couch contrasts with the colors of the cats, making them stand out in the image. The remote control adds a touch of modernity to the scene. The high angle of the shot gives a sense of depth and perspective to the image. Overall, the image captures a peaceful and cozy moment of two cats sleeping together. cats tabby tiger-striped pink couch remote control high angle depth perspective peaceful cozy tabby tiger-striped pink couch remote control high angle
With an abrupt stop.
I tried using many various sampling arguments exposed by transformers like length_penalty, repetition_penalty, num_beams=[1..5], do_sample=True with various top_p, top_k, and temperature.
Was your sample doing post processing to simply select the first n sentences, or am I missing something?