...To analyze how a token in the input prompt influenced the generation, you can study the token attribution scores. You can also check all the images that the diffusion process generated at the end of each step. Gradient checkpointing also reduces GPU usage, but makes computations a bit slower.