I wonder if it benefits because it can attend to individual tokens of the prompt while generating, compared to typical diffusion models that just get a static v

I wonder if it benefits because it can attend to individual tokens of the prompt while generating, compared to typical diffusion models that just get a static vector embedding of the prompt.