Does the model learn or am I doing something wrong?
retained text
I am using comfyUI and the basic template Minimax H3 workflow, and not quite understanding what I'm seeing. The output seems to be "learning" or retaining information, despite me changing it. For example, if I start generating a video of a "50 year old japanese man" he will look distinctly japanese. If I go back to the prompt and change nothing else, but change "japanese" to "filipino" the man still looks distinctly japanese, almost nothing changes. I can change this to "vietnamese", "korean", "asian" and almost nothing will change about the man. Similarly, the man starts out looking 50 years old, but if I change it to "40" then "30" then "18" and re-run it, almost nothing changes, the man still look more than middle aged. Similarly, I generated a video where the camera is on the exterior of a kitchen window, watching the man preparing sushi. I ran a few versions of this prompt, and then changed it completely so that instead the camera is now inside the kitchen. Despite the prompt not containing the words "outside" or "window" at all, all the output is still from outside looking in. Even though in my subject prompt I say "exactly one japanese man" and in my shot description i say "there are no other people in the shot" it is generating the man from the first set of prompts, viewed through kitchen window, plus the new man working at the kitchen island instead of the counter viewed through the window. I'm working from the default workflow, I've changed nothing except turning "turbo" to "true" and I'm editing nothing but the prompt. Is this expected behaviour? How can I make it "forget" previous prompts?
document fetched 2026-09-27 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1wqxwwq/does_the_model_learn_or_am_i_doing_something_wrong