Negative prompting?
I noticed that this model seems to use true CFG, according to the codebase. Even the versions labelled as "distilled" still use a guidance_scale parameter above 1, so they also use true CFG.
However, I do not see any examples that make use of a negative prompt, and the official demo interface also does not expose the negative prompt. Is it because the model is trained to not use negative prompts at all, or are there some undocumented tips regarding negative prompting for this model?
I downloaded the model and the gradio demo code to run locally, and when I modify the demo code to take a negative prompt, there is a noticeable, though inconsistent improvement in quality. I highly recommend your team to expose the negative prompt in your official demo, because it is basically a free quality win can be at least somewhat useful.
Thanks for the info!
To answer your question, the model was not explicitly trained with negative prompts but these can be used during the CFG as usual.
May I ask what negative prompt you've been using? I will update the demo if it is general and improves the quality.
I will expose the negative prompt anyway but I'm interested if you have one that works well.
It was simply a generic, "intuitive" negative prompt because I am not a prompt engineering expert when it come to diffusion models. I do not remember the exact negative prompt I tried out, but it was something along the lines of "low quality, blurry, low resolution, distorted, disfigured, JPEG artifacts, grainy, multiple limbs, monochrome, faded" (faded causes the model to bias towards more saturated outputs, so I am not sure if it can cause excessive saturation)
I spent about an hour testing this myself, but I am not sure if this is the "optimal" negative prompt, so maybe your team might want to explore this as well.