guidance_scale < 1 crashes
Clement-Lelievre opened this issue · 0 comments
hello,
thank you for your model.
I have come across a bug: Using guidance_scale < 1
breaks.
Repro steps: launch the gradio app, select any params, and guidance scale < 1
This is because in the call to the pipe, the prompt encoding method, specifically the concatenation of the prompt embeddings and the negative prompt embeddings, fails here due to negative_prompt_embeds
being None
.
This is because we are not in a classifier-free guidance context, again due to the value of guidance_scale
being < 1.
Can you please advise on the best course of action to correct this?
The intuitive solution is to set guidance_scale >=1 , but I'd like to ascertain this is OK. I had a look at the Imagen paper, and it seems they never use guidance < 1.
Thanks,
FWIW:
diffusers==0.27.0
torch==2.3.1
transformers==4.42.4
python==3.12.0
cc @zengyh1900