Text-to-image negative prompt differs from the reference implementation
Thanks for the quick ComfyUI support for Fibo 2.0!
It seems that there is an inconsistency with the original implementation in how the negative prompt is handled.
Reference (briaai/Fibo-2.0-dev): the model card uses two negative prompts: NEGATIVE_T2I for text-to-image and NEGATIVE_EDIT for editing. NEGATIVE_T2I improves fine detail and overall image quality, and NEGATIVE_EDIT is the null caption the model was trained with. (Before going into the text encoder, each is only wrapped in the chat template.)
ComfyUI (TextEncodeFibo2): by default, the negative prompt is the null caption, i.e. NEGATIVE_EDIT, for both text-to-image and editing. Text-to-image therefore runs without NEGATIVE_T2I and loses the detail and quality it adds.
Thank you!
Mentioning all-powerful @Kijai helps too
It was already fixed before you mentioned them XD https://github.com/Comfy-Org/ComfyUI/pull/17285