There are bug in diffusers with sdxl

[Johnson-yue]

software version :

diffusers == 0. 29.2

Bug1:

AttributeError: 'CLIPTextModelOutput' object has no attribute 'pooler_output'

Because text_encoder is CLIPTextModel , it output is transformers.modeling_outputs.BaseModelOutputWithPooling, so the output contain pooler_output .

But in text_encoder_2 is CLIPTextModelWithProjection, it output is transformers.models.clip.modeling_clip.CLIPTextModelOutput, it not have 'pooler_output' ,and in your code

LayerDiffuse_DiffusersCLI/diffusers_kdiffusion_sdxl.py

Line 101 in 3061d9a

pooled_prompt_embeds = prompt_embeds.pooler_output

, you try to get pooler_output in every text_encoder (both text_encoder and text_encoder_2), so it break!!