SDXL prompt embeddings not working

[adhikjoshi]

Describe the bug

pipe = StableDiffusionXLPipeline.from_pretrained(
                    model_id, torch_dtype=torch.float16, use_safetensors=True, variant="fp16", use_auth_token=hf_token).to("cuda")
cond, uncond = text_embeddings(
    self.pipe, args_dict["prompt"], args_dict["negative_prompt"], clip_skip=2)
del args_dict["prompt"]
del args_dict["negative_prompt"]
args_dict["prompt_embeds"] = cond
args_dict["negative_prompt_embeds"] = uncond
pipe(**args_dict)

Gives error

File "/opt/conda/envs/ldm/lib/python3.11/site-packages/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py", line 402, in encode_prompt
pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
                           ^^^^^^^^^^^^^^^^^^^^
 UnboundLocalError: cannot access local variable 'pooled_prompt_embeds' where it is not associated with a value

Reproduction

SDXL

Logs

No response

System Info

pip3 install git+https://github.com/huggingface/diffusers.git

Who can help?

@patrickvonplaten

[bghira]

it's true. they do not work. and img2img can't use a negative prompt at all?

[sayakpaul]

It's a known issue and we're working on it. Cc: @patrickvonplaten

[patrickvonplaten]

We're working with Compel to help the user have a nice UX here