Description
Describe the bug
The code crashes in the newest version.
The minimum viable code is part of the official Huggingface BlipDiffusion example found here, and the full example works on diffusers v0.32.2:
from diffusers.pipelines import BlipDiffusionPipeline
from diffusers import DiffusionPipeline
from diffusers.utils import load_image
import torch
blip_diffusion_pipe = BlipDiffusionPipeline.from_pretrained(
"Salesforce/blipdiffusion", torch_dtype=torch.float16
).to("cuda")
# blip_diffusion_pipe = DiffusionPipeline.from_pretrained(
# "Salesforce/blipdiffusion" , torch_dtype=torch.float16
# ).to("cuda") # my attempt to see if DiffusionPipeline will fix the problem
cond_subject = "dog"
tgt_subject = "dog"
text_prompt_input = "swimming underwater"
cond_image = load_image(
"https://huggingface.co/datasets/ayushtues/blipdiffusion_images/resolve/main/dog.jpg"
)
guidance_scale = 7.5
num_inference_steps = 25
negative_prompt = "over-exposure, under-exposure, saturated, duplicate, out of frame, lowres, cropped, worst quality, low quality, jpeg artifacts, morbid, mutilated, out of frame, ugly, bad anatomy, bad proportions, deformed, blurry, duplicate"
output = blip_diffusion_pipe(
text_prompt_input,
cond_image,
cond_subject,
tgt_subject,
guidance_scale=guidance_scale,
num_inference_steps=num_inference_steps,
neg_prompt=negative_prompt,
height=512,
width=512,
).images
# output[0].save("image.png")
output[0]Output / Error:
Warnings produce the output below, and are triggered at
cannot get type annotation for Parameter mean of <class 'diffusers.pipelines.blip_diffusion.pipeline_blip_diffusion.BlipDiffusionPipeline'>.
cannot get type annotation for Parameter std of <class 'diffusers.pipelines.blip_diffusion.pipeline_blip_diffusion.BlipDiffusionPipeline'>.
Error => KeyError: 'mean' on
Did a quick check, seems like init_kwargs have 2 tensors mean and std, in addition to vae, etc. But didn't investigate fully what the cause was.
Reproduction
Error manifested when trying to load BlipDiffusion, minimal viable example on diffusers v0.33.1:
from diffusers.pipelines import BlipDiffusionPipeline
from diffusers.utils import load_image
import torch
blip_diffusion_pipe = BlipDiffusionPipeline.from_pretrained(
"Salesforce/blipdiffusion", torch_dtype=torch.float16
).to("cuda")Logs
System Info
Environment information (same system, 2 environments):
Python (Working): 3.10.16, (Not working): 3.10.17
Transformer (Working): 4.48.2 (Not working): 4.46.3
Diffusers (Working): 0.32.2 (Not working): 0.33.1
Torch (Working): 2.6.0+cu124, (Not working): 2.6.0+cu124
Who can help?
No response