does the image_embeds in the diffusers inpaint pipeline come from the prior?
#6
by
Winne
- opened
the inpaint pipeline from in diffusers requires an argument image_embeds, does this embedding come from the prior like the text-to-image pipeline like the following codes
pipe_prior = KandinskyV22PriorPipeline.from_pretrained(prior_path, torch_dtype=torch.float16)
image_emb, zero_image_emb = pipe_prior(prompt, return_dict=False)
out = pipe(
image_emb,
negative_image_embeds=zero_image_emb,
image=image,
mask_image=mask_image,
height=image.height,
width=image.width,
num_inference_steps=150,
)
out_image = out.images[0]