Skip to content

Improve post-processing performance #10159

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 1 commit into from
Closed

Improve post-processing performance #10159

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
37 changes: 24 additions & 13 deletions src/diffusers/image_processor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -236,7 +236,7 @@ def denormalize(images: Union[np.ndarray, torch.Tensor]) -> Union[np.ndarray, to
`np.ndarray` or `torch.Tensor`:
The denormalized image array.
"""
return (images / 2 + 0.5).clamp(0, 1)
return (images * 0.5 + 0.5).clamp(0, 1)

@staticmethod
def convert_to_rgb(image: PIL.Image.Image) -> PIL.Image.Image:
Expand Down Expand Up @@ -537,6 +537,27 @@ def binarize(self, image: PIL.Image.Image) -> PIL.Image.Image:

return image

def _denormalize_conditionally(
self, images: torch.Tensor, do_denormalize: Optional[List[bool]] = None
) -> torch.Tensor:
r"""
Denormalize a batch of images based on a condition list.

Args:
images (`torch.Tensor`):
The input image tensor.
do_denormalize (`Optional[List[bool]`, *optional*, defaults to `None`):
A list of booleans indicating whether to denormalize each image in the batch. If `None`, will use the
value of `do_normalize` in the `VaeImageProcessor` config.
"""
if do_denormalize is None:
return self.denormalize(images) if self.config.do_normalize else images

# De-normalizing a batch and selectively torch.stack'ing the results turns out to be
# significantly faster than performing a lot of smaller denormalizations
denormalized = self.denormalize(images)
Copy link
Collaborator

@yiyixuxu yiyixuxu Dec 9, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think there is some context to that, each image in the batch may have a different value for do_normalize for sd1.5, see the code here

diffusers/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

Line 1075 in 6131a93

image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)

Since most of the new pipelines (sdxl, sd3, flux), we do not pass the do_normalize from the pipeline, i.e. do_normalize is None here , see SDXL https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py#L1304 ,

and we already did this so it will be batch processed for all the new pipelines already, I think this line is sufficient here

        if do_denormalize is None:
            return self.denormalize(images) if self.config.do_normalize else images

return torch.stack([denormalized[i] if do_denormalize[i] else images[i] for i in range(images.shape[0])])

def get_default_height_width(
self,
image: Union[PIL.Image.Image, np.ndarray, torch.Tensor],
Expand Down Expand Up @@ -752,12 +773,7 @@ def postprocess(
if output_type == "latent":
return image

if do_denormalize is None:
do_denormalize = [self.config.do_normalize] * image.shape[0]

image = torch.stack(
[self.denormalize(image[i]) if do_denormalize[i] else image[i] for i in range(image.shape[0])]
)
image = self._denormalize_conditionally(image, do_denormalize)
Copy link
Collaborator

@yiyixuxu yiyixuxu Dec 9, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more. 

if do_denormalize is None:
            image = self.denormalize(images) if self.config.do_normalize 
else: 
        image = torch.stack(
            [self.denormalize(image[i]) if do_denormalize[i] else image[i] for i in range(image.shape[0])]
        )


if output_type == "pt":
return image
Expand Down Expand Up @@ -966,12 +982,7 @@ def postprocess(
deprecate("Unsupported output_type", "1.0.0", deprecation_message, standard_warn=False)
output_type = "np"

if do_denormalize is None:
do_denormalize = [self.config.do_normalize] * image.shape[0]

image = torch.stack(
[self.denormalize(image[i]) if do_denormalize[i] else image[i] for i in range(image.shape[0])]
)
image = self._denormalize_conditionally(image, do_denormalize)

image = self.pt_to_numpy(image)

Expand Down
Loading