Skip to content

Fix prepare latent image ids and vae sample generators for flux #9981

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

Merged
merged 2 commits into from
Nov 21, 2024
Merged

Fix prepare latent image ids and vae sample generators for flux #9981

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9b0cdc8
fix
a-r-r-o-w Nov 21, 2024
4d42be4
update expected slice
a-r-r-o-w Nov 21, 2024
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
2 changes: 1 addition & 1 deletion src/diffusers/pipelines/flux/pipeline_flux.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -513,7 +513,7 @@ def prepare_latents(
shape = (batch_size, num_channels_latents, height, width)

if latents is not None:
latent_image_ids = self._prepare_latent_image_ids(batch_size, height, width, device, dtype)
latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype)
return latents.to(device=device, dtype=dtype), latent_image_ids

if isinstance(generator, list) and len(generator) != batch_size:
Expand Down
20 changes: 17 additions & 3 deletions src/diffusers/pipelines/flux/pipeline_flux_controlnet.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -97,6 +97,20 @@ def calculate_shift(
return mu


# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
def retrieve_latents(
encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
):
if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
return encoder_output.latent_dist.sample(generator)
elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
return encoder_output.latent_dist.mode()
elif hasattr(encoder_output, "latents"):
return encoder_output.latents
else:
raise AttributeError("Could not access latents of provided encoder_output")


# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
def retrieve_timesteps(
scheduler,
Expand Down Expand Up @@ -512,7 +526,7 @@ def prepare_latents(
shape = (batch_size, num_channels_latents, height, width)

if latents is not None:
latent_image_ids = self._prepare_latent_image_ids(batch_size, height, width, device, dtype)
latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype)
Copy link
Member

Choose a reason for hiding this comment

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

I can recollect two PRs here:

Gently pinging @DN6

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Oh, do you mean I am reverting a previous change that was necessary for something? I have been trying to pass latents for testing purposes but it always fails at this point unless the height and width are correctly adjusted -- it makes sense to me when comparing with other pipelines and doing rough calculation with tensor shapes, but LMK if this is wrong

Copy link
Collaborator

@DN6 DN6 Nov 21, 2024 •
edited
Loading

Choose a reason for hiding this comment

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

No this is fine. We missed scaling the height and width when passing in latents. We do it here when latents aren't passed. Change is good 👍🏽

diffusers/src/diffusers/pipelines/flux/pipeline_flux.py

Line 528 in e564abe

latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype)

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Oh okay. Then I'll update the slices and merge once CI is green

return latents.to(device=device, dtype=dtype), latent_image_ids

if isinstance(generator, list) and len(generator) != batch_size:
Expand Down Expand Up @@ -772,7 +786,7 @@ def __call__(
controlnet_blocks_repeat = False if self.controlnet.input_hint_block is None else True
if self.controlnet.input_hint_block is None:
# vae encode
control_image = self.vae.encode(control_image).latent_dist.sample()
control_image = retrieve_latents(self.vae.encode(control_image), generator=generator)
control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor

# pack
Expand Down Expand Up @@ -810,7 +824,7 @@ def __call__(

if self.controlnet.nets[0].input_hint_block is None:
# vae encode
control_image_ = self.vae.encode(control_image_).latent_dist.sample()
control_image_ = retrieve_latents(self.vae.encode(control_image_), generator=generator)
control_image_ = (control_image_ - self.vae.config.shift_factor) * self.vae.config.scaling_factor

# pack
Expand Down
4 changes: 2 additions & 2 deletions src/diffusers/pipelines/flux/pipeline_flux_controlnet_image_to_image.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -801,7 +801,7 @@ def __call__(
)
height, width = control_image.shape[-2:]

control_image = self.vae.encode(control_image).latent_dist.sample()
control_image = retrieve_latents(self.vae.encode(control_image), generator=generator)
control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor

height_control_image, width_control_image = control_image.shape[2:]
Expand Down Expand Up @@ -832,7 +832,7 @@ def __call__(
)
height, width = control_image_.shape[-2:]

control_image_ = self.vae.encode(control_image_).latent_dist.sample()
control_image_ = retrieve_latents(self.vae.encode(control_image_), generator=generator)
control_image_ = (control_image_ - self.vae.config.shift_factor) * self.vae.config.scaling_factor

height_control_image, width_control_image = control_image_.shape[2:]
Expand Down
4 changes: 2 additions & 2 deletions src/diffusers/pipelines/flux/pipeline_flux_controlnet_inpainting.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -942,7 +942,7 @@ def __call__(
controlnet_blocks_repeat = False if self.controlnet.input_hint_block is None else True
if self.controlnet.input_hint_block is None:
# vae encode
control_image = self.vae.encode(control_image).latent_dist.sample()
control_image = retrieve_latents(self.vae.encode(control_image), generator=generator)
control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor

# pack
Expand Down Expand Up @@ -979,7 +979,7 @@ def __call__(

if self.controlnet.nets[0].input_hint_block is None:
# vae encode
control_image_ = self.vae.encode(control_image_).latent_dist.sample()
control_image_ = retrieve_latents(self.vae.encode(control_image_), generator=generator)
control_image_ = (control_image_ - self.vae.config.shift_factor) * self.vae.config.scaling_factor

# pack
Expand Down
Loading