Skip to content

GitLab

"test/vscode:/vscode.git/clone" did not exist on "3c699772c9fa27b734bf91d466306a4ec1b628b4"
Unverified Commit fa633ed6 authored by elucida's avatar elucida Committed by GitHub
Browse files

refactor: move model helper function in pipeline to a mixin class (#6571) 



* move model helper function in pipeline to EfficiencyMixin

---------
Co-authored-by: default avatarYiYi Xu <yixu310@gmail.com>
Co-authored-by: default avatarSayak Paul <spsayakpaul@gmail.com>
parent 2cad1a84
Show whitespace changes
Inline Side-by-side
Showing with 56 additions and 1179 deletions
examples/community/pipeline_stable_diffusion_xl_controlnet_adapter.py
View file @ fa633ed6
......@@ -33,7 +33,7 @@ from diffusers.models.attention_processor import (
)
from diffusers.models.lora import adjust_lora_scale_text_encoder
from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
......@@ -158,7 +158,11 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
class StableDiffusionXLControlNetAdapterPipeline(
DiffusionPipeline, FromSingleFileMixin, StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin
DiffusionPipeline,
StableDiffusionMixin,
FromSingleFileMixin,
StableDiffusionXLLoraLoaderMixin,
TextualInversionLoaderMixin,
):
r"""
Pipeline for text-to-image generation using Stable Diffusion augmented with T2I-Adapter
......@@ -234,39 +238,6 @@ class StableDiffusionXLControlNetAdapterPipeline(
)
self.default_sample_size = self.unet.config.sample_size
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
def enable_vae_tiling(self):
r"""
Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
processing larger images.
"""
self.vae.enable_tiling()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
def disable_vae_tiling(self):
r"""
Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
computing decoding in one step.
"""
self.vae.disable_tiling()
# Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
def encode_prompt(
self,
......@@ -863,34 +834,6 @@ class StableDiffusionXLControlNetAdapterPipeline(
return height, width
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
The suffixes after the scaling factors represent the stages where they are being applied.
Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
Args:
s1 (`float`):
Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
mitigate "oversmoothing effect" in the enhanced denoising process.
s2 (`float`):
Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
mitigate "oversmoothing effect" in the enhanced denoising process.
b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
"""
if not hasattr(self, "unet"):
raise ValueError("The pipeline must have `unet` for using FreeU.")
self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
def disable_freeu(self):
"""Disables the FreeU mechanism if enabled."""
self.unet.disable_freeu()
def prepare_control_image(
self,
image,
......
examples/community/pipeline_stable_diffusion_xl_controlnet_adapter_inpaint.py
View file @ fa633ed6
......@@ -52,6 +52,7 @@ from diffusers.models.attention_processor import (
)
from diffusers.models.lora import adjust_lora_scale_text_encoder
from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
......@@ -303,7 +304,9 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
return noise_cfg
class StableDiffusionXLControlNetAdapterInpaintPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoaderMixin):
class StableDiffusionXLControlNetAdapterInpaintPipeline(
DiffusionPipeline, StableDiffusionMixin, FromSingleFileMixin, LoraLoaderMixin
):
r"""
Pipeline for text-to-image generation using Stable Diffusion augmented with T2I-Adapter
https://arxiv.org/abs/2302.08453
......@@ -383,39 +386,6 @@ class StableDiffusionXLControlNetAdapterInpaintPipeline(DiffusionPipeline, FromS
)
self.default_sample_size = self.unet.config.sample_size
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
def enable_vae_tiling(self):
r"""
Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
processing larger images.
"""
self.vae.enable_tiling()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
def disable_vae_tiling(self):
r"""
Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
computing decoding in one step.
"""
self.vae.disable_tiling()
# Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
def encode_prompt(
self,
......@@ -1207,34 +1177,6 @@ class StableDiffusionXLControlNetAdapterInpaintPipeline(DiffusionPipeline, FromS
return height, width
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
The suffixes after the scaling factors represent the stages where they are being applied.
Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
Args:
s1 (`float`):
Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
mitigate "oversmoothing effect" in the enhanced denoising process.
s2 (`float`):
Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
mitigate "oversmoothing effect" in the enhanced denoising process.
b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
"""
if not hasattr(self, "unet"):
raise ValueError("The pipeline must have `unet` for using FreeU.")
self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
def disable_freeu(self):
"""Disables the FreeU mechanism if enabled."""
self.unet.disable_freeu()
def prepare_control_image(
self,
image,
......
examples/community/pipeline_stable_diffusion_xl_ipex.py
View file @ fa633ed6
......@@ -267,39 +267,6 @@ class StableDiffusionXLPipelineIpex(
else:
self.watermark = None
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
def enable_vae_tiling(self):
r"""
Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
processing larger images.
"""
self.vae.enable_tiling()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
def disable_vae_tiling(self):
r"""
Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
computing decoding in one step.
"""
self.vae.disable_tiling()
def encode_prompt(
self,
prompt: str,
......@@ -701,34 +668,6 @@ class StableDiffusionXLPipelineIpex(
self.vae.decoder.conv_in.to(dtype)
self.vae.decoder.mid_block.to(dtype)
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
The suffixes after the scaling factors represent the stages where they are being applied.
Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
Args:
s1 (`float`):
Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
mitigate "oversmoothing effect" in the enhanced denoising process.
s2 (`float`):
Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
mitigate "oversmoothing effect" in the enhanced denoising process.
b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
"""
if not hasattr(self, "unet"):
raise ValueError("The pipeline must have `unet` for using FreeU.")
self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
def disable_freeu(self):
"""Disables the FreeU mechanism if enabled."""
self.unet.disable_freeu()
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
"""
......
examples/community/pipeline_zero1to3.py
View file @ fa633ed6
......@@ -22,18 +22,16 @@ from transformers import CLIPFeatureExtractor, CLIPVisionModelWithProjection
# randn_tensor,
# replace_example_docstring,
# )
# from ..pipeline_utils import DiffusionPipeline
# from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
# from . import StableDiffusionPipelineOutput
# from .safety_checker import StableDiffusionSafetyChecker
from diffusers import AutoencoderKL, DiffusionPipeline, UNet2DConditionModel
from diffusers import AutoencoderKL, DiffusionPipeline, StableDiffusionMixin, UNet2DConditionModel
from diffusers.configuration_utils import ConfigMixin, FrozenDict
from diffusers.models.modeling_utils import ModelMixin
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
deprecate,
is_accelerate_available,
is_accelerate_version,
logging,
replace_example_docstring,
)
......@@ -68,7 +66,7 @@ class CCProjection(ModelMixin, ConfigMixin):
return self.projection(x)
class Zero1to3StableDiffusionPipeline(DiffusionPipeline):
class Zero1to3StableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
r"""
Pipeline for single view conditioned novel view generation using Zero1to3.
......@@ -187,109 +185,6 @@ class Zero1to3StableDiffusionPipeline(DiffusionPipeline):
self.register_to_config(requires_safety_checker=requires_safety_checker)
# self.model_mode = None
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding.
When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
def enable_vae_tiling(self):
r"""
Enable tiled VAE decoding.
When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in
several steps. This is useful to save a large amount of memory and to allow the processing of larger images.
"""
self.vae.enable_tiling()
def disable_vae_tiling(self):
r"""
Disable tiled VAE decoding. If `enable_vae_tiling` was previously invoked, this method will go back to
computing decoding in one step.
"""
self.vae.disable_tiling()
def enable_sequential_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
`torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
Note that offloading happens on a submodule basis. Memory savings are higher than with
`enable_model_cpu_offload`, but performance is lower.
"""
if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
from accelerate import cpu_offload
else:
raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
device = torch.device(f"cuda:{gpu_id}")
if self.device.type != "cpu":
self.to("cpu", silence_dtype_warnings=True)
torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist)
for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
cpu_offload(cpu_offloaded_model, device)
if self.safety_checker is not None:
cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
def enable_model_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
`enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
"""
if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
from accelerate import cpu_offload_with_hook
else:
raise ImportError("`enable_model_offload` requires `accelerate v0.17.0` or higher.")
device = torch.device(f"cuda:{gpu_id}")
if self.device.type != "cpu":
self.to("cpu", silence_dtype_warnings=True)
torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist)
hook = None
for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
_, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
if self.safety_checker is not None:
_, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
# We'll offload the last model manually.
self.final_offload_hook = hook
@property
def _execution_device(self):
r"""
Returns the device on which the pipeline's models will be executed. After calling
`pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
hooks.
"""
if not hasattr(self.unet, "_hf_hook"):
return self.device
for module in self.unet.modules():
if (
hasattr(module, "_hf_hook")
and hasattr(module._hf_hook, "execution_device")
and module._hf_hook.execution_device is not None
):
return torch.device(module._hf_hook.execution_device)
return self.device
def _encode_prompt(
self,
prompt,
......
examples/community/sd_text2img_k_diffusion.py
View file @ fa633ed6
......@@ -19,9 +19,9 @@ from typing import Callable, List, Optional, Union
import torch
from k_diffusion.external import CompVisDenoiser, CompVisVDenoiser
from diffusers import DiffusionPipeline, LMSDiscreteScheduler
from diffusers import DiffusionPipeline, LMSDiscreteScheduler, StableDiffusionMixin
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from diffusers.utils import is_accelerate_available, logging
from diffusers.utils import logging
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
......@@ -41,7 +41,7 @@ class ModelWrapper:
return self.model(*args, encoder_hidden_states=encoder_hidden_states, **kwargs).sample
class StableDiffusionPipeline(DiffusionPipeline):
class StableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
r"""
Pipeline for text-to-image generation using Stable Diffusion.
......@@ -120,68 +120,6 @@ class StableDiffusionPipeline(DiffusionPipeline):
sampling = getattr(library, "sampling")
self.sampler = getattr(sampling, scheduler_type)
def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
r"""
Enable sliced attention computation.
When this option is enabled, the attention module will split the input tensor in slices, to compute attention
in several steps. This is useful to save some memory in exchange for a small speed decrease.
Args:
slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
`attention_head_dim` must be a multiple of `slice_size`.
"""
if slice_size == "auto":
# half the attention head size is usually a good trade-off between
# speed and memory
slice_size = self.unet.config.attention_head_dim // 2
self.unet.set_attention_slice(slice_size)
def disable_attention_slicing(self):
r"""
Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
back to computing attention in one step.
"""
# set slice_size = `None` to disable `attention slicing`
self.enable_attention_slicing(None)
def enable_sequential_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
`torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
"""
if is_accelerate_available():
from accelerate import cpu_offload
else:
raise ImportError("Please install accelerate via `pip install accelerate`")
device = torch.device(f"cuda:{gpu_id}")
for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.safety_checker]:
if cpu_offloaded_model is not None:
cpu_offload(cpu_offloaded_model, device)
@property
def _execution_device(self):
r"""
Returns the device on which the pipeline's models will be executed. After calling
`pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
hooks.
"""
if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
return self.device
for module in self.unet.modules():
if (
hasattr(module, "_hf_hook")
and hasattr(module._hf_hook, "execution_device")
and module._hf_hook.execution_device is not None
):
return torch.device(module._hf_hook.execution_device)
return self.device
def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt):
r"""
Encodes the prompt into text encoder hidden states.
......
examples/community/seed_resize_stable_diffusion.py
View file @ fa633ed6
......@@ -9,6 +9,7 @@ from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
from diffusers import DiffusionPipeline
from diffusers.models import AutoencoderKL, UNet2DConditionModel
from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
......@@ -18,7 +19,7 @@ from diffusers.utils import logging
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
class SeedResizeStableDiffusionPipeline(DiffusionPipeline):
class SeedResizeStableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
r"""
Pipeline for text-to-image generation using Stable Diffusion.
......@@ -67,33 +68,6 @@ class SeedResizeStableDiffusionPipeline(DiffusionPipeline):
feature_extractor=feature_extractor,
)
def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
r"""
Enable sliced attention computation.
When this option is enabled, the attention module will split the input tensor in slices, to compute attention
in several steps. This is useful to save some memory in exchange for a small speed decrease.
Args:
slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
`attention_head_dim` must be a multiple of `slice_size`.
"""
if slice_size == "auto":
# half the attention head size is usually a good trade-off between
# speed and memory
slice_size = self.unet.config.attention_head_dim // 2
self.unet.set_attention_slice(slice_size)
def disable_attention_slicing(self):
r"""
Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
back to computing attention in one step.
"""
# set slice_size = `None` to disable `attention slicing`
self.enable_attention_slicing(None)
@torch.no_grad()
def __call__(
self,
......
examples/community/speech_to_image_diffusion.py
View file @ fa633ed6
......@@ -18,6 +18,7 @@ from diffusers import (
PNDMScheduler,
UNet2DConditionModel,
)
from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
from diffusers.utils import logging
......@@ -26,7 +27,7 @@ from diffusers.utils import logging
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
class SpeechToImagePipeline(DiffusionPipeline):
class SpeechToImagePipeline(DiffusionPipeline, StableDiffusionMixin):
def __init__(
self,
speech_model: WhisperForConditionalGeneration,
......@@ -62,14 +63,6 @@ class SpeechToImagePipeline(DiffusionPipeline):
feature_extractor=feature_extractor,
)
def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
if slice_size == "auto":
slice_size = self.unet.config.attention_head_dim // 2
self.unet.set_attention_slice(slice_size)
def disable_attention_slicing(self):
self.enable_attention_slicing(None)
@torch.no_grad()
def __call__(
self,
......
examples/community/stable_diffusion_comparison.py
View file @ fa633ed6
......@@ -12,6 +12,7 @@ from diffusers import (
StableDiffusionPipeline,
UNet2DConditionModel,
)
from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
......@@ -22,7 +23,7 @@ pipe3_model_id = "CompVis/stable-diffusion-v1-3"
pipe4_model_id = "CompVis/stable-diffusion-v1-4"
class StableDiffusionComparisonPipeline(DiffusionPipeline):
class StableDiffusionComparisonPipeline(DiffusionPipeline, StableDiffusionMixin):
r"""
Pipeline for parallel comparison of Stable Diffusion v1-v4
This pipeline inherits from DiffusionPipeline and depends on the use of an Auth Token for
......@@ -83,31 +84,6 @@ class StableDiffusionComparisonPipeline(DiffusionPipeline):
def layers(self) -> Dict[str, Any]:
return {k: getattr(self, k) for k in self.config.keys() if not k.startswith("_")}
def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
r"""
Enable sliced attention computation.
When this option is enabled, the attention module will split the input tensor in slices, to compute attention
in several steps. This is useful to save some memory in exchange for a small speed decrease.
Args:
slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
`attention_head_dim` must be a multiple of `slice_size`.
"""
if slice_size == "auto":
# half the attention head size is usually a good trade-off between
# speed and memory
slice_size = self.unet.config.attention_head_dim // 2
self.unet.set_attention_slice(slice_size)
def disable_attention_slicing(self):
r"""
Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
back to computing attention in one step.
"""
# set slice_size = `None` to disable `attention slicing`
self.enable_attention_slicing(None)
@torch.no_grad()
def text2img_sd1_1(
self,
......
examples/community/stable_diffusion_controlnet_img2img.py
View file @ fa633ed6
......@@ -8,14 +8,13 @@ import PIL.Image
import torch
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
from diffusers import AutoencoderKL, ControlNetModel, DiffusionPipeline, UNet2DConditionModel, logging
from diffusers import AutoencoderKL, ControlNetModel, UNet2DConditionModel, logging
from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
PIL_INTERPOLATION,
is_accelerate_available,
is_accelerate_version,
replace_example_docstring,
)
from diffusers.utils.torch_utils import randn_tensor
......@@ -130,7 +129,7 @@ def prepare_controlnet_conditioning_image(
return controlnet_conditioning_image
class StableDiffusionControlNetImg2ImgPipeline(DiffusionPipeline):
class StableDiffusionControlNetImg2ImgPipeline(DiffusionPipeline, StableDiffusionMixin):
"""
Inspired by: https://github.com/haofanwang/ControlNet-for-Diffusers/
"""
......@@ -183,89 +182,6 @@ class StableDiffusionControlNetImg2ImgPipeline(DiffusionPipeline):
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
self.register_to_config(requires_safety_checker=requires_safety_checker)
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding.
When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
def enable_sequential_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
text_encoder, vae, controlnet, and safety checker have their state dicts saved to CPU and then are moved to a
`torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
Note that offloading happens on a submodule basis. Memory savings are higher than with
`enable_model_cpu_offload`, but performance is lower.
"""
if is_accelerate_available():
from accelerate import cpu_offload
else:
raise ImportError("Please install accelerate via `pip install accelerate`")
device = torch.device(f"cuda:{gpu_id}")
for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.controlnet]:
cpu_offload(cpu_offloaded_model, device)
if self.safety_checker is not None:
cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
def enable_model_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
`enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
"""
if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
from accelerate import cpu_offload_with_hook
else:
raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
device = torch.device(f"cuda:{gpu_id}")
hook = None
for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
_, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
if self.safety_checker is not None:
# the safety checker can offload the vae again
_, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
# control net hook has be manually offloaded as it alternates with unet
cpu_offload_with_hook(self.controlnet, device)
# We'll offload the last model manually.
self.final_offload_hook = hook
@property
def _execution_device(self):
r"""
Returns the device on which the pipeline's models will be executed. After calling
`pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
hooks.
"""
if not hasattr(self.unet, "_hf_hook"):
return self.device
for module in self.unet.modules():
if (
hasattr(module, "_hf_hook")
and hasattr(module._hf_hook, "execution_device")
and module._hf_hook.execution_device is not None
):
return torch.device(module._hf_hook.execution_device)
return self.device
def _encode_prompt(
self,
prompt,
......
examples/community/stable_diffusion_controlnet_inpaint.py
View file @ fa633ed6
......@@ -9,14 +9,13 @@ import torch
import torch.nn.functional as F
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
from diffusers import AutoencoderKL, ControlNetModel, DiffusionPipeline, UNet2DConditionModel, logging
from diffusers import AutoencoderKL, ControlNetModel, UNet2DConditionModel, logging
from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
PIL_INTERPOLATION,
is_accelerate_available,
is_accelerate_version,
replace_example_docstring,
)
from diffusers.utils.torch_utils import randn_tensor
......@@ -228,7 +227,7 @@ def prepare_controlnet_conditioning_image(
return controlnet_conditioning_image
class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline):
class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, StableDiffusionMixin):
"""
Inspired by: https://github.com/haofanwang/ControlNet-for-Diffusers/
"""
......@@ -282,89 +281,6 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline):
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
self.register_to_config(requires_safety_checker=requires_safety_checker)
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding.
When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
def enable_sequential_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
text_encoder, vae, controlnet, and safety checker have their state dicts saved to CPU and then are moved to a
`torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
Note that offloading happens on a submodule basis. Memory savings are higher than with
`enable_model_cpu_offload`, but performance is lower.
"""
if is_accelerate_available():
from accelerate import cpu_offload
else:
raise ImportError("Please install accelerate via `pip install accelerate`")
device = torch.device(f"cuda:{gpu_id}")
for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.controlnet]:
cpu_offload(cpu_offloaded_model, device)
if self.safety_checker is not None:
cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
def enable_model_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
`enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
"""
if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
from accelerate import cpu_offload_with_hook
else:
raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
device = torch.device(f"cuda:{gpu_id}")
hook = None
for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
_, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
if self.safety_checker is not None:
# the safety checker can offload the vae again
_, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
# control net hook has be manually offloaded as it alternates with unet
cpu_offload_with_hook(self.controlnet, device)
# We'll offload the last model manually.
self.final_offload_hook = hook
@property
def _execution_device(self):
r"""
Returns the device on which the pipeline's models will be executed. After calling
`pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
hooks.
"""
if not hasattr(self.unet, "_hf_hook"):
return self.device
for module in self.unet.modules():
if (
hasattr(module, "_hf_hook")
and hasattr(module._hf_hook, "execution_device")
and module._hf_hook.execution_device is not None
):
return torch.device(module._hf_hook.execution_device)
return self.device
def _encode_prompt(
self,
prompt,
......
examples/community/stable_diffusion_controlnet_inpaint_img2img.py
View file @ fa633ed6
......@@ -9,13 +9,12 @@ import torch
import torch.nn.functional as F
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
from diffusers import AutoencoderKL, ControlNetModel, DiffusionPipeline, UNet2DConditionModel, logging
from diffusers import AutoencoderKL, ControlNetModel, UNet2DConditionModel, logging
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
PIL_INTERPOLATION,
is_accelerate_available,
is_accelerate_version,
replace_example_docstring,
)
from diffusers.utils.torch_utils import randn_tensor
......@@ -217,7 +216,7 @@ def prepare_controlnet_conditioning_image(
return controlnet_conditioning_image
class StableDiffusionControlNetInpaintImg2ImgPipeline(DiffusionPipeline):
class StableDiffusionControlNetInpaintImg2ImgPipeline(DiffusionPipeline, StableDiffusionMixin):
"""
Inspired by: https://github.com/haofanwang/ControlNet-for-Diffusers/
"""
......@@ -267,89 +266,6 @@ class StableDiffusionControlNetInpaintImg2ImgPipeline(DiffusionPipeline):
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
self.register_to_config(requires_safety_checker=requires_safety_checker)
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding.
When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
def enable_sequential_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
text_encoder, vae, controlnet, and safety checker have their state dicts saved to CPU and then are moved to a
`torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
Note that offloading happens on a submodule basis. Memory savings are higher than with
`enable_model_cpu_offload`, but performance is lower.
"""
if is_accelerate_available():
from accelerate import cpu_offload
else:
raise ImportError("Please install accelerate via `pip install accelerate`")
device = torch.device(f"cuda:{gpu_id}")
for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.controlnet]:
cpu_offload(cpu_offloaded_model, device)
if self.safety_checker is not None:
cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
def enable_model_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
`enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
"""
if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
from accelerate import cpu_offload_with_hook
else:
raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
device = torch.device(f"cuda:{gpu_id}")
hook = None
for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
_, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
if self.safety_checker is not None:
# the safety checker can offload the vae again
_, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
# control net hook has be manually offloaded as it alternates with unet
cpu_offload_with_hook(self.controlnet, device)
# We'll offload the last model manually.
self.final_offload_hook = hook
@property
def _execution_device(self):
r"""
Returns the device on which the pipeline's models will be executed. After calling
`pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
hooks.
"""
if not hasattr(self.unet, "_hf_hook"):
return self.device
for module in self.unet.modules():
if (
hasattr(module, "_hf_hook")
and hasattr(module._hf_hook, "execution_device")
and module._hf_hook.execution_device is not None
):
return torch.device(module._hf_hook.execution_device)
return self.device
def _encode_prompt(
self,
prompt,
......
examples/community/stable_diffusion_ipex.py
View file @ fa633ed6
......@@ -23,14 +23,12 @@ from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
from diffusers.configuration_utils import FrozenDict
from diffusers.loaders import LoraLoaderMixin, TextualInversionLoaderMixin
from diffusers.models import AutoencoderKL, UNet2DConditionModel
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
deprecate,
is_accelerate_available,
is_accelerate_version,
logging,
replace_example_docstring,
)
......@@ -62,7 +60,9 @@ EXAMPLE_DOC_STRING = """
"""
class StableDiffusionIPEXPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
class StableDiffusionIPEXPipeline(
DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin
):
r"""
Pipeline for text-to-image generation using Stable Diffusion on IPEX.
......@@ -304,109 +304,6 @@ class StableDiffusionIPEXPipeline(DiffusionPipeline, TextualInversionLoaderMixin
ave_decoder_trace_model = torch.jit.freeze(ave_decoder_trace_model)
self.vae.decoder.forward = ave_decoder_trace_model.forward
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding.
When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
def enable_vae_tiling(self):
r"""
Enable tiled VAE decoding.
When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in
several steps. This is useful to save a large amount of memory and to allow the processing of larger images.
"""
self.vae.enable_tiling()
def disable_vae_tiling(self):
r"""
Disable tiled VAE decoding. If `enable_vae_tiling` was previously invoked, this method will go back to
computing decoding in one step.
"""
self.vae.disable_tiling()
def enable_sequential_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
`torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
Note that offloading happens on a submodule basis. Memory savings are higher than with
`enable_model_cpu_offload`, but performance is lower.
"""
if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
from accelerate import cpu_offload
else:
raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
device = torch.device(f"cuda:{gpu_id}")
if self.device.type != "cpu":
self.to("cpu", silence_dtype_warnings=True)
torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist)
for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
cpu_offload(cpu_offloaded_model, device)
if self.safety_checker is not None:
cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
def enable_model_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
`enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
"""
if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
from accelerate import cpu_offload_with_hook
else:
raise ImportError("`enable_model_offload` requires `accelerate v0.17.0` or higher.")
device = torch.device(f"cuda:{gpu_id}")
if self.device.type != "cpu":
self.to("cpu", silence_dtype_warnings=True)
torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist)
hook = None
for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
_, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
if self.safety_checker is not None:
_, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
# We'll offload the last model manually.
self.final_offload_hook = hook
@property
def _execution_device(self):
r"""
Returns the device on which the pipeline's models will be executed. After calling
`pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
hooks.
"""
if not hasattr(self.unet, "_hf_hook"):
return self.device
for module in self.unet.modules():
if (
hasattr(module, "_hf_hook")
and hasattr(module._hf_hook, "execution_device")
and module._hf_hook.execution_device is not None
):
return torch.device(module._hf_hook.execution_device)
return self.device
def _encode_prompt(
self,
prompt,
......
examples/community/stable_diffusion_mega.py
View file @ fa633ed6
......@@ -16,6 +16,7 @@ from diffusers import (
UNet2DConditionModel,
)
from diffusers.configuration_utils import FrozenDict
from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
from diffusers.utils import deprecate, logging
......@@ -23,7 +24,7 @@ from diffusers.utils import deprecate, logging
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
class StableDiffusionMegaPipeline(DiffusionPipeline):
class StableDiffusionMegaPipeline(DiffusionPipeline, StableDiffusionMixin):
r"""
Pipeline for text-to-image generation using Stable Diffusion.
......@@ -94,33 +95,6 @@ class StableDiffusionMegaPipeline(DiffusionPipeline):
def components(self) -> Dict[str, Any]:
return {k: getattr(self, k) for k in self.config.keys() if not k.startswith("_")}
def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
r"""
Enable sliced attention computation.
When this option is enabled, the attention module will split the input tensor in slices, to compute attention
in several steps. This is useful to save some memory in exchange for a small speed decrease.
Args:
slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
`attention_head_dim` must be a multiple of `slice_size`.
"""
if slice_size == "auto":
# half the attention head size is usually a good trade-off between
# speed and memory
slice_size = self.unet.config.attention_head_dim // 2
self.unet.set_attention_slice(slice_size)
def disable_attention_slicing(self):
r"""
Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
back to computing attention in one step.
"""
# set slice_size = `None` to disable `attention slicing`
self.enable_attention_slicing(None)
@torch.no_grad()
def inpaint(
self,
......
examples/community/stable_diffusion_repaint.py
View file @ fa633ed6
......@@ -24,14 +24,13 @@ from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
from diffusers import AutoencoderKL, DiffusionPipeline, UNet2DConditionModel
from diffusers.configuration_utils import FrozenDict, deprecate
from diffusers.loaders import LoraLoaderMixin, TextualInversionLoaderMixin
from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from diffusers.pipelines.stable_diffusion.safety_checker import (
StableDiffusionSafetyChecker,
)
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
is_accelerate_available,
is_accelerate_version,
logging,
)
from diffusers.utils.torch_utils import randn_tensor
......@@ -140,7 +139,9 @@ def prepare_mask_and_masked_image(image, mask):
return mask, masked_image
class StableDiffusionRepaintPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
class StableDiffusionRepaintPipeline(
DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin
):
r"""
Pipeline for text-guided image inpainting using Stable Diffusion. *This is an experimental feature*.
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
......@@ -276,80 +277,6 @@ class StableDiffusionRepaintPipeline(DiffusionPipeline, TextualInversionLoaderMi
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
self.register_to_config(requires_safety_checker=requires_safety_checker)
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_sequential_cpu_offload
def enable_sequential_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
`torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
Note that offloading happens on a submodule basis. Memory savings are higher than with
`enable_model_cpu_offload`, but performance is lower.
"""
if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
from accelerate import cpu_offload
else:
raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
device = torch.device(f"cuda:{gpu_id}")
if self.device.type != "cpu":
self.to("cpu", silence_dtype_warnings=True)
torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist)
for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
cpu_offload(cpu_offloaded_model, device)
if self.safety_checker is not None:
cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_model_cpu_offload
def enable_model_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
`enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
"""
if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
from accelerate import cpu_offload_with_hook
else:
raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
device = torch.device(f"cuda:{gpu_id}")
if self.device.type != "cpu":
self.to("cpu", silence_dtype_warnings=True)
torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist)
hook = None
for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
_, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
if self.safety_checker is not None:
_, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
# We'll offload the last model manually.
self.final_offload_hook = hook
@property
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
def _execution_device(self):
r"""
Returns the device on which the pipeline's models will be executed. After calling
`pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
hooks.
"""
if not hasattr(self.unet, "_hf_hook"):
return self.device
for module in self.unet.modules():
if (
hasattr(module, "_hf_hook")
and hasattr(module._hf_hook, "execution_device")
and module._hf_hook.execution_device is not None
):
return torch.device(module._hf_hook.execution_device)
return self.device
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
def _encode_prompt(
self,
......
examples/community/text_inpainting.py
View file @ fa633ed6
......@@ -13,16 +13,17 @@ from transformers import (
from diffusers import DiffusionPipeline
from diffusers.configuration_utils import FrozenDict
from diffusers.models import AutoencoderKL, UNet2DConditionModel
from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
from diffusers.pipelines.stable_diffusion import StableDiffusionInpaintPipeline
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
from diffusers.utils import deprecate, is_accelerate_available, logging
from diffusers.utils import deprecate, logging
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
class TextInpainting(DiffusionPipeline):
class TextInpainting(DiffusionPipeline, StableDiffusionMixin):
r"""
Pipeline for text based inpainting using Stable Diffusion.
Uses CLIPSeg to get a mask from the given text, then calls the Inpainting pipeline with the generated mask
......@@ -120,69 +121,6 @@ class TextInpainting(DiffusionPipeline):
feature_extractor=feature_extractor,
)
def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
r"""
Enable sliced attention computation.
When this option is enabled, the attention module will split the input tensor in slices, to compute attention
in several steps. This is useful to save some memory in exchange for a small speed decrease.
Args:
slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
`attention_head_dim` must be a multiple of `slice_size`.
"""
if slice_size == "auto":
# half the attention head size is usually a good trade-off between
# speed and memory
slice_size = self.unet.config.attention_head_dim // 2
self.unet.set_attention_slice(slice_size)
def disable_attention_slicing(self):
r"""
Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go
back to computing attention in one step.
"""
# set slice_size = `None` to disable `attention slicing`
self.enable_attention_slicing(None)
def enable_sequential_cpu_offload(self):
r"""
Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
`torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
"""
if is_accelerate_available():
from accelerate import cpu_offload
else:
raise ImportError("Please install accelerate via `pip install accelerate`")
device = torch.device("cuda")
for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.safety_checker]:
if cpu_offloaded_model is not None:
cpu_offload(cpu_offloaded_model, device)
@property
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
def _execution_device(self):
r"""
Returns the device on which the pipeline's models will be executed. After calling
`pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
hooks.
"""
if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
return self.device
for module in self.unet.modules():
if (
hasattr(module, "_hf_hook")
and hasattr(module._hf_hook, "execution_device")
and module._hf_hook.execution_device is not None
):
return torch.device(module._hf_hook.execution_device)
return self.device
@torch.no_grad()
def __call__(
self,
......
examples/community/unclip_image_interpolation.py
View file @ fa633ed6
......@@ -19,7 +19,7 @@ from diffusers import (
UNet2DModel,
)
from diffusers.pipelines.unclip import UnCLIPTextProjModel
from diffusers.utils import is_accelerate_available, logging
from diffusers.utils import logging
from diffusers.utils.torch_utils import randn_tensor
......@@ -204,50 +204,6 @@ class UnCLIPImageInterpolationPipeline(DiffusionPipeline):
return image_embeddings
# Copied from diffusers.pipelines.unclip.pipeline_unclip_image_variation.UnCLIPImageVariationPipeline.enable_sequential_cpu_offload
def enable_sequential_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
when their specific submodule has its `forward` method called.
"""
if is_accelerate_available():
from accelerate import cpu_offload
else:
raise ImportError("Please install accelerate via `pip install accelerate`")
device = torch.device(f"cuda:{gpu_id}")
models = [
self.decoder,
self.text_proj,
self.text_encoder,
self.super_res_first,
self.super_res_last,
]
for cpu_offloaded_model in models:
if cpu_offloaded_model is not None:
cpu_offload(cpu_offloaded_model, device)
@property
# Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline._execution_device
def _execution_device(self):
r"""
Returns the device on which the pipeline's models will be executed. After calling
`pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
hooks.
"""
if self.device != torch.device("meta") or not hasattr(self.decoder, "_hf_hook"):
return self.device
for module in self.decoder.modules():
if (
hasattr(module, "_hf_hook")
and hasattr(module._hf_hook, "execution_device")
and module._hf_hook.execution_device is not None
):
return torch.device(module._hf_hook.execution_device)
return self.device
@torch.no_grad()
def __call__(
self,
......
examples/community/unclip_text_interpolation.py
View file @ fa633ed6
......@@ -15,7 +15,7 @@ from diffusers import (
UNet2DModel,
)
from diffusers.pipelines.unclip import UnCLIPTextProjModel
from diffusers.utils import is_accelerate_available, logging
from diffusers.utils import logging
from diffusers.utils.torch_utils import randn_tensor
......@@ -212,51 +212,6 @@ class UnCLIPTextInterpolationPipeline(DiffusionPipeline):
return prompt_embeds, text_encoder_hidden_states, text_mask
# Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.enable_sequential_cpu_offload
def enable_sequential_cpu_offload(self, gpu_id=0):
r"""
Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
when their specific submodule has its `forward` method called.
"""
if is_accelerate_available():
from accelerate import cpu_offload
else:
raise ImportError("Please install accelerate via `pip install accelerate`")
device = torch.device(f"cuda:{gpu_id}")
# TODO: self.prior.post_process_latents is not covered by the offload hooks, so it fails if added to the list
models = [
self.decoder,
self.text_proj,
self.text_encoder,
self.super_res_first,
self.super_res_last,
]
for cpu_offloaded_model in models:
if cpu_offloaded_model is not None:
cpu_offload(cpu_offloaded_model, device)
@property
# Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline._execution_device
def _execution_device(self):
r"""
Returns the device on which the pipeline's models will be executed. After calling
`pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
hooks.
"""
if self.device != torch.device("meta") or not hasattr(self.decoder, "_hf_hook"):
return self.device
for module in self.decoder.modules():
if (
hasattr(module, "_hf_hook")
and hasattr(module._hf_hook, "execution_device")
and module._hf_hook.execution_device is not None
):
return torch.device(module._hf_hook.execution_device)
return self.device
@torch.no_grad()
def __call__(
self,
......
examples/community/wildcard_stable_diffusion.py
View file @ fa633ed6
......@@ -8,9 +8,9 @@ from typing import Callable, Dict, List, Optional, Union
import torch
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
from diffusers import DiffusionPipeline
from diffusers.configuration_utils import FrozenDict
from diffusers.models import AutoencoderKL, UNet2DConditionModel
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
......@@ -63,7 +63,7 @@ class WildcardStableDiffusionOutput(StableDiffusionPipelineOutput):
prompts: List[str]
class WildcardStableDiffusionPipeline(DiffusionPipeline):
class WildcardStableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
r"""
Example Usage:
pipe = WildcardStableDiffusionPipeline.from_pretrained(
......
examples/research_projects/controlnetxs/pipeline_controlnet_xs.py
View file @ fa633ed6
......@@ -26,7 +26,7 @@ from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
from diffusers.models import AutoencoderKL, UNet2DConditionModel
from diffusers.models.lora import adjust_lora_scale_text_encoder
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
from diffusers.schedulers import KarrasDiffusionSchedulers
......@@ -44,7 +44,7 @@ logger = logging.get_logger(__name__) # pylint: disable=invalid-name
class StableDiffusionControlNetXSPipeline(
DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
):
r"""
Pipeline for text-to-image generation using Stable Diffusion with ControlNet-XS guidance.
......@@ -139,39 +139,6 @@ class StableDiffusionControlNetXSPipeline(
)
self.register_to_config(requires_safety_checker=requires_safety_checker)
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
def enable_vae_tiling(self):
r"""
Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
processing larger images.
"""
self.vae.enable_tiling()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
def disable_vae_tiling(self):
r"""
Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
computing decoding in one step.
"""
self.vae.disable_tiling()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
def _encode_prompt(
self,
......@@ -596,34 +563,6 @@ class StableDiffusionControlNetXSPipeline(
latents = latents * self.scheduler.init_noise_sigma
return latents
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
The suffixes after the scaling factors represent the stages where they are being applied.
Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
Args:
s1 (`float`):
Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
mitigate "oversmoothing effect" in the enhanced denoising process.
s2 (`float`):
Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
mitigate "oversmoothing effect" in the enhanced denoising process.
b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
"""
if not hasattr(self, "unet"):
raise ValueError("The pipeline must have `unet` for using FreeU.")
self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
def disable_freeu(self):
"""Disables the FreeU mechanism if enabled."""
self.unet.disable_freeu()
@torch.no_grad()
def __call__(
self,
......
examples/research_projects/controlnetxs/pipeline_controlnet_xs_sd_xl.py
View file @ fa633ed6
......@@ -31,7 +31,7 @@ from diffusers.models.attention_processor import (
XFormersAttnProcessor,
)
from diffusers.models.lora import adjust_lora_scale_text_encoder
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
......@@ -52,7 +52,11 @@ logger = logging.get_logger(__name__) # pylint: disable=invalid-name
class StableDiffusionXLControlNetXSPipeline(
DiffusionPipeline, TextualInversionLoaderMixin, StableDiffusionXLLoraLoaderMixin, FromSingleFileMixin
DiffusionPipeline,
StableDiffusionMixin,
TextualInversionLoaderMixin,
StableDiffusionXLLoraLoaderMixin,
FromSingleFileMixin,
):
r"""
Pipeline for text-to-image generation using Stable Diffusion XL with ControlNet-XS guidance.
......@@ -145,39 +149,6 @@ class StableDiffusionXLControlNetXSPipeline(
self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
def enable_vae_tiling(self):
r"""
Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
processing larger images.
"""
self.vae.enable_tiling()
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
def disable_vae_tiling(self):
r"""
Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
computing decoding in one step.
"""
self.vae.disable_tiling()
# Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
def encode_prompt(
self,
......@@ -661,34 +632,6 @@ class StableDiffusionXLControlNetXSPipeline(
self.vae.decoder.conv_in.to(dtype)
self.vae.decoder.mid_block.to(dtype)
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu
def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
The suffixes after the scaling factors represent the stages where they are being applied.
Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
Args:
s1 (`float`):
Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
mitigate "oversmoothing effect" in the enhanced denoising process.
s2 (`float`):
Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
mitigate "oversmoothing effect" in the enhanced denoising process.
b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
"""
if not hasattr(self, "unet"):
raise ValueError("The pipeline must have `unet` for using FreeU.")
self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu
def disable_freeu(self):
"""Disables the FreeU mechanism if enabled."""
self.unet.disable_freeu()
@torch.no_grad()
def __call__(
self,
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment