Mirrors
/
diffusers
mirror of https://github.com/ShivamShrirao/diffusers.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

update composable diffusion for an updated diffuser library (#1697) 

* update composable diffusion for an updated diffuser library

* fix style/quality for code

* Revert "fix style/quality for code"

This reverts commit 71f23497639fe69de00d93cf91edc31b08dcd7a4.

* update style

* reduce memory usage by computing score sequentially
This commit is contained in:
Nan Liu 2022-12-19 11:03:40 -06:00 committed by GitHub
parent a5edb981a7
commit 6f15026330
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 484 additions and 229 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

52
examples/community/README.md
View File

@ -355,43 +355,45 @@ out = pipe(
import torch as th
import numpy as np
import torchvision.utils as tvu
from diffusers import DiffusionPipeline
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--prompt", type=str, default="mystical trees | A magical pond | dark",
help="use '|' as the delimiter to compose separate sentences.")
parser.add_argument("--steps", type=int, default=50)
parser.add_argument("--scale", type=float, default=7.5)
parser.add_argument("--weights", type=str, default="7.5 | 7.5 | -7.5")
parser.add_argument("--seed", type=int, default=2)
parser.add_argument("--model_path", type=str, default="CompVis/stable-diffusion-v1-4")
parser.add_argument("--num_images", type=int, default=1)
args = parser.parse_args()
has_cuda = th.cuda.is_available()
device = th.device('cpu' if not has_cuda else 'cuda')
prompt = args.prompt
scale = args.scale
steps = args.steps
pipe = DiffusionPipeline.from_pretrained(
"CompVis/stable-diffusion-v1-4",
use_auth_token=True,
args.model_path,
custom_pipeline="composable_stable_diffusion",
).to(device)
def dummy(images, **kwargs):
return images, False
pipe.safety_checker = dummy
pipe.safety_checker = None
images = []
generator = torch.Generator("cuda").manual_seed(0)
generator = th.Generator("cuda").manual_seed(args.seed)
for i in range(args.num_images):
image = pipe(prompt, guidance_scale=scale, num_inference_steps=steps,
weights=args.weights, generator=generator).images[0]
images.append(th.from_numpy(np.array(image)).permute(2, 0, 1) / 255.)
grid = tvu.make_grid(th.stack(images, dim=0), nrow=4, padding=0)
tvu.save_image(grid, f'{prompt}_{args.weights}' + '.png')
seed = 0
prompt = "a forest | a camel"
weights = " 1 | 1" # Equal weight to each prompt. Can be negative
images = []
for i in range(4):
res = pipe(
prompt,
guidance_scale=7.5,
num_inference_steps=50,
weights=weights,
generator=generator)
image = res.images[0]
images.append(image)
for i, img in enumerate(images):
img.save(f"./composable_diffusion/image_{i}.png")
```
### Imagic Stable Diffusion

661
examples/community/composable_stable_diffusion.py
View File

@ -1,25 +1,52 @@
"""
modified based on diffusion library from Huggingface: https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py
"""
# Copyright 2022 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import inspect
import warnings
from typing import List, Optional, Union
from typing import Callable, List, Optional, Union
import torch
from diffusers.models import AutoencoderKL, UNet2DConditionModel
from diffusers.pipeline_utils import DiffusionPipeline
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
from diffusers.utils import is_accelerate_available
from packaging import version
from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
from ...configuration_utils import FrozenDict
from ...models import AutoencoderKL, UNet2DConditionModel
from ...pipeline_utils import DiffusionPipeline
from ...schedulers import (
DDIMScheduler,
DPMSolverMultistepScheduler,
EulerAncestralDiscreteScheduler,
EulerDiscreteScheduler,
LMSDiscreteScheduler,
PNDMScheduler,
)
from ...utils import deprecate, logging
from . import StableDiffusionPipelineOutput
from .safety_checker import StableDiffusionSafetyChecker
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
class ComposableStableDiffusionPipeline(DiffusionPipeline):
r"""
Pipeline for text-to-image generation using Stable Diffusion.
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
Args:
vae ([`AutoencoderKL`]):
Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
@ -35,11 +62,12 @@ class ComposableStableDiffusionPipeline(DiffusionPipeline):
A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
[`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
safety_checker ([`StableDiffusionSafetyChecker`]):
Classification module that estimates whether generated images could be considered offsensive or harmful.
Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details.
Classification module that estimates whether generated images could be considered offensive or harmful.
Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
feature_extractor ([`CLIPFeatureExtractor`]):
Model that extracts features from generated images to be used as inputs for the `safety_checker`.
"""
_optional_components = ["safety_checker", "feature_extractor"]
def __init__(
self,
@ -47,11 +75,84 @@ class ComposableStableDiffusionPipeline(DiffusionPipeline):
text_encoder: CLIPTextModel,
tokenizer: CLIPTokenizer,
unet: UNet2DConditionModel,
scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
scheduler: Union[
DDIMScheduler,
PNDMScheduler,
LMSDiscreteScheduler,
EulerDiscreteScheduler,
EulerAncestralDiscreteScheduler,
DPMSolverMultistepScheduler,
],
safety_checker: StableDiffusionSafetyChecker,
feature_extractor: CLIPFeatureExtractor,
requires_safety_checker: bool = True,
):
super().__init__()
if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
deprecation_message = (
f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
"to update the config accordingly as leaving `steps_offset` might led to incorrect results"
" in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
" it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
" file"
)
deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
new_config = dict(scheduler.config)
new_config["steps_offset"] = 1
scheduler._internal_dict = FrozenDict(new_config)
if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
deprecation_message = (
f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
" `clip_sample` should be set to False in the configuration file. Please make sure to update the"
" config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
" future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
" nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
)
deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
new_config = dict(scheduler.config)
new_config["clip_sample"] = False
scheduler._internal_dict = FrozenDict(new_config)
if safety_checker is None and requires_safety_checker:
logger.warning(
f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
" that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
" results in services or applications open to the public. Both the diffusers team and Hugging Face"
" strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
" it only for use-cases that involve analyzing network behavior or auditing its results. For more"
" information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
)
if safety_checker is not None and feature_extractor is None:
raise ValueError(
"Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
" checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
)
is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
version.parse(unet.config._diffusers_version).base_version
) < version.parse("0.9.0.dev0")
is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
deprecation_message = (
"The configuration file of the unet has set the default `sample_size` to smaller than"
" 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
" following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
" CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
" \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
" configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
" in the config might lead to incorrect results in future versions. If you have downloaded this"
" checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
" the `unet/config.json` file"
)
deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
new_config = dict(unet.config)
new_config["sample_size"] = 64
unet._internal_dict = FrozenDict(new_config)
self.register_modules(
vae=vae,
text_encoder=text_encoder,
@ -61,56 +162,265 @@ class ComposableStableDiffusionPipeline(DiffusionPipeline):
safety_checker=safety_checker,
feature_extractor=feature_extractor,
)
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_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
def enable_vae_slicing(self):
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.
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 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]:
if cpu_offloaded_model is not None:
cpu_offload(cpu_offloaded_model, device)
if self.safety_checker is not None:
# TODO(Patrick) - there is currently a bug with cpu offload of nn.Parameter in accelerate
# fix by only offloading self.safety_checker for now
cpu_offload(self.safety_checker.vision_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.
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`.
prompt (`str` or `list(int)`):
prompt to be encoded
device: (`torch.device`):
torch device
num_images_per_prompt (`int`):
number of images that should be generated per prompt
do_classifier_free_guidance (`bool`):
whether to use classifier free guidance or not
negative_prompt (`str` or `List[str]`):
The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
if `guidance_scale` is less than `1`).
"""
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)
batch_size = len(prompt) if isinstance(prompt, list) else 1
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)
text_inputs = self.tokenizer(
prompt,
padding="max_length",
max_length=self.tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
logger.warning(
"The following part of your input was truncated because CLIP can only handle sequences up to"
f" {self.tokenizer.model_max_length} tokens: {removed_text}"
)
if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
attention_mask = text_inputs.attention_mask.to(device)
else:
attention_mask = None
text_embeddings = self.text_encoder(
text_input_ids.to(device),
attention_mask=attention_mask,
)
text_embeddings = text_embeddings[0]
# duplicate text embeddings for each generation per prompt, using mps friendly method
bs_embed, seq_len, _ = text_embeddings.shape
text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
# get unconditional embeddings for classifier free guidance
if do_classifier_free_guidance:
uncond_tokens: List[str]
if negative_prompt is None:
uncond_tokens = [""] * batch_size
elif type(prompt) is not type(negative_prompt):
raise TypeError(
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
f" {type(prompt)}."
)
elif isinstance(negative_prompt, str):
uncond_tokens = [negative_prompt]
elif batch_size != len(negative_prompt):
raise ValueError(
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
" the batch size of `prompt`."
)
else:
uncond_tokens = negative_prompt
max_length = text_input_ids.shape[-1]
uncond_input = self.tokenizer(
uncond_tokens,
padding="max_length",
max_length=max_length,
truncation=True,
return_tensors="pt",
)
if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
attention_mask = uncond_input.attention_mask.to(device)
else:
attention_mask = None
uncond_embeddings = self.text_encoder(
uncond_input.input_ids.to(device),
attention_mask=attention_mask,
)
uncond_embeddings = uncond_embeddings[0]
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
seq_len = uncond_embeddings.shape[1]
uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
# For classifier free guidance, we need to do two forward passes.
# Here we concatenate the unconditional and text embeddings into a single batch
# to avoid doing two forward passes
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
return text_embeddings
def run_safety_checker(self, image, device, dtype):
if self.safety_checker is not None:
safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device)
image, has_nsfw_concept = self.safety_checker(
images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
)
else:
has_nsfw_concept = None
return image, has_nsfw_concept
def decode_latents(self, latents):
latents = 1 / 0.18215 * latents
image = self.vae.decode(latents).sample
image = (image / 2 + 0.5).clamp(0, 1)
# we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
return image
def prepare_extra_step_kwargs(self, generator, eta):
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
# and should be between [0, 1]
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
extra_step_kwargs = {}
if accepts_eta:
extra_step_kwargs["eta"] = eta
# check if the scheduler accepts generator
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
if accepts_generator:
extra_step_kwargs["generator"] = generator
return extra_step_kwargs
def check_inputs(self, prompt, height, width, callback_steps):
if not isinstance(prompt, str) and not isinstance(prompt, list):
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
if height % 8 != 0 or width % 8 != 0:
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
if (callback_steps is None) or (
callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
):
raise ValueError(
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
f" {type(callback_steps)}."
)
def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
if latents is None:
if device.type == "mps":
# randn does not work reproducibly on mps
latents = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device)
else:
latents = torch.randn(shape, generator=generator, device=device, dtype=dtype)
else:
if latents.shape != shape:
raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
latents = latents.to(device)
# scale the initial noise by the standard deviation required by the scheduler
latents = latents * self.scheduler.init_noise_sigma
return latents
@torch.no_grad()
def __call__(
self,
prompt: Union[str, List[str]],
height: Optional[int] = 512,
width: Optional[int] = 512,
num_inference_steps: Optional[int] = 50,
guidance_scale: Optional[float] = 7.5,
eta: Optional[float] = 0.0,
height: Optional[int] = None,
width: Optional[int] = None,
num_inference_steps: int = 50,
guidance_scale: float = 7.5,
negative_prompt: Optional[Union[str, List[str]]] = None,
num_images_per_prompt: Optional[int] = 1,
eta: float = 0.0,
generator: Optional[torch.Generator] = None,
latents: Optional[torch.FloatTensor] = None,
output_type: Optional[str] = "pil",
return_dict: bool = True,
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
callback_steps: Optional[int] = 1,
weights: Optional[str] = "",
**kwargs,
):
r"""
Function invoked when calling the pipeline for generation.
Args:
prompt (`str` or `List[str]`):
The prompt or prompts to guide the image generation.
height (`int`, *optional*, defaults to 512):
height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
The height in pixels of the generated image.
width (`int`, *optional*, defaults to 512):
width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
The width in pixels of the generated image.
num_inference_steps (`int`, *optional*, defaults to 50):
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
@ -121,6 +431,11 @@ class ComposableStableDiffusionPipeline(DiffusionPipeline):
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
usually at the expense of lower image quality.
negative_prompt (`str` or `List[str]`, *optional*):
The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
if `guidance_scale` is less than `1`).
num_images_per_prompt (`int`, *optional*, defaults to 1):
The number of images to generate per prompt.
eta (`float`, *optional*, defaults to 0.0):
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
[`schedulers.DDIMScheduler`], will be ignored for others.
@ -137,6 +452,13 @@ class ComposableStableDiffusionPipeline(DiffusionPipeline):
return_dict (`bool`, *optional*, defaults to `True`):
Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
plain tuple.
callback (`Callable`, *optional*):
A function that will be called every `callback_steps` steps during inference. The function will be
called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
callback_steps (`int`, *optional*, defaults to 1):
The frequency at which the `callback` function will be called. If not specified, the callback will be
called at every step.
Returns:
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
@ -144,182 +466,113 @@ class ComposableStableDiffusionPipeline(DiffusionPipeline):
list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
(nsfw) content, according to the `safety_checker`.
"""
# 0. Default height and width to unet
height = height or self.unet.config.sample_size * self.vae_scale_factor
width = width or self.unet.config.sample_size * self.vae_scale_factor
if "torch_device" in kwargs:
device = kwargs.pop("torch_device")
warnings.warn(
"`torch_device` is deprecated as an input argument to `__call__` and will be removed in v0.3.0."
" Consider using `pipe.to(torch_device)` instead."
)
# 1. Check inputs. Raise error if not correct
self.check_inputs(prompt, height, width, callback_steps)
# Set device as before (to be removed in 0.3.0)
if device is None:
device = "cuda" if torch.cuda.is_available() else "cpu"
self.to(device)
if isinstance(prompt, str):
batch_size = 1
elif isinstance(prompt, list):
batch_size = len(prompt)
else:
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
if height % 8 != 0 or width % 8 != 0:
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
# 2. Define call parameters
batch_size = 1 if isinstance(prompt, str) else len(prompt)
device = self._execution_device
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
# corresponds to doing no classifier free guidance.
do_classifier_free_guidance = guidance_scale > 1.0
if "|" in prompt:
prompt = [x.strip() for x in prompt.split("|")]
print(f"composing {prompt}...")
# get prompt text embeddings
text_input = self.tokenizer(
prompt,
padding="max_length",
max_length=self.tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_embeddings = self.text_encoder(text_input.input_ids.to(self.device))[0]
if not weights:
# specify weights for prompts (excluding the unconditional score)
print("using equal weights for all prompts...")
pos_weights = torch.tensor(
[1 / (text_embeddings.shape[0] - 1)] * (text_embeddings.shape[0] - 1), device=self.device
).reshape(-1, 1, 1, 1)
neg_weights = torch.tensor([1.0], device=self.device).reshape(-1, 1, 1, 1)
mask = torch.tensor([False] + [True] * pos_weights.shape[0], dtype=torch.bool)
else:
# set prompt weight for each
num_prompts = len(prompt) if isinstance(prompt, list) else 1
weights = [float(w.strip()) for w in weights.split("|")]
if len(weights) < num_prompts:
weights.append(1.0)
weights = torch.tensor(weights, device=self.device)
assert len(weights) == text_embeddings.shape[0], "weights specified are not equal to the number of prompts"
pos_weights = []
neg_weights = []
mask = [] # first one is unconditional score
for w in weights:
if w > 0:
pos_weights.append(w)
mask.append(True)
else:
neg_weights.append(abs(w))
mask.append(False)
# normalize the weights
pos_weights = torch.tensor(pos_weights, device=self.device).reshape(-1, 1, 1, 1)
pos_weights = pos_weights / pos_weights.sum()
neg_weights = torch.tensor(neg_weights, device=self.device).reshape(-1, 1, 1, 1)
neg_weights = neg_weights / neg_weights.sum()
mask = torch.tensor(mask, device=self.device, dtype=torch.bool)
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
# corresponds to doing no classifier free guidance.
do_classifier_free_guidance = guidance_scale > 1.0
# get unconditional embeddings for classifier free guidance
if do_classifier_free_guidance:
max_length = text_input.input_ids.shape[-1]
if torch.all(mask):
# no negative prompts, so we use empty string as the negative prompt
uncond_input = self.tokenizer(
[""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt"
)
uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
# For classifier free guidance, we need to do two forward passes.
# Here we concatenate the unconditional and text embeddings into a single batch
# to avoid doing two forward passes
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
# update negative weights
neg_weights = torch.tensor([1.0], device=self.device)
mask = torch.tensor([False] + mask.detach().tolist(), device=self.device, dtype=torch.bool)
# get the initial random noise unless the user supplied it
# Unlike in other pipelines, latents need to be generated in the target device
# for 1-to-1 results reproducibility with the CompVis implementation.
# However this currently doesn't work in `mps`.
latents_device = "cpu" if self.device.type == "mps" else self.device
latents_shape = (batch_size, self.unet.in_channels, height // 8, width // 8)
if latents is None:
latents = torch.randn(
latents_shape,
generator=generator,
device=latents_device,
)
else:
if latents.shape != latents_shape:
raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
latents = latents.to(self.device)
# set timesteps
accepts_offset = "offset" in set(inspect.signature(self.scheduler.set_timesteps).parameters.keys())
extra_set_kwargs = {}
if accepts_offset:
extra_set_kwargs["offset"] = 1
self.scheduler.set_timesteps(num_inference_steps, **extra_set_kwargs)
# if we use LMSDiscreteScheduler, let's make sure latents are multiplied by sigmas
if isinstance(self.scheduler, LMSDiscreteScheduler):
latents = latents * self.scheduler.sigmas[0]
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
# and should be between [0, 1]
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
extra_step_kwargs = {}
if accepts_eta:
extra_step_kwargs["eta"] = eta
for i, t in enumerate(self.progress_bar(self.scheduler.timesteps)):
# expand the latents if we are doing classifier free guidance
latent_model_input = (
torch.cat([latents] * text_embeddings.shape[0]) if do_classifier_free_guidance else latents
)
if isinstance(self.scheduler, LMSDiscreteScheduler):
sigma = self.scheduler.sigmas[i]
# the model input needs to be scaled to match the continuous ODE formulation in K-LMS
latent_model_input = latent_model_input / ((sigma**2 + 1) ** 0.5)
# reduce memory by predicting each score sequentially
noise_preds = []
# predict the noise residual
for latent_in, text_embedding_in in zip(
torch.chunk(latent_model_input, chunks=latent_model_input.shape[0], dim=0),
torch.chunk(text_embeddings, chunks=text_embeddings.shape[0], dim=0),
):
noise_preds.append(self.unet(latent_in, t, encoder_hidden_states=text_embedding_in).sample)
noise_preds = torch.cat(noise_preds, dim=0)
# perform guidance
if do_classifier_free_guidance:
noise_pred_uncond = (noise_preds[~mask] * neg_weights).sum(dim=0, keepdims=True)
noise_pred_text = (noise_preds[mask] * pos_weights).sum(dim=0, keepdims=True)
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
# compute the previous noisy sample x_t -> x_t-1
if isinstance(self.scheduler, LMSDiscreteScheduler):
latents = self.scheduler.step(noise_pred, i, latents, **extra_step_kwargs).prev_sample
if not weights:
# specify weights for prompts (excluding the unconditional score)
print("using equal positive weights (conjunction) for all prompts...")
weights = torch.tensor([guidance_scale] * len(prompt), device=self.device).reshape(-1, 1, 1, 1)
else:
# set prompt weight for each
num_prompts = len(prompt) if isinstance(prompt, list) else 1
weights = [float(w.strip()) for w in weights.split("|")]
# guidance scale as the default
if len(weights) < num_prompts:
weights.append(guidance_scale)
else:
weights = weights[:num_prompts]
assert len(weights) == len(prompt), "weights specified are not equal to the number of prompts"
weights = torch.tensor(weights, device=self.device).reshape(-1, 1, 1, 1)
else:
weights = guidance_scale
# 3. Encode input prompt
text_embeddings = self._encode_prompt(
prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
)
# 4. Prepare timesteps
self.scheduler.set_timesteps(num_inference_steps, device=device)
timesteps = self.scheduler.timesteps
# 5. Prepare latent variables
num_channels_latents = self.unet.in_channels
latents = self.prepare_latents(
batch_size * num_images_per_prompt,
num_channels_latents,
height,
width,
text_embeddings.dtype,
device,
generator,
latents,
)
# composable diffusion
if isinstance(prompt, list) and batch_size == 1:
# remove extra unconditional embedding
# N = one unconditional embed + conditional embeds
text_embeddings = text_embeddings[len(prompt) - 1 :]
# 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
# 7. Denoising loop
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
with self.progress_bar(total=num_inference_steps) as progress_bar:
for i, t in enumerate(timesteps):
# expand the latents if we are doing classifier free guidance
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
# predict the noise residual
noise_pred = []
for j in range(text_embeddings.shape[0]):
noise_pred.append(
self.unet(latent_model_input[:1], t, encoder_hidden_states=text_embeddings[j : j + 1]).sample
)
noise_pred = torch.cat(noise_pred, dim=0)
# perform guidance
if do_classifier_free_guidance:
noise_pred_uncond, noise_pred_text = noise_pred[:1], noise_pred[1:]
noise_pred = noise_pred_uncond + (weights * (noise_pred_text - noise_pred_uncond)).sum(
dim=0, keepdims=True
)
# compute the previous noisy sample x_t -> x_t-1
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
# scale and decode the image latents with vae
latents = 1 / 0.18215 * latents
image = self.vae.decode(latents).sample
# call the callback, if provided
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
progress_bar.update()
if callback is not None and i % callback_steps == 0:
callback(i, t, latents)
image = (image / 2 + 0.5).clamp(0, 1)
image = image.cpu().permute(0, 2, 3, 1).numpy()
# 8. Post-processing
image = self.decode_latents(latents)
# run safety checker
safety_cheker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(self.device)
image, has_nsfw_concept = self.safety_checker(images=image, clip_input=safety_cheker_input.pixel_values)
# 9. Run safety checker
image, has_nsfw_concept = self.run_safety_checker(image, device, text_embeddings.dtype)
# 10. Convert to PIL
if output_type == "pil":
image = self.numpy_to_pil(image)