mirror of https://github.com/vladmandic/automatic
872 lines
40 KiB
Python
872 lines
40 KiB
Python
# Credits: @ukaprch <https://github.com/huggingface/diffusers/issues/9924>
|
|
|
|
import math
|
|
from dataclasses import dataclass
|
|
from typing import List, Optional, Tuple, Union
|
|
|
|
import numpy as np
|
|
import torch
|
|
import torchsde
|
|
|
|
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
|
from diffusers.utils import BaseOutput
|
|
from diffusers.utils.torch_utils import randn_tensor
|
|
from diffusers.schedulers.scheduling_utils import SchedulerMixin
|
|
import scipy.stats
|
|
|
|
|
|
class BatchedBrownianTree:
|
|
"""A wrapper around torchsde.BrownianTree that enables batches of entropy."""
|
|
|
|
def __init__(self, x, t0, t1, seed=None, **kwargs):
|
|
t0, t1, self.sign = self.sort(t0, t1)
|
|
w0 = kwargs.get("w0", torch.zeros_like(x))
|
|
if seed is None:
|
|
seed = [torch.randint(0, 2**63 - 1, []).item()]
|
|
seed = [s.initial_seed() if isinstance(s, torch.Generator) else s for s in seed]
|
|
self.batched = True
|
|
try:
|
|
assert len(seed) == x.shape[0]
|
|
w0 = w0[0]
|
|
except TypeError:
|
|
seed = [seed]
|
|
self.batched = False
|
|
self.trees = [
|
|
torchsde.BrownianInterval(
|
|
t0=t0,
|
|
t1=t1,
|
|
size=w0.shape,
|
|
dtype=w0.dtype,
|
|
device=w0.device,
|
|
entropy=s,
|
|
tol=1e-6,
|
|
pool_size=24,
|
|
halfway_tree=True,
|
|
)
|
|
for s in seed
|
|
]
|
|
|
|
@staticmethod
|
|
def sort(a, b):
|
|
return (a, b, 1) if a < b else (b, a, -1)
|
|
|
|
def __call__(self, t0, t1):
|
|
t0, t1, sign = self.sort(t0, t1)
|
|
w = torch.stack([tree(t0, t1) for tree in self.trees]) * (self.sign * sign)
|
|
return w if self.batched else w[0]
|
|
|
|
class BrownianTreeNoiseSampler:
|
|
"""A noise sampler backed by a torchsde.BrownianTree.
|
|
|
|
Args:
|
|
x (Tensor): The tensor whose shape, device and dtype to use to generate
|
|
random samples.
|
|
sigma_min (float): The low end of the valid interval.
|
|
sigma_max (float): The high end of the valid interval.
|
|
seed (int or List[int]): The random seed. If a list of seeds is
|
|
supplied instead of a single integer, then the noise sampler will use one BrownianTree per batch item, each
|
|
with its own seed.
|
|
transform (callable): A function that maps sigma to the sampler's
|
|
internal timestep.
|
|
"""
|
|
|
|
def __init__(self, x, sigma_min, sigma_max, seed=None, transform=lambda x: x):
|
|
self.transform = transform
|
|
t0, t1 = self.transform(torch.as_tensor(sigma_min)), self.transform(torch.as_tensor(sigma_max))
|
|
self.tree = BatchedBrownianTree(x, t0, t1, seed)
|
|
|
|
def __call__(self, sigma, sigma_next):
|
|
t0, t1 = self.transform(torch.as_tensor(sigma)), self.transform(torch.as_tensor(sigma_next))
|
|
return self.tree(t0, t1) / (t1 - t0).abs().sqrt()
|
|
|
|
@dataclass
|
|
class FlowMatchDPMSolverMultistepSchedulerOutput(BaseOutput):
|
|
"""
|
|
Output class for the scheduler's `step` function output.
|
|
|
|
Args:
|
|
prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
|
|
Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
|
|
denoising loop.
|
|
"""
|
|
|
|
prev_sample: torch.FloatTensor
|
|
|
|
class FlowMatchDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
|
|
"""
|
|
`DPMSolverMultistepScheduler` is a fast dedicated high-order solver for diffusion ODEs.
|
|
|
|
This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
|
|
methods the library implements for all schedulers such as loading and saving.
|
|
|
|
Args:
|
|
num_train_timesteps (`int`, defaults to 1000):
|
|
The number of diffusion steps to train the model.
|
|
beta_start (`float`, defaults to 0.0001):
|
|
The starting `beta` value of inference.
|
|
beta_end (`float`, defaults to 0.02):
|
|
The final `beta` value.
|
|
beta_schedule (`str`, defaults to `"scaled linear"`):
|
|
The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from `linear` or `scaled_linear`.
|
|
trained_betas (`np.ndarray`, *optional*):
|
|
Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
|
|
solver_order (`int`, defaults to 2):
|
|
The DPMSolver order which can be `2` or `3`. It is recommended to use `solver_order=2` for guided
|
|
sampling, and `solver_order=3` for unconditional sampling.
|
|
algorithm_type (`str`, defaults to `dpmsolver++2M`):
|
|
Algorithm type for the solver; can be `dpmsolver2`, `dpmsolver2A`, `dpmsolver++2M`, `dpmsolver++2S`, `dpmsolver++sde`, `dpmsolver++2Msde`,
|
|
or `dpmsolver++3Msde`.
|
|
solver_type (`str`, defaults to `midpoint`):
|
|
Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the
|
|
sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers.
|
|
sigma_schedule (`str`, *optional*, defaults to None (beta)): Sigma schedule to compute the `sigmas`. Optionally, we use
|
|
the schedule "karras" introduced in the EDM paper (https://arxiv.org/abs/2206.00364). Other acceptable values are
|
|
"exponential". The exponential schedule was incorporated in this model: https://huggingface.co/stabilityai/cosxl.
|
|
Other acceptable values are "lambdas". The uniform-logSNR for step sizes proposed by Lu's DPM-Solver in the
|
|
noise schedule during the sampling process. The sigmas and time steps are determined according to a sequence of `lambda(t)`.
|
|
"betas" for step sizes in the noise schedule during the sampling process. Refer to [Beta
|
|
Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information.
|
|
use_noise_sampler for BrownianTreeNoiseSampler (only valid for `dpmsolver++2S`, `dpmsolver++sde`, `dpmsolver++2Msde`, or `dpmsolver++3Msde`.
|
|
A noise sampler backed by a torchsde increasing the stability of convergence. Default strategy
|
|
(random noise) has it jumping all over the place, but Brownian sampling is more stable. Utilizes the model generation seed provided.
|
|
midpoint_ratio (`float`, *optional*, range: 0.4 to 0.6, default=0.5): Only valid for (`dpmsolver++sde`, `dpmsolver++2S`).
|
|
Higher values may result in smoothing, more vivid colors and less noise at the expense of more detail and effect.
|
|
s_noise (`float`, *optional*, defaults to 1.0): Sigma noise strength: range 0 - 1.1 (only valid for `dpmsolver++2S`, `dpmsolver++sde`,
|
|
`dpmsolver++2Msde`, or `dpmsolver++3Msde`). The amount of additional noise to counteract loss of detail during sampling. A
|
|
reasonable range is [1.000, 1.011]. Defaults to 1.0 from the original implementation.
|
|
use_beta_sigmas: (`bool` defaults to False for FLUX and True for SD3). Based on original interpretation of using beta values for determining sigmas.
|
|
use_dynamic_shifting (`bool` defaults to False for SD3 and True for FLUX). When `True`, shift is ignored.
|
|
shift (`float`, defaults to 3.0): The shift value for the timestep schedule for SD3 when not using dynamic shifting.
|
|
The remaining args are specific to Flux's dynamic shifting based on resolution.
|
|
"""
|
|
|
|
_compatibles = []
|
|
order = 1
|
|
|
|
@register_to_config
|
|
def __init__(
|
|
self,
|
|
num_train_timesteps: int = 1000,
|
|
beta_start: float = 0.00085,
|
|
beta_end: float = 0.012,
|
|
beta_schedule: str = "scaled linear",
|
|
trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
|
|
solver_order: int = 2,
|
|
algorithm_type: str = "dpmsolver++2M",
|
|
solver_type: str = "midpoint",
|
|
sigma_schedule: Optional[str] = None,
|
|
shift: float = 3.0,
|
|
midpoint_ratio: Optional[float] = 0.5,
|
|
s_noise: Optional[float] = 1.0,
|
|
use_noise_sampler: Optional[bool] = True,
|
|
use_beta_sigmas: Optional[bool] = False,
|
|
use_dynamic_shifting=False,
|
|
base_shift: Optional[float] = 0.5,
|
|
max_shift: Optional[float] = 1.15,
|
|
base_image_seq_len: Optional[int] = 256,
|
|
max_image_seq_len: Optional[int] = 4096,
|
|
):
|
|
# settings for DPM-Solver
|
|
if algorithm_type not in ["dpmsolver2", "dpmsolver2A", "dpmsolver++2M", "dpmsolver++2S", "dpmsolver++sde", "dpmsolver++2Msde", "dpmsolver++3Msde"]:
|
|
raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}")
|
|
|
|
if solver_type not in ["midpoint", "heun"]:
|
|
raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}")
|
|
|
|
if sigma_schedule not in [None, "karras", "exponential", "lambdas", "betas"]:
|
|
raise NotImplementedError(f"{sigma_schedule} is not implemented for {self.__class__}")
|
|
|
|
if beta_schedule not in ["linear", "scaled linear"]:
|
|
raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}")
|
|
|
|
# setable values
|
|
timesteps = np.linspace(1, num_train_timesteps, num_train_timesteps, dtype=np.float32)[::-1].copy()
|
|
timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32)
|
|
|
|
sigmas = timesteps / num_train_timesteps
|
|
if not use_dynamic_shifting:
|
|
# when use_dynamic_shifting is True, we apply the timestep shifting on the fly based on the image resolution
|
|
sigmas = shift * sigmas / (1 + (shift - 1) * sigmas)
|
|
|
|
self.timesteps = sigmas * num_train_timesteps
|
|
self.h_last = None
|
|
self.h_1 = None
|
|
self.h_2 = None
|
|
self.noise_sampler = None
|
|
self._step_index = None
|
|
self._begin_index = None
|
|
self.sigmas = sigmas.to("cpu") # to avoid too much CPU/GPU communication
|
|
self.model_outputs = [None] * solver_order
|
|
|
|
@property
|
|
def step_index(self):
|
|
"""
|
|
The index counter for current timestep. It will increase 1 after each scheduler step.
|
|
"""
|
|
return self._step_index
|
|
|
|
@property
|
|
def begin_index(self):
|
|
"""
|
|
The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
|
|
"""
|
|
return self._begin_index
|
|
|
|
def set_begin_index(self, begin_index: int = 0):
|
|
"""
|
|
Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
|
|
|
|
Args:
|
|
begin_index (`int`):
|
|
The begin index for the scheduler.
|
|
"""
|
|
self._begin_index = begin_index
|
|
|
|
def time_shift(self, mu: float, sigma: float, t: torch.FloatTensor):
|
|
return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
|
|
|
|
def set_timesteps(self,
|
|
num_inference_steps: int = None,
|
|
device: Union[str, torch.device] = None,
|
|
sigmas: Optional[List[float]] = None,
|
|
mu: Optional[float] = None,
|
|
timesteps: Optional[torch.Tensor] = None,
|
|
):
|
|
"""
|
|
Sets the discrete timesteps used for the diffusion chain (to be run before inference).
|
|
|
|
Args:
|
|
num_inference_steps (`int`):
|
|
The number of diffusion steps used when generating samples with a pre-trained model.
|
|
device (`str` or `torch.device`, *optional*):
|
|
The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
|
|
"""
|
|
if self.config.use_dynamic_shifting and mu is None:
|
|
raise ValueError(" you have a pass a value for `mu` when `use_dynamic_shifting` is set to be `True`")
|
|
|
|
if sigmas is None:
|
|
self.use_beta_sigmas = True
|
|
self.num_inference_steps = num_inference_steps
|
|
beta_start = self.config.beta_start
|
|
beta_end = self.config.beta_end
|
|
if self.config.trained_betas is not None:
|
|
betas = torch.tensor(self.config.trained_betas, dtype=torch.float64)
|
|
elif self.config.beta_schedule == "linear":
|
|
betas = torch.linspace(beta_start, beta_end, self.config.num_train_timesteps, dtype=torch.float64)
|
|
elif self.config.beta_schedule == "scaled linear":
|
|
# this schedule is very specific to the latent diffusion model.
|
|
betas = torch.linspace(beta_start**0.5, beta_end**0.5, self.config.num_train_timesteps, dtype=torch.float64) ** 2
|
|
else:
|
|
raise NotImplementedError(f"{self.config.beta_schedule} is not implemented for {self.__class__}")
|
|
alphas = 1.0 - betas
|
|
alphas_cumprod = torch.cumprod(alphas, dim=0)
|
|
sigmas = np.array(((1 - alphas_cumprod) / alphas_cumprod) ** 0.5)
|
|
del alphas_cumprod
|
|
del alphas
|
|
del betas
|
|
elif self.use_beta_sigmas:
|
|
num_inference_steps = len(sigmas)
|
|
self.num_inference_steps = num_inference_steps
|
|
beta_start = self.config.beta_start
|
|
beta_end = self.config.beta_end
|
|
if self.config.trained_betas is not None:
|
|
betas = torch.tensor(self.config.trained_betas, dtype=torch.float64)
|
|
elif self.config.beta_schedule == "linear":
|
|
betas = torch.linspace(beta_start, beta_end, self.config.num_train_timesteps, dtype=torch.float64)
|
|
elif self.config.beta_schedule == "scaled linear":
|
|
# this schedule is very specific to the latent diffusion model.
|
|
betas = torch.linspace(beta_start**0.5, beta_end**0.5, self.config.num_train_timesteps, dtype=torch.float64) ** 2
|
|
else:
|
|
raise NotImplementedError(f"{self.config.beta_schedule} is not implemented for {self.__class__}")
|
|
alphas = 1.0 - betas
|
|
alphas_cumprod = torch.cumprod(alphas, dim=0)
|
|
sigmas = np.array(((1 - alphas_cumprod) / alphas_cumprod) ** 0.5)
|
|
del alphas_cumprod
|
|
del alphas
|
|
del betas
|
|
else:
|
|
num_inference_steps = len(sigmas)
|
|
self.num_inference_steps = num_inference_steps
|
|
|
|
if self.config.sigma_schedule == "exponential":
|
|
if self.use_beta_sigmas:
|
|
sigmas = np.flip(sigmas).copy()
|
|
sigma_min = sigmas[-1]
|
|
sigma_max = sigmas[0]
|
|
sigmas = self._convert_to_exponential(sigma_min, sigma_max, num_inference_steps=num_inference_steps)
|
|
OldRange = sigma_max - sigma_min
|
|
NewRange = 1.0 - sigma_min
|
|
sigmas = (((sigmas - sigma_min) * NewRange) / OldRange) + sigma_min
|
|
else:
|
|
sigma_min = sigmas[-1]
|
|
sigma_max = sigmas[0]
|
|
sigmas = self._convert_to_exponential(sigma_min, sigma_max, num_inference_steps=num_inference_steps)
|
|
elif self.config.sigma_schedule == "karras":
|
|
if self.use_beta_sigmas:
|
|
sigmas = np.flip(sigmas).copy()
|
|
sigma_min = sigmas[-1]
|
|
sigma_max = sigmas[0]
|
|
sigmas = self._convert_to_karras(sigma_min, sigma_max, num_inference_steps=num_inference_steps)
|
|
OldRange = sigma_max - sigma_min
|
|
NewRange = 1.0 - sigma_min
|
|
sigmas = (((sigmas - sigma_min) * NewRange) / OldRange) + sigma_min
|
|
else:
|
|
sigma_min = sigmas[-1]
|
|
sigma_max = sigmas[0]
|
|
sigmas = self._convert_to_karras(sigma_min, sigma_max, num_inference_steps=num_inference_steps)
|
|
sigmas = torch.from_numpy(sigmas).to(dtype=torch.float64, device=device)
|
|
elif self.config.sigma_schedule == "lambdas":
|
|
if self.use_beta_sigmas:
|
|
log_sigmas = np.log(sigmas)
|
|
lambdas = np.flip(log_sigmas.copy())
|
|
lambdas = self._convert_to_lu(in_lambdas=lambdas, num_inference_steps=num_inference_steps)
|
|
sigmas = np.exp(lambdas)
|
|
sigma_min = sigmas[-1]
|
|
sigma_max = sigmas[0]
|
|
OldRange = sigma_max - sigma_min
|
|
NewRange = 1.0 - sigma_min
|
|
sigmas = (((sigmas - sigma_min) * NewRange) / OldRange) + sigma_min
|
|
del lambdas
|
|
del log_sigmas
|
|
else:
|
|
log_sigmas = np.log(sigmas)
|
|
lambdas = log_sigmas.copy()
|
|
lambdas = self._convert_to_lu(in_lambdas=lambdas, num_inference_steps=num_inference_steps)
|
|
sigmas = np.exp(lambdas)
|
|
del lambdas
|
|
del log_sigmas
|
|
sigmas = torch.from_numpy(sigmas).to(dtype=torch.float64, device=device)
|
|
elif self.config.sigma_schedule == "betas":
|
|
if self.use_beta_sigmas:
|
|
sigmas = np.flip(sigmas).copy()
|
|
sigma_min = sigmas[-1]
|
|
sigma_max = sigmas[0]
|
|
sigmas = self._convert_to_beta(sigma_min, sigma_max, num_inference_steps=num_inference_steps, device=device)
|
|
OldRange = sigma_max - sigma_min
|
|
NewRange = 1.0 - sigma_min
|
|
sigmas = (((sigmas - sigma_min) * NewRange) / OldRange) + sigma_min
|
|
else:
|
|
sigmas = np.flip(sigmas).copy()
|
|
sigma_min = sigmas[-1]
|
|
sigmas = np.linspace(1.0, sigma_min, num_inference_steps)
|
|
sigmas = torch.from_numpy(sigmas).to(dtype=torch.float64, device=device)
|
|
else:
|
|
if self.use_beta_sigmas:
|
|
sigmas = np.flip(sigmas).copy()
|
|
sigma_min = sigmas[-1]
|
|
sigmas = np.linspace(1.0, sigma_min, num_inference_steps)
|
|
sigmas = torch.from_numpy(sigmas).to(dtype=torch.float64, device=device)
|
|
|
|
if self.config.use_dynamic_shifting:
|
|
sigmas = self.time_shift(mu, 1.0, sigmas)
|
|
else:
|
|
sigmas = self.config.shift * sigmas / (1 + (self.config.shift - 1) * sigmas)
|
|
|
|
timesteps = sigmas * self.config.num_train_timesteps
|
|
self.timesteps = timesteps.to(device=device)
|
|
self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
|
|
self.h_last = None
|
|
self.h_1 = None
|
|
self.h_2 = None
|
|
self.noise_sampler = None
|
|
self.model_outputs = [None] * self.config.solver_order
|
|
self._step_index = None
|
|
self._begin_index = None
|
|
|
|
# Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta
|
|
def _convert_to_beta(self, sigma_min, sigma_max, num_inference_steps, device: Union[str, torch.device] = None, alpha: float = 0.6, beta: float = 0.6) -> torch.Tensor:
|
|
"""From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)"""
|
|
sigmas = torch.Tensor(
|
|
[
|
|
sigma_min + (ppf * (sigma_max - sigma_min))
|
|
for ppf in [
|
|
scipy.stats.beta.ppf(timestep, alpha, beta)
|
|
for timestep in 1 - np.linspace(0, 1, num_inference_steps).astype(np.float64)
|
|
]
|
|
]
|
|
).to(dtype=torch.float64, device=device)
|
|
return sigmas
|
|
|
|
def _convert_to_lu(self, in_lambdas: torch.Tensor, num_inference_steps) -> torch.Tensor:
|
|
"""Constructs the noise schedule of Lu et al. (2022)."""
|
|
|
|
lambda_min: float = in_lambdas[-1].item()
|
|
lambda_max: float = in_lambdas[0].item()
|
|
|
|
rho = 1.0 # 1.0 is the value used in the paper
|
|
ramp = np.linspace(0, 1, num_inference_steps)
|
|
min_inv_rho = lambda_min ** (1 / rho)
|
|
max_inv_rho = lambda_max ** (1 / rho)
|
|
lambdas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
|
|
return lambdas
|
|
|
|
# Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
|
|
def _convert_to_karras(self, sigma_min, sigma_max, num_inference_steps) -> torch.Tensor:
|
|
rho = 7.0 # 7.0 is the value used in the paper
|
|
ramp = np.linspace(0, 1, num_inference_steps)
|
|
min_inv_rho = sigma_min ** (1 / rho)
|
|
max_inv_rho = sigma_max ** (1 / rho)
|
|
sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
|
|
return sigmas
|
|
|
|
def _convert_to_exponential(self, sigma_min, sigma_max, num_inference_steps) -> torch.Tensor:
|
|
sigmas = torch.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps).exp()
|
|
return sigmas
|
|
|
|
def index_for_timestep(self, timestep, schedule_timesteps=None):
|
|
if schedule_timesteps is None:
|
|
schedule_timesteps = self.timesteps
|
|
|
|
indices = (schedule_timesteps == timestep).nonzero()
|
|
|
|
# The sigma index that is taken for the **very** first `step`
|
|
# is always the second index (or the last index if there is only 1)
|
|
# This way we can ensure we don't accidentally skip a sigma in
|
|
# case we start in the middle of the denoising schedule (e.g. for image-to-image)
|
|
pos = 1 if len(indices) > 1 else 0
|
|
|
|
return indices[pos].item()
|
|
|
|
def _init_step_index(self, timestep):
|
|
if self.begin_index is None:
|
|
if isinstance(timestep, torch.Tensor):
|
|
timestep = timestep.to(self.timesteps.device)
|
|
self._step_index = self.index_for_timestep(timestep)
|
|
else:
|
|
self._step_index = self._begin_index
|
|
|
|
def step(
|
|
self,
|
|
model_output: torch.FloatTensor,
|
|
timestep: Union[float, torch.FloatTensor],
|
|
sample: torch.FloatTensor,
|
|
generator: Optional[torch.Generator] = None,
|
|
variance_noise: Optional[torch.FloatTensor] = None,
|
|
return_dict: bool = True,
|
|
) -> Union[FlowMatchDPMSolverMultistepSchedulerOutput, Tuple]:
|
|
"""
|
|
Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
|
|
the multistep DPMSolver.
|
|
|
|
Args:
|
|
model_output (`torch.Tensor`):
|
|
The direct output from learned diffusion model.
|
|
timestep (`int`):
|
|
The current discrete timestep in the diffusion chain.
|
|
sample (`torch.Tensor`):
|
|
A current instance of a sample created by the diffusion process.
|
|
generator (`torch.Generator`, *optional*):
|
|
A random number generator.
|
|
variance_noise (`torch.Tensor`):
|
|
Alternative to generating noise with `generator` by directly providing the noise for the variance
|
|
itself. Useful for methods such as [`LEdits++`].
|
|
return_dict (`bool`):
|
|
Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
|
|
|
|
Returns:
|
|
[`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
|
|
If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
|
|
tuple is returned where the first element is the sample tensor.
|
|
|
|
"""
|
|
if self.num_inference_steps is None:
|
|
raise ValueError(
|
|
"Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
|
|
)
|
|
|
|
if self.step_index is None:
|
|
self._init_step_index(timestep)
|
|
|
|
if self.config.algorithm_type in ["dpmsolver2", "dpmsolver2A"]:
|
|
pass
|
|
else:
|
|
# Flow Match needs to solve an integral of the data prediction model.
|
|
sigma = self.sigmas[self.step_index]
|
|
model_output = sample - sigma * model_output
|
|
for i in range(self.config.solver_order - 1):
|
|
self.model_outputs[i] = self.model_outputs[i + 1]
|
|
self.model_outputs[-1] = model_output
|
|
|
|
# Upcast to avoid precision issues when computing prev_sample
|
|
if sample.dtype != model_output.dtype:
|
|
sample = sample.to(model_output.dtype)
|
|
|
|
if self.config.algorithm_type in ["dpmsolver2A", "dpmsolver++2S", "dpmsolver++sde", "dpmsolver++2Msde", "dpmsolver++3Msde"] and variance_noise is None:
|
|
# Create a noise sampler if it hasn't been created yet
|
|
if self.config.use_noise_sampler:
|
|
if self.noise_sampler is None:
|
|
min_sigma, max_sigma = self.sigmas[self.sigmas > 0].min(), self.sigmas.max()
|
|
self.noise_sampler = BrownianTreeNoiseSampler(sample, min_sigma, max_sigma, generator)
|
|
else:
|
|
noise = randn_tensor(model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype)
|
|
elif self.config.algorithm_type in ["dpmsolver2A", "dpmsolver++2S", "dpmsolver++sde", "dpmsolver++2Msde", "dpmsolver++3Msde"]:
|
|
noise = variance_noise.to(device=model_output.device, dtype=model_output.dtype)
|
|
else:
|
|
noise = None
|
|
|
|
def sigma_fn(_t: torch.Tensor) -> torch.Tensor:
|
|
return _t.neg().exp()
|
|
def t_fn(_sigma: torch.Tensor) -> torch.Tensor:
|
|
return _sigma.log().neg()
|
|
sigma = self.sigmas[self.step_index]
|
|
try:
|
|
sigma_next = self.sigmas[self.step_index + 1]
|
|
except Exception:
|
|
sigma_next = self.sigmas[-1]
|
|
sigma_prev = self.sigmas[self.step_index - 1]
|
|
if self.config.algorithm_type == "dpmsolver2":
|
|
if self.config.solver_order == 2:
|
|
if sigma_next == 0:
|
|
# Euler method
|
|
model_output = sample - sigma * model_output
|
|
d = (sample - model_output) / sigma
|
|
dt = sigma_next - sigma
|
|
sample = sample + d * dt
|
|
else:
|
|
# DPM-Solver2
|
|
sigma_mid = sigma.log().lerp(sigma_next.log(), 0.5).exp()
|
|
|
|
#using epsilon for new model output:
|
|
pred_original_sample = sample - sigma * model_output
|
|
# 2. Convert to an ODE derivative for 1st order
|
|
d = (sample - pred_original_sample) / sigma
|
|
# 3. delta timestep
|
|
dt = sigma_mid - sigma
|
|
x_2 = sample + d * dt
|
|
|
|
#using epsilon for new model output:
|
|
denoised_2 = x_2 - sigma_mid * model_output
|
|
# 2. Convert to an ODE derivative for 2nd order
|
|
d = (x_2 - denoised_2) / sigma_mid
|
|
|
|
# 3. delta timestep
|
|
dt = sigma_next - sigma
|
|
sample = sample + d * dt
|
|
|
|
del pred_original_sample
|
|
del denoised_2
|
|
del x_2
|
|
del d
|
|
elif self.config.algorithm_type == "dpmsolver2A":
|
|
if self.config.solver_order == 2:
|
|
# get ancestral step
|
|
sigma_from = sigma
|
|
sigma_to = sigma_next
|
|
su = min(sigma_to, (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5)
|
|
sd = (sigma_to**2 - su**2) ** 0.5
|
|
if sd == 0:
|
|
# Euler method
|
|
model_output = sample - sigma * model_output
|
|
d = (sample - model_output) / sigma
|
|
dt = sd - sigma
|
|
sample = sample + d * dt
|
|
else:
|
|
# DPM-Solver2A
|
|
sigma_mid = sigma.log().lerp(sd.log(), 0.5).exp()
|
|
|
|
#using epsilon for new model output:
|
|
model_output = sample - sigma * model_output
|
|
# 2. Convert to an ODE derivative for 1st order
|
|
d = (sample - model_output) / sigma
|
|
dt = sd - sigma
|
|
sample = sample + d * dt
|
|
|
|
#using epsilon for new model output:
|
|
pred_original_sample = sample - sigma * model_output
|
|
# 2. Convert to an ODE derivative for 1st order
|
|
d = (sample - pred_original_sample) / sigma
|
|
# 3. delta timestep
|
|
dt_1 = sigma_mid - sigma
|
|
x_2 = sample + d * dt_1
|
|
|
|
#using epsilon for new model output:
|
|
denoised_2 = x_2 - sigma_mid * model_output
|
|
# 2. Convert to an ODE derivative for 2nd order
|
|
d_2 = (x_2 - denoised_2) / sigma_mid
|
|
|
|
# 3. delta timestep
|
|
dt_2 = sd - sigma_mid
|
|
sample = sample + d_2 * dt_2
|
|
|
|
if self.config.use_noise_sampler:
|
|
sample = sample + self.noise_sampler(sigma, sigma_next) * self.config.s_noise * su
|
|
else:
|
|
sample = sample + noise * self.config.s_noise * su
|
|
|
|
del pred_original_sample
|
|
del denoised_2
|
|
del x_2
|
|
del d
|
|
elif self.config.algorithm_type == "dpmsolver++2M":
|
|
if self.config.solver_order == 2:
|
|
t, t_next = t_fn(sigma), t_fn(sigma_next)
|
|
h = t_next - t
|
|
if self.model_outputs[-2] is None or sigma_next == 0:
|
|
sample = (sigma_fn(t_next) / sigma_fn(t)) * sample - (-h).expm1() * model_output
|
|
else:
|
|
# DPM-Solver++(2M)
|
|
h_last = t - t_fn(sigma_prev)
|
|
r = h_last / h
|
|
denoised_d = (1 + 1 / (2 * r)) * model_output - (1 / (2 * r)) * self.model_outputs[-2]
|
|
sample = (sigma_fn(t_next) / sigma_fn(t)) * sample - (-h).expm1() * denoised_d
|
|
del denoised_d
|
|
elif self.config.algorithm_type == "dpmsolver++2S":
|
|
if self.config.solver_order == 2:
|
|
# get ancestral step
|
|
sigma_from = sigma
|
|
sigma_to = sigma_next
|
|
su = min(sigma_to, (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5)
|
|
sd = (sigma_to**2 - su**2) ** 0.5
|
|
if sd == 0:
|
|
# Euler method
|
|
d = (sample - model_output) / sigma
|
|
dt = sd - sigma
|
|
sample = sample + d * dt
|
|
else:
|
|
# DPM-Solver++(2S)
|
|
t, t_next = t_fn(sigma), t_fn(sd)
|
|
r = self.config.midpoint_ratio
|
|
h = t_next - t
|
|
s = t + r * h
|
|
|
|
# Euler method
|
|
d = (sample - model_output) / sigma
|
|
dt = sd - sigma
|
|
sample = sample + d * dt
|
|
|
|
x_2 = (sigma_fn(s) / sigma_fn(t)) * sample - (-h * r).expm1() * model_output
|
|
|
|
#using epsilon for new model output:
|
|
denoised_2 = x_2 - sigma_fn(s) * model_output
|
|
# 2. Convert to an ODE derivative for 2nd order
|
|
d = (x_2 - denoised_2) / sigma_fn(s)
|
|
dt = sd - sigma_next
|
|
sample = sample + d * dt
|
|
|
|
del x_2
|
|
del denoised_2
|
|
del d
|
|
# Noise addition
|
|
if sigma_next > 0:
|
|
if self.config.use_noise_sampler:
|
|
sample = sample + self.noise_sampler(sigma, sigma_next) * self.config.s_noise * su
|
|
else:
|
|
sample = sample + noise * self.config.s_noise * su
|
|
elif self.config.algorithm_type == "dpmsolver++sde":
|
|
if self.config.solver_order == 2:
|
|
if sigma_next == 0:
|
|
# Euler method
|
|
d = (sample - model_output) / sigma
|
|
dt = sigma_next - sigma
|
|
sample = sample + d * dt
|
|
else:
|
|
# DPM-Solver++(SDE)
|
|
t, t_next = t_fn(sigma), t_fn(sigma_next)
|
|
r = self.config.midpoint_ratio
|
|
h = t_next - t
|
|
s = t + r * h
|
|
|
|
# Euler method
|
|
d = (sample - model_output) / sigma
|
|
dt = sigma_next - sigma
|
|
sample = sample + d * dt
|
|
|
|
# Step 1
|
|
# get ancestral step
|
|
sigma_from = sigma_fn(t)
|
|
sigma_to = sigma_fn(s)
|
|
su = min(sigma_to, (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5)
|
|
sd = (sigma_to**2 - su**2) ** 0.5
|
|
|
|
# Euler method
|
|
d = (sample - model_output) / sigma
|
|
dt = sd - sigma
|
|
sample = sample + d * dt
|
|
|
|
s_ = t_fn(sd)
|
|
x_2 = (sigma_fn(s_) / sigma_fn(t)) * sample - (t - s_).expm1() * model_output
|
|
if self.config.use_noise_sampler:
|
|
x_2 = x_2 + self.noise_sampler(sigma_fn(t), sigma_fn(s)) * self.config.s_noise * su
|
|
else:
|
|
x_2 = x_2 + noise * self.config.s_noise * su
|
|
|
|
# Step 2
|
|
# get ancestral step
|
|
sigma_from = sigma_fn(t)
|
|
sigma_to = sigma_fn(t_next)
|
|
su = min(sigma_to, (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5)
|
|
sd = (sigma_to**2 - su**2) ** 0.5
|
|
|
|
#using epsilon for new model output:
|
|
denoised_2 = x_2 - sigma_fn(s) * model_output
|
|
# 2. Convert to an ODE derivative for 2nd order
|
|
d = (x_2 - denoised_2) / sigma_fn(s)
|
|
dt = sd - sigma_next
|
|
sample = sample + d * dt
|
|
|
|
if self.config.use_noise_sampler:
|
|
sample = sample + self.noise_sampler(sigma_fn(t), sigma_fn(t_next)) * self.config.s_noise * su
|
|
else:
|
|
sample = sample + noise * self.config.s_noise * su
|
|
del x_2
|
|
del denoised_2
|
|
del d
|
|
elif self.config.algorithm_type == "dpmsolver++2Msde":
|
|
if self.config.solver_order == 2:
|
|
if sigma_next == 0:
|
|
sample = model_output
|
|
else:
|
|
# DPM-Solver++(2M) SDE
|
|
t, s = -sigma.log(), -sigma_next.log()
|
|
h = s - t
|
|
eta_h = h * 1
|
|
|
|
# 3. Delta timestep
|
|
dt = sigma_next - sigma
|
|
sample = sample + model_output * dt
|
|
|
|
sample = sigma_next / sigma * (-eta_h).exp() * sample + (-h - eta_h).expm1().neg() * model_output
|
|
|
|
if self.model_outputs[-2] is not None:
|
|
r = self.h_last / h
|
|
if self.solver_type == 'heun':
|
|
sample = sample + ((-h - eta_h).expm1().neg() / (-h - eta_h) + 1) * (1 / r) * (model_output - self.model_outputs[-2])
|
|
elif self.solver_type == 'midpoint':
|
|
sample = sample + 0.5 * (-h - eta_h).expm1().neg() * (1 / r) * (model_output - self.model_outputs[-2])
|
|
|
|
if self.config.use_noise_sampler:
|
|
sample = sample + self.noise_sampler(sigma, sigma_next) * sigma_next * (-2 * eta_h).expm1().neg().sqrt() * self.config.s_noise
|
|
else:
|
|
sample = sample + noise * sigma_next * (-2 * eta_h).expm1().neg().sqrt() * self.config.s_noise
|
|
|
|
self.h_last = h
|
|
elif self.config.algorithm_type == "dpmsolver++3Msde":
|
|
if self.config.solver_order == 3:
|
|
if sigma_next == 0:
|
|
sample = model_output
|
|
else:
|
|
# DPM-Solver++(3M) SDE
|
|
t, s = -sigma.log(), -sigma_next.log()
|
|
h = s - t
|
|
h_eta = h * 2
|
|
|
|
# 3. Delta timestep
|
|
dt = sigma_next - sigma
|
|
sample = sample + model_output * dt
|
|
|
|
sample = torch.exp(-h_eta) * sample + (-h_eta).expm1().neg() * model_output
|
|
|
|
if self.h_2 is not None:
|
|
r0 = self.h_1 / h
|
|
r1 = self.h_2 / h
|
|
d1_0 = (model_output - self.model_outputs[-2]) / r0
|
|
d1_1 = (self.model_outputs[-2] - self.model_outputs[-3]) / r1
|
|
d1 = d1_0 + (d1_0 - d1_1) * r0 / (r0 + r1)
|
|
d2 = (d1_0 - d1_1) / (r0 + r1)
|
|
phi_2 = h_eta.neg().expm1() / h_eta + 1
|
|
phi_3 = phi_2 / h_eta - 0.5
|
|
sample = sample + phi_2 * d1 - phi_3 * d2
|
|
del d1_0
|
|
del d1_1
|
|
del d1
|
|
del d2
|
|
del phi_2
|
|
del phi_3
|
|
elif self.h_1 is not None:
|
|
r = self.h_1 / h
|
|
d = (model_output - self.model_outputs[-2]) / r
|
|
phi_2 = h_eta.neg().expm1() / h_eta + 1
|
|
sample = sample + phi_2 * d
|
|
del d
|
|
del phi_2
|
|
|
|
if self.config.use_noise_sampler:
|
|
sample = sample + self.noise_sampler(sigma, sigma_next) * sigma_next * (-2 * h).expm1().neg().sqrt() * self.config.s_noise
|
|
else:
|
|
sample = sample + noise * sigma_next * (-2 * h).expm1().neg().sqrt() * self.config.s_noise
|
|
|
|
self.h_2 = self.h_1
|
|
self.h_1 = h
|
|
if not self.config.use_noise_sampler and noise is not None:
|
|
del noise
|
|
prev_sample = sample
|
|
|
|
# Cast sample back to expected dtype
|
|
prev_sample = prev_sample.to(model_output.dtype)
|
|
|
|
# upon completion increase step index by one
|
|
self._step_index += 1
|
|
|
|
torch.cuda.empty_cache()
|
|
|
|
if not return_dict:
|
|
return (prev_sample,)
|
|
|
|
return FlowMatchDPMSolverMultistepSchedulerOutput(prev_sample=prev_sample)
|
|
|
|
def scale_model_input(self, sample: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor:
|
|
"""
|
|
Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
|
|
current timestep.
|
|
|
|
Args:
|
|
sample (`torch.Tensor`):
|
|
The input sample.
|
|
|
|
Returns:
|
|
`torch.Tensor`:
|
|
A scaled input sample.
|
|
"""
|
|
return sample
|
|
|
|
def scale_noise(
|
|
self,
|
|
sample: torch.FloatTensor,
|
|
timestep: Union[float, torch.FloatTensor],
|
|
noise: Optional[torch.FloatTensor] = None,
|
|
) -> torch.FloatTensor:
|
|
"""
|
|
Forward process in flow-matching
|
|
|
|
Args:
|
|
sample (`torch.FloatTensor`):
|
|
The input sample.
|
|
timestep (`int`, *optional*):
|
|
The current timestep in the diffusion chain.
|
|
|
|
Returns:
|
|
`torch.FloatTensor`:
|
|
A scaled input sample.
|
|
"""
|
|
# Make sure sigmas and timesteps have the same device and dtype as original_samples
|
|
sigmas = self.sigmas.to(device=sample.device, dtype=sample.dtype)
|
|
|
|
if sample.device.type == "mps" and torch.is_floating_point(timestep):
|
|
# mps does not support float64
|
|
schedule_timesteps = self.timesteps.to(sample.device, dtype=torch.float32)
|
|
timestep = timestep.to(sample.device, dtype=torch.float32)
|
|
else:
|
|
schedule_timesteps = self.timesteps.to(sample.device)
|
|
timestep = timestep.to(sample.device)
|
|
|
|
# self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index
|
|
if self.begin_index is None:
|
|
step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timestep]
|
|
elif self.step_index is not None:
|
|
# add_noise is called after first denoising step (for inpainting)
|
|
step_indices = [self.step_index] * timestep.shape[0]
|
|
else:
|
|
# add noise is called before first denoising step to create initial latent(img2img)
|
|
step_indices = [self.begin_index] * timestep.shape[0]
|
|
|
|
sigma = sigmas[step_indices].flatten()
|
|
while len(sigma.shape) < len(sample.shape):
|
|
sigma = sigma.unsqueeze(-1)
|
|
|
|
sample = sigma * noise + (1.0 - sigma) * sample
|
|
|
|
return sample
|
|
|
|
def __len__(self):
|
|
return self.config.num_train_timesteps
|