| import math |
| from dataclasses import dataclass |
| from typing import Literal |
|
|
| import torch |
|
|
| from ..configuration_utils import ConfigMixin, register_to_config |
| from ..utils import BaseOutput |
| from ..utils.torch_utils import randn_tensor |
| from .scheduling_utils import SchedulerMixin |
|
|
|
|
| |
| def betas_for_alpha_bar( |
| num_diffusion_timesteps: int, |
| max_beta: float = 0.999, |
| alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", |
| ) -> torch.Tensor: |
| """ |
| Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of |
| (1-beta) over time from t = [0,1]. |
| |
| Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up |
| to that part of the diffusion process. |
| |
| Args: |
| num_diffusion_timesteps (`int`): |
| The number of betas to produce. |
| max_beta (`float`, defaults to `0.999`): |
| The maximum beta to use; use values lower than 1 to avoid numerical instability. |
| alpha_transform_type (`str`, defaults to `"cosine"`): |
| The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. |
| |
| Returns: |
| `torch.Tensor`: |
| The betas used by the scheduler to step the model outputs. |
| """ |
| if alpha_transform_type == "cosine": |
|
|
| def alpha_bar_fn(t): |
| return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 |
|
|
| elif alpha_transform_type == "laplace": |
|
|
| def alpha_bar_fn(t): |
| lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) |
| snr = math.exp(lmb) |
| return math.sqrt(snr / (1 + snr)) |
|
|
| elif alpha_transform_type == "exp": |
|
|
| def alpha_bar_fn(t): |
| return math.exp(t * -12.0) |
|
|
| else: |
| raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") |
|
|
| betas = [] |
| for i in range(num_diffusion_timesteps): |
| t1 = i / num_diffusion_timesteps |
| t2 = (i + 1) / num_diffusion_timesteps |
| betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) |
| return torch.tensor(betas, dtype=torch.float32) |
|
|
|
|
| @dataclass |
| class ConsistencyDecoderSchedulerOutput(BaseOutput): |
| """ |
| Output class for the scheduler's `step` function. |
| |
| Args: |
| prev_sample (`torch.Tensor` 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.Tensor |
|
|
|
|
| class ConsistencyDecoderScheduler(SchedulerMixin, ConfigMixin): |
| """ |
| A scheduler for the consistency decoder used in Stable Diffusion pipelines. |
| |
| This scheduler implements a two-step denoising process using consistency models for decoding latent representations |
| into images. |
| |
| 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`, *optional*, defaults to `1024`): |
| The number of diffusion steps to train the model. |
| sigma_data (`float`, *optional*, defaults to `0.5`): |
| The standard deviation of the data distribution. Used for computing the skip and output scaling factors. |
| """ |
|
|
| order = 1 |
|
|
| @register_to_config |
| def __init__( |
| self, |
| num_train_timesteps: int = 1024, |
| sigma_data: float = 0.5, |
| ) -> None: |
| betas = betas_for_alpha_bar(num_train_timesteps) |
|
|
| alphas = 1.0 - betas |
| alphas_cumprod = torch.cumprod(alphas, dim=0) |
|
|
| self.sqrt_alphas_cumprod = torch.sqrt(alphas_cumprod) |
| self.sqrt_one_minus_alphas_cumprod = torch.sqrt(1.0 - alphas_cumprod) |
|
|
| sigmas = torch.sqrt(1.0 / alphas_cumprod - 1) |
|
|
| sqrt_recip_alphas_cumprod = torch.sqrt(1.0 / alphas_cumprod) |
|
|
| self.c_skip = sqrt_recip_alphas_cumprod * sigma_data**2 / (sigmas**2 + sigma_data**2) |
| self.c_out = sigmas * sigma_data / (sigmas**2 + sigma_data**2) ** 0.5 |
| self.c_in = sqrt_recip_alphas_cumprod / (sigmas**2 + sigma_data**2) ** 0.5 |
|
|
| def set_timesteps( |
| self, |
| num_inference_steps: int | None = None, |
| device: str | torch.device = None, |
| ): |
| if num_inference_steps != 2: |
| raise ValueError("Currently more than 2 inference steps are not supported.") |
|
|
| self.timesteps = torch.tensor([1008, 512], dtype=torch.long, device=device) |
| self.sqrt_alphas_cumprod = self.sqrt_alphas_cumprod.to(device) |
| self.sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod.to(device) |
| self.c_skip = self.c_skip.to(device) |
| self.c_out = self.c_out.to(device) |
| self.c_in = self.c_in.to(device) |
|
|
| @property |
| def init_noise_sigma(self) -> torch.Tensor: |
| """ |
| Return the standard deviation of the initial noise distribution. |
| |
| Returns: |
| `torch.Tensor`: |
| The initial noise sigma value from the precomputed `sqrt_one_minus_alphas_cumprod` at the first |
| timestep. |
| """ |
| return self.sqrt_one_minus_alphas_cumprod[self.timesteps[0]] |
|
|
| def scale_model_input(self, sample: torch.Tensor, timestep: int | None = None) -> torch.Tensor: |
| """ |
| Ensures interchangeability with schedulers that need to scale the denoising model input depending on the |
| current timestep. |
| |
| Args: |
| sample (`torch.Tensor`): |
| The input sample. |
| timestep (`int`, *optional*): |
| The current timestep in the diffusion chain. |
| |
| Returns: |
| `torch.Tensor`: |
| A scaled input sample. |
| """ |
| return sample * self.c_in[timestep] |
|
|
| def step( |
| self, |
| model_output: torch.Tensor, |
| timestep: float | torch.Tensor, |
| sample: torch.Tensor, |
| generator: torch.Generator | None = None, |
| return_dict: bool = True, |
| ) -> ConsistencyDecoderSchedulerOutput | tuple: |
| """ |
| Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion |
| process from the learned model outputs (most often the predicted noise). |
| |
| Args: |
| model_output (`torch.Tensor`): |
| The direct output from the learned diffusion model. |
| timestep (`float` or `torch.Tensor`): |
| The current 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 for reproducibility. |
| return_dict (`bool`, *optional*, defaults to `True`): |
| Whether or not to return a |
| [`~schedulers.scheduling_consistency_decoder.ConsistencyDecoderSchedulerOutput`] or `tuple`. |
| |
| Returns: |
| [`~schedulers.scheduling_consistency_decoder.ConsistencyDecoderSchedulerOutput`] or `tuple`: |
| If `return_dict` is `True`, |
| [`~schedulers.scheduling_consistency_decoder.ConsistencyDecoderSchedulerOutput`] is returned, otherwise |
| a tuple is returned where the first element is the sample tensor. |
| """ |
| x_0 = self.c_out[timestep] * model_output + self.c_skip[timestep] * sample |
|
|
| timestep_idx = torch.where(self.timesteps == timestep)[0] |
|
|
| if timestep_idx == len(self.timesteps) - 1: |
| prev_sample = x_0 |
| else: |
| noise = randn_tensor(x_0.shape, generator=generator, dtype=x_0.dtype, device=x_0.device) |
| prev_sample = ( |
| self.sqrt_alphas_cumprod[self.timesteps[timestep_idx + 1]].to(x_0.dtype) * x_0 |
| + self.sqrt_one_minus_alphas_cumprod[self.timesteps[timestep_idx + 1]].to(x_0.dtype) * noise |
| ) |
|
|
| if not return_dict: |
| return (prev_sample,) |
|
|
| return ConsistencyDecoderSchedulerOutput(prev_sample=prev_sample) |
|
|
