diffusers 0.17.1__py3-none-any.whl → 0.18.2__py3-none-any.whl

diffusers/schedulers/scheduling_deis_multistep.py

@@ -26,7 +26,11 @@ from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, Schedul
 
 
 # Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
-def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999):
+def betas_for_alpha_bar(
+    num_diffusion_timesteps,
+    max_beta=0.999,
+    alpha_transform_type="cosine",
+):
     """
     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].
@@ -39,19 +43,30 @@ def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999):
         num_diffusion_timesteps (`int`): the number of betas to produce.
         max_beta (`float`): the maximum beta to use; use values lower than 1 to
                      prevent singularities.
+        alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+                     Choose from `cosine` or `exp`
 
     Returns:
         betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
     """
+    if alpha_transform_type == "cosine":
 
-    def alpha_bar(time_step):
-        return math.cos((time_step + 0.008) / 1.008 * math.pi / 2) ** 2
+        def alpha_bar_fn(t):
+            return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+    elif alpha_transform_type == "exp":
+
+        def alpha_bar_fn(t):
+            return math.exp(t * -12.0)
+
+    else:
+        raise ValueError(f"Unsupported alpha_tranform_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(t2) / alpha_bar(t1), max_beta))
+        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
     return torch.tensor(betas, dtype=torch.float32)
 
 
@@ -103,7 +118,17 @@ class DEISMultistepScheduler(SchedulerMixin, ConfigMixin):
         lower_order_final (`bool`, default `True`):
             whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. We empirically
             find this trick can stabilize the sampling of DEIS for steps < 15, especially for steps <= 10.
-
+        use_karras_sigmas (`bool`, *optional*, defaults to `False`):
+             This parameter controls whether to use Karras sigmas (Karras et al. (2022) scheme) for step sizes in the
+             noise schedule during the sampling process. If True, the sigmas will be determined according to a sequence
+             of noise levels {σi} as defined in Equation (5) of the paper https://arxiv.org/pdf/2206.00364.pdf.
+        timestep_spacing (`str`, default `"linspace"`):
+            The way the timesteps should be scaled. Refer to Table 2. of [Common Diffusion Noise Schedules and Sample
+            Steps are Flawed](https://arxiv.org/abs/2305.08891) for more information.
+        steps_offset (`int`, default `0`):
+            an offset added to the inference steps. You can use a combination of `offset=1` and
+            `set_alpha_to_one=False`, to make the last step use step 0 for the previous alpha product, as done in
+            stable diffusion.
     """
 
     _compatibles = [e.name for e in KarrasDiffusionSchedulers]
@@ -125,6 +150,9 @@ class DEISMultistepScheduler(SchedulerMixin, ConfigMixin):
         algorithm_type: str = "deis",
         solver_type: str = "logrho",
         lower_order_final: bool = True,
+        use_karras_sigmas: Optional[bool] = False,
+        timestep_spacing: str = "linspace",
+        steps_offset: int = 0,
     ):
         if trained_betas is not None:
             self.betas = torch.tensor(trained_betas, dtype=torch.float32)
@@ -181,12 +209,39 @@ class DEISMultistepScheduler(SchedulerMixin, ConfigMixin):
             device (`str` or `torch.device`, optional):
                 the device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
         """
-        timesteps = (
-            np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps + 1)
-            .round()[::-1][:-1]
-            .copy()
-            .astype(np.int64)
-        )
+        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+        if self.config.timestep_spacing == "linspace":
+            timesteps = (
+                np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps + 1)
+                .round()[::-1][:-1]
+                .copy()
+                .astype(np.int64)
+            )
+        elif self.config.timestep_spacing == "leading":
+            step_ratio = self.config.num_train_timesteps // (num_inference_steps + 1)
+            # creates integer timesteps by multiplying by ratio
+            # casting to int to avoid issues when num_inference_step is power of 3
+            timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(np.int64)
+            timesteps += self.config.steps_offset
+        elif self.config.timestep_spacing == "trailing":
+            step_ratio = self.config.num_train_timesteps / num_inference_steps
+            # creates integer timesteps by multiplying by ratio
+            # casting to int to avoid issues when num_inference_step is power of 3
+            timesteps = np.arange(self.config.num_train_timesteps, 0, -step_ratio).round().copy().astype(np.int64)
+            timesteps -= 1
+        else:
+            raise ValueError(
+                f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+            )
+
+        sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+        if self.config.use_karras_sigmas:
+            log_sigmas = np.log(sigmas)
+            sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
+            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
+            timesteps = np.flip(timesteps).copy().astype(np.int64)
+
+        self.sigmas = torch.from_numpy(sigmas)
 
         # when num_inference_steps == num_train_timesteps, we can end up with
         # duplicates in timesteps.

diffusers/schedulers/scheduling_dpmsolver_multistep.py

@@ -26,7 +26,11 @@ from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, Schedul
 
 
 # Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
-def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999):
+def betas_for_alpha_bar(
+    num_diffusion_timesteps,
+    max_beta=0.999,
+    alpha_transform_type="cosine",
+):
     """
     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].
@@ -39,19 +43,30 @@ def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999):
         num_diffusion_timesteps (`int`): the number of betas to produce.
         max_beta (`float`): the maximum beta to use; use values lower than 1 to
                      prevent singularities.
+        alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+                     Choose from `cosine` or `exp`
 
     Returns:
         betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
     """
+    if alpha_transform_type == "cosine":
 
-    def alpha_bar(time_step):
-        return math.cos((time_step + 0.008) / 1.008 * math.pi / 2) ** 2
+        def alpha_bar_fn(t):
+            return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+    elif alpha_transform_type == "exp":
+
+        def alpha_bar_fn(t):
+            return math.exp(t * -12.0)
+
+    else:
+        raise ValueError(f"Unsupported alpha_tranform_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(t2) / alpha_bar(t1), max_beta))
+        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
     return torch.tensor(betas, dtype=torch.float32)
 
 
@@ -134,6 +149,13 @@ class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
             guided-diffusion (https://github.com/openai/guided-diffusion) predicts both mean and variance of the
             Gaussian distribution in the model's output. DPM-Solver only needs the "mean" output because it is based on
             diffusion ODEs.
+        timestep_spacing (`str`, default `"linspace"`):
+            The way the timesteps should be scaled. Refer to Table 2. of [Common Diffusion Noise Schedules and Sample
+            Steps are Flawed](https://arxiv.org/abs/2305.08891) for more information.
+        steps_offset (`int`, default `0`):
+            an offset added to the inference steps. You can use a combination of `offset=1` and
+            `set_alpha_to_one=False`, to make the last step use step 0 for the previous alpha product, as done in
+            stable diffusion.
     """
 
     _compatibles = [e.name for e in KarrasDiffusionSchedulers]
@@ -158,6 +180,8 @@ class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         use_karras_sigmas: Optional[bool] = False,
         lambda_min_clipped: float = -float("inf"),
         variance_type: Optional[str] = None,
+        timestep_spacing: str = "linspace",
+        steps_offset: int = 0,
     ):
         if trained_betas is not None:
             self.betas = torch.tensor(trained_betas, dtype=torch.float32)
@@ -203,7 +227,6 @@ class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         self.timesteps = torch.from_numpy(timesteps)
         self.model_outputs = [None] * solver_order
         self.lower_order_nums = 0
-        self.use_karras_sigmas = use_karras_sigmas
 
     def set_timesteps(self, num_inference_steps: int = None, device: Union[str, torch.device] = None):
         """
@@ -218,20 +241,39 @@ class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         # Clipping the minimum of all lambda(t) for numerical stability.
         # This is critical for cosine (squaredcos_cap_v2) noise schedule.
         clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.config.lambda_min_clipped)
-        timesteps = (
-            np.linspace(0, self.config.num_train_timesteps - 1 - clipped_idx, num_inference_steps + 1)
-            .round()[::-1][:-1]
-            .copy()
-            .astype(np.int64)
-        )
+        last_timestep = ((self.config.num_train_timesteps - clipped_idx).numpy()).item()
+
+        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+        if self.config.timestep_spacing == "linspace":
+            timesteps = (
+                np.linspace(0, last_timestep - 1, num_inference_steps + 1).round()[::-1][:-1].copy().astype(np.int64)
+            )
+        elif self.config.timestep_spacing == "leading":
+            step_ratio = last_timestep // (num_inference_steps + 1)
+            # creates integer timesteps by multiplying by ratio
+            # casting to int to avoid issues when num_inference_step is power of 3
+            timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(np.int64)
+            timesteps += self.config.steps_offset
+        elif self.config.timestep_spacing == "trailing":
+            step_ratio = self.config.num_train_timesteps / num_inference_steps
+            # creates integer timesteps by multiplying by ratio
+            # casting to int to avoid issues when num_inference_step is power of 3
+            timesteps = np.arange(last_timestep, 0, -step_ratio).round().copy().astype(np.int64)
+            timesteps -= 1
+        else:
+            raise ValueError(
+                f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+            )
 
-        if self.use_karras_sigmas:
-            sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+        sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+        if self.config.use_karras_sigmas:
             log_sigmas = np.log(sigmas)
             sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
             timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
             timesteps = np.flip(timesteps).copy().astype(np.int64)
 
+        self.sigmas = torch.from_numpy(sigmas)
+
         # when num_inference_steps == num_train_timesteps, we can end up with
         # duplicates in timesteps.
         _, unique_indices = np.unique(timesteps, return_index=True)

diffusers/schedulers/scheduling_dpmsolver_multistep_inverse.py

@@ -26,7 +26,11 @@ from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, Schedul
 
 
 # Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
-def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999):
+def betas_for_alpha_bar(
+    num_diffusion_timesteps,
+    max_beta=0.999,
+    alpha_transform_type="cosine",
+):
     """
     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].
@@ -39,19 +43,30 @@ def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999):
         num_diffusion_timesteps (`int`): the number of betas to produce.
         max_beta (`float`): the maximum beta to use; use values lower than 1 to
                      prevent singularities.
+        alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+                     Choose from `cosine` or `exp`
 
     Returns:
         betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
     """
+    if alpha_transform_type == "cosine":
 
-    def alpha_bar(time_step):
-        return math.cos((time_step + 0.008) / 1.008 * math.pi / 2) ** 2
+        def alpha_bar_fn(t):
+            return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+    elif alpha_transform_type == "exp":
+
+        def alpha_bar_fn(t):
+            return math.exp(t * -12.0)
+
+    else:
+        raise ValueError(f"Unsupported alpha_tranform_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(t2) / alpha_bar(t1), max_beta))
+        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
     return torch.tensor(betas, dtype=torch.float32)
 
 

diffusers/schedulers/scheduling_dpmsolver_sde.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 import math
+from collections import defaultdict
 from typing import List, Optional, Tuple, Union
 
 import numpy as np
@@ -76,7 +77,11 @@ class BrownianTreeNoiseSampler:
 
 
 # Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
-def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999) -> torch.Tensor:
+def betas_for_alpha_bar(
+    num_diffusion_timesteps,
+    max_beta=0.999,
+    alpha_transform_type="cosine",
+):
     """
     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].
@@ -89,19 +94,30 @@ def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999) -> torch.Tensor
         num_diffusion_timesteps (`int`): the number of betas to produce.
         max_beta (`float`): the maximum beta to use; use values lower than 1 to
                      prevent singularities.
+        alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+                     Choose from `cosine` or `exp`
 
     Returns:
         betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
     """
+    if alpha_transform_type == "cosine":
 
-    def alpha_bar(time_step):
-        return math.cos((time_step + 0.008) / 1.008 * math.pi / 2) ** 2
+        def alpha_bar_fn(t):
+            return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+    elif alpha_transform_type == "exp":
+
+        def alpha_bar_fn(t):
+            return math.exp(t * -12.0)
+
+    else:
+        raise ValueError(f"Unsupported alpha_tranform_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(t2) / alpha_bar(t1), max_beta))
+        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
     return torch.tensor(betas, dtype=torch.float32)
 
 
@@ -133,6 +149,13 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
              of noise levels {σi} as defined in Equation (5) of the paper https://arxiv.org/pdf/2206.00364.pdf.
         noise_sampler_seed (`int`, *optional*, defaults to `None`):
             The random seed to use for the noise sampler. If `None`, a random seed will be generated.
+        timestep_spacing (`str`, default `"linspace"`):
+            The way the timesteps should be scaled. Refer to Table 2. of [Common Diffusion Noise Schedules and Sample
+            Steps are Flawed](https://arxiv.org/abs/2305.08891) for more information.
+        steps_offset (`int`, default `0`):
+            an offset added to the inference steps. You can use a combination of `offset=1` and
+            `set_alpha_to_one=False`, to make the last step use step 0 for the previous alpha product, as done in
+            stable diffusion.
     """
 
     _compatibles = [e.name for e in KarrasDiffusionSchedulers]
@@ -149,6 +172,8 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
         prediction_type: str = "epsilon",
         use_karras_sigmas: Optional[bool] = False,
         noise_sampler_seed: Optional[int] = None,
+        timestep_spacing: str = "linspace",
+        steps_offset: int = 0,
     ):
         if trained_betas is not None:
             self.betas = torch.tensor(trained_betas, dtype=torch.float32)
@@ -181,12 +206,26 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
 
         indices = (schedule_timesteps == timestep).nonzero()
 
-        if self.state_in_first_order:
-            pos = -1
+        # 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)
+        if len(self._index_counter) == 0:
+            pos = 1 if len(indices) > 1 else 0
         else:
-            pos = 0
+            timestep_int = timestep.cpu().item() if torch.is_tensor(timestep) else timestep
+            pos = self._index_counter[timestep_int]
+
         return indices[pos].item()
 
+    @property
+    def init_noise_sigma(self):
+        # standard deviation of the initial noise distribution
+        if self.config.timestep_spacing in ["linspace", "trailing"]:
+            return self.sigmas.max()
+
+        return (self.sigmas.max() ** 2 + 1) ** 0.5
+
     def scale_model_input(
         self,
         sample: torch.FloatTensor,
@@ -226,7 +265,25 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
 
         num_train_timesteps = num_train_timesteps or self.config.num_train_timesteps
 
-        timesteps = np.linspace(0, num_train_timesteps - 1, num_inference_steps, dtype=float)[::-1].copy()
+        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+        if self.config.timestep_spacing == "linspace":
+            timesteps = np.linspace(0, num_train_timesteps - 1, num_inference_steps, dtype=float)[::-1].copy()
+        elif self.config.timestep_spacing == "leading":
+            step_ratio = num_train_timesteps // self.num_inference_steps
+            # creates integer timesteps by multiplying by ratio
+            # casting to int to avoid issues when num_inference_step is power of 3
+            timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(float)
+            timesteps += self.config.steps_offset
+        elif self.config.timestep_spacing == "trailing":
+            step_ratio = num_train_timesteps / self.num_inference_steps
+            # creates integer timesteps by multiplying by ratio
+            # casting to int to avoid issues when num_inference_step is power of 3
+            timesteps = (np.arange(num_train_timesteps, 0, -step_ratio)).round().copy().astype(float)
+            timesteps -= 1
+        else:
+            raise ValueError(
+                f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+            )
 
         sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
         log_sigmas = np.log(sigmas)
@@ -242,9 +299,6 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
         sigmas = torch.from_numpy(sigmas).to(device=device)
         self.sigmas = torch.cat([sigmas[:1], sigmas[1:-1].repeat_interleave(2), sigmas[-1:]])
 
-        # standard deviation of the initial noise distribution
-        self.init_noise_sigma = self.sigmas.max()
-
         timesteps = torch.from_numpy(timesteps)
         second_order_timesteps = torch.from_numpy(second_order_timesteps)
         timesteps = torch.cat([timesteps[:1], timesteps[1:].repeat_interleave(2)])
@@ -260,6 +314,10 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
         self.sample = None
         self.mid_point_sigma = None
 
+        # for exp beta schedules, such as the one for `pipeline_shap_e.py`
+        # we need an index counter
+        self._index_counter = defaultdict(int)
+
     def _second_order_timesteps(self, sigmas, log_sigmas):
         def sigma_fn(_t):
             return np.exp(-_t)
@@ -341,6 +399,10 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
         """
         step_index = self.index_for_timestep(timestep)
 
+        # advance index counter by 1
+        timestep_int = timestep.cpu().item() if torch.is_tensor(timestep) else timestep
+        self._index_counter[timestep_int] += 1
+
         # Create a noise sampler if it hasn't been created yet
         if self.noise_sampler is None:
             min_sigma, max_sigma = self.sigmas[self.sigmas > 0].min(), self.sigmas.max()

diffusers/schedulers/scheduling_dpmsolver_singlestep.py

@@ -29,7 +29,11 @@ logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
 # Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
-def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999):
+def betas_for_alpha_bar(
+    num_diffusion_timesteps,
+    max_beta=0.999,
+    alpha_transform_type="cosine",
+):
     """
     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].
@@ -42,19 +46,30 @@ def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999):
         num_diffusion_timesteps (`int`): the number of betas to produce.
         max_beta (`float`): the maximum beta to use; use values lower than 1 to
                      prevent singularities.
+        alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+                     Choose from `cosine` or `exp`
 
     Returns:
         betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
     """
+    if alpha_transform_type == "cosine":
 
-    def alpha_bar(time_step):
-        return math.cos((time_step + 0.008) / 1.008 * math.pi / 2) ** 2
+        def alpha_bar_fn(t):
+            return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+    elif alpha_transform_type == "exp":
+
+        def alpha_bar_fn(t):
+            return math.exp(t * -12.0)
+
+    else:
+        raise ValueError(f"Unsupported alpha_tranform_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(t2) / alpha_bar(t1), max_beta))
+        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
     return torch.tensor(betas, dtype=torch.float32)
 
 
@@ -202,7 +217,6 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
         self.model_outputs = [None] * solver_order
         self.sample = None
         self.order_list = self.get_order_list(num_train_timesteps)
-        self.use_karras_sigmas = use_karras_sigmas
 
     def get_order_list(self, num_inference_steps: int) -> List[int]:
         """
@@ -259,13 +273,15 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
             .astype(np.int64)
         )
 
-        if self.use_karras_sigmas:
-            sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+        sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+        if self.config.use_karras_sigmas:
             log_sigmas = np.log(sigmas)
             sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
             timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
             timesteps = np.flip(timesteps).copy().astype(np.int64)
 
+        self.sigmas = torch.from_numpy(sigmas)
+
         self.timesteps = torch.from_numpy(timesteps).to(device)
         self.model_outputs = [None] * self.config.solver_order
         self.sample = None

diffusers/schedulers/scheduling_euler_ancestral_discrete.py

@@ -47,7 +47,11 @@ class EulerAncestralDiscreteSchedulerOutput(BaseOutput):
 
 
 # Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
-def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999) -> torch.Tensor:
+def betas_for_alpha_bar(
+    num_diffusion_timesteps,
+    max_beta=0.999,
+    alpha_transform_type="cosine",
+):
     """
     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].
@@ -60,19 +64,30 @@ def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999) -> torch.Tensor
         num_diffusion_timesteps (`int`): the number of betas to produce.
         max_beta (`float`): the maximum beta to use; use values lower than 1 to
                      prevent singularities.
+        alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+                     Choose from `cosine` or `exp`
 
     Returns:
         betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
     """
+    if alpha_transform_type == "cosine":
 
-    def alpha_bar(time_step):
-        return math.cos((time_step + 0.008) / 1.008 * math.pi / 2) ** 2
+        def alpha_bar_fn(t):
+            return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+    elif alpha_transform_type == "exp":
+
+        def alpha_bar_fn(t):
+            return math.exp(t * -12.0)
+
+    else:
+        raise ValueError(f"Unsupported alpha_tranform_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(t2) / alpha_bar(t1), max_beta))
+        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
     return torch.tensor(betas, dtype=torch.float32)
 
 
@@ -99,7 +114,13 @@ class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
             prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion
             process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4
             https://imagen.research.google/video/paper.pdf)
-
+        timestep_spacing (`str`, default `"linspace"`):
+            The way the timesteps should be scaled. Refer to Table 2. of [Common Diffusion Noise Schedules and Sample
+            Steps are Flawed](https://arxiv.org/abs/2305.08891) for more information.
+        steps_offset (`int`, default `0`):
+            an offset added to the inference steps. You can use a combination of `offset=1` and
+            `set_alpha_to_one=False`, to make the last step use step 0 for the previous alpha product, as done in
+            stable diffusion.
     """
 
     _compatibles = [e.name for e in KarrasDiffusionSchedulers]
@@ -114,6 +135,8 @@ class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
         beta_schedule: str = "linear",
         trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
         prediction_type: str = "epsilon",
+        timestep_spacing: str = "linspace",
+        steps_offset: int = 0,
     ):
         if trained_betas is not None:
             self.betas = torch.tensor(trained_betas, dtype=torch.float32)
@@ -137,15 +160,20 @@ class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
         sigmas = np.concatenate([sigmas[::-1], [0.0]]).astype(np.float32)
         self.sigmas = torch.from_numpy(sigmas)
 
-        # standard deviation of the initial noise distribution
-        self.init_noise_sigma = self.sigmas.max()
-
         # setable values
         self.num_inference_steps = None
         timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=float)[::-1].copy()
         self.timesteps = torch.from_numpy(timesteps)
         self.is_scale_input_called = False
 
+    @property
+    def init_noise_sigma(self):
+        # standard deviation of the initial noise distribution
+        if self.config.timestep_spacing in ["linspace", "trailing"]:
+            return self.sigmas.max()
+
+        return (self.sigmas.max() ** 2 + 1) ** 0.5
+
     def scale_model_input(
         self, sample: torch.FloatTensor, timestep: Union[float, torch.FloatTensor]
     ) -> torch.FloatTensor:
@@ -179,7 +207,28 @@ class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
         """
         self.num_inference_steps = num_inference_steps
 
-        timesteps = np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps, dtype=float)[::-1].copy()
+        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+        if self.config.timestep_spacing == "linspace":
+            timesteps = np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps, dtype=float)[
+                ::-1
+            ].copy()
+        elif self.config.timestep_spacing == "leading":
+            step_ratio = self.config.num_train_timesteps // self.num_inference_steps
+            # creates integer timesteps by multiplying by ratio
+            # casting to int to avoid issues when num_inference_step is power of 3
+            timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(float)
+            timesteps += self.config.steps_offset
+        elif self.config.timestep_spacing == "trailing":
+            step_ratio = self.config.num_train_timesteps / self.num_inference_steps
+            # creates integer timesteps by multiplying by ratio
+            # casting to int to avoid issues when num_inference_step is power of 3
+            timesteps = (np.arange(self.config.num_train_timesteps, 0, -step_ratio)).round().copy().astype(float)
+            timesteps -= 1
+        else:
+            raise ValueError(
+                f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+            )
+
         sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
         sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
         sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32)