diffusers 0.33.1__py3-none-any.whl → 0.34.0__py3-none-any.whl

diffusers/schedulers/scheduling_dpmsolver_sde.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Katherine Crowson, The HuggingFace Team and hlky. All rights reserved.
+# Copyright 2025 Katherine Crowson, The HuggingFace Team and hlky. 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.
@@ -352,7 +352,7 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
 
         num_train_timesteps = num_train_timesteps or self.config.num_train_timesteps
 
-        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/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":

diffusers/schedulers/scheduling_dpmsolver_singlestep.py

@@ -1,4 +1,4 @@
-# Copyright 2024 TSAIL Team and The HuggingFace Team. All rights reserved.
+# Copyright 2025 TSAIL Team and 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.
@@ -218,7 +218,7 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
 
         if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"] and final_sigmas_type == "zero":
             raise ValueError(
-                f"`final_sigmas_type` {final_sigmas_type} is not supported for `algorithm_type` {algorithm_type}. Please chooose `sigma_min` instead."
+                f"`final_sigmas_type` {final_sigmas_type} is not supported for `algorithm_type` {algorithm_type}. Please choose `sigma_min` instead."
             )
 
         # setable values
@@ -410,7 +410,7 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
         pixels from saturation at each step. We find that dynamic thresholding results in significantly better
         photorealism as well as better image-text alignment, especially when using very large guidance weights."
 
-        https://arxiv.org/abs/2205.11487
+        https://huggingface.co/papers/2205.11487
         """
         dtype = sample.dtype
         batch_size, channels, *remaining_dims = sample.shape
@@ -584,7 +584,7 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
             if len(args) > 1:
                 sample = args[1]
             else:
-                raise ValueError("missing `sample` as a required keyward argument")
+                raise ValueError("missing `sample` as a required keyword argument")
         if timestep is not None:
             deprecate(
                 "timesteps",
@@ -681,7 +681,7 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
             if len(args) > 2:
                 sample = args[2]
             else:
-                raise ValueError(" missing `sample` as a required keyward argument")
+                raise ValueError("missing `sample` as a required keyword argument")
         if timestep is not None:
             deprecate(
                 "timesteps",
@@ -746,7 +746,7 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
             if len(args) > 2:
                 sample = args[2]
             else:
-                raise ValueError(" missing `sample` as a required keyward argument")
+                raise ValueError("missing `sample` as a required keyword argument")
         if timestep_list is not None:
             deprecate(
                 "timestep_list",
@@ -780,7 +780,7 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
         r0 = h_0 / h
         D0, D1 = m1, (1.0 / r0) * (m0 - m1)
         if self.config.algorithm_type == "dpmsolver++":
-            # See https://arxiv.org/abs/2211.01095 for detailed derivations
+            # See https://huggingface.co/papers/2211.01095 for detailed derivations
             if self.config.solver_type == "midpoint":
                 x_t = (
                     (sigma_t / sigma_s1) * sample
@@ -794,7 +794,7 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
                     + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
                 )
         elif self.config.algorithm_type == "dpmsolver":
-            # See https://arxiv.org/abs/2206.00927 for detailed derivations
+            # See https://huggingface.co/papers/2206.00927 for detailed derivations
             if self.config.solver_type == "midpoint":
                 x_t = (
                     (alpha_t / alpha_s1) * sample
@@ -858,7 +858,7 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
             if len(args) > 2:
                 sample = args[2]
             else:
-                raise ValueError(" missing`sample` as a required keyward argument")
+                raise ValueError("missing `sample` as a required keyword argument")
         if timestep_list is not None:
             deprecate(
                 "timestep_list",
@@ -899,7 +899,7 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
         D1 = (r0 * D1_0 - r1 * D1_1) / (r0 - r1)
         D2 = 2.0 * (D1_1 - D1_0) / (r0 - r1)
         if self.config.algorithm_type == "dpmsolver++":
-            # See https://arxiv.org/abs/2206.00927 for detailed derivations
+            # See https://huggingface.co/papers/2206.00927 for detailed derivations
             if self.config.solver_type == "midpoint":
                 x_t = (
                     (sigma_t / sigma_s2) * sample
@@ -914,7 +914,7 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
                     - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2
                 )
         elif self.config.algorithm_type == "dpmsolver":
-            # See https://arxiv.org/abs/2206.00927 for detailed derivations
+            # See https://huggingface.co/papers/2206.00927 for detailed derivations
             if self.config.solver_type == "midpoint":
                 x_t = (
                     (alpha_t / alpha_s2) * sample
@@ -981,12 +981,12 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
             if len(args) > 2:
                 sample = args[2]
             else:
-                raise ValueError(" missing`sample` as a required keyward argument")
+                raise ValueError("missing `sample` as a required keyword argument")
         if order is None:
             if len(args) > 3:
                 order = args[3]
             else:
-                raise ValueError(" missing `order` as a required keyward argument")
+                raise ValueError("missing `order` as a required keyword argument")
         if timestep_list is not None:
             deprecate(
                 "timestep_list",

diffusers/schedulers/scheduling_edm_dpmsolver_multistep.py

@@ -1,4 +1,4 @@
-# Copyright 2024 TSAIL Team and The HuggingFace Team. All rights reserved.
+# Copyright 2025 TSAIL Team and 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.
@@ -31,7 +31,7 @@ class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
     `EDMDPMSolverMultistepScheduler` is a fast dedicated high-order solver for diffusion ODEs.
 
     [1] Karras, Tero, et al. "Elucidating the Design Space of Diffusion-Based Generative Models."
-    https://arxiv.org/abs/2206.00364
+    https://huggingface.co/papers/2206.00364
 
     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.
@@ -47,8 +47,8 @@ class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
             The standard deviation of the data distribution. This is set to 0.5 in the EDM paper [1].
         sigma_schedule (`str`, *optional*, defaults to `karras`):
             Sigma schedule to compute the `sigmas`. By default, we the schedule introduced in the EDM paper
-            (https://arxiv.org/abs/2206.00364). Other acceptable value is "exponential". The exponential schedule was
-            incorporated in this model: https://huggingface.co/stabilityai/cosxl.
+            (https://huggingface.co/papers/2206.00364). Other acceptable value is "exponential". The exponential
+            schedule was incorporated in this model: https://huggingface.co/stabilityai/cosxl.
         num_train_timesteps (`int`, defaults to 1000):
             The number of diffusion steps to train the model.
         solver_order (`int`, defaults to 2):
@@ -176,7 +176,7 @@ class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
 
     # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_inputs
     def precondition_inputs(self, sample, sigma):
-        c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5)
+        c_in = self._get_conditioning_c_in(sigma)
         scaled_sample = sample * c_in
         return scaled_sample
 
@@ -305,7 +305,7 @@ class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         pixels from saturation at each step. We find that dynamic thresholding results in significantly better
         photorealism as well as better image-text alignment, especially when using very large guidance weights."
 
-        https://arxiv.org/abs/2205.11487
+        https://huggingface.co/papers/2205.11487
         """
         dtype = sample.dtype
         batch_size, channels, *remaining_dims = sample.shape
@@ -472,7 +472,7 @@ class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         r0 = h_0 / h
         D0, D1 = m0, (1.0 / r0) * (m0 - m1)
         if self.config.algorithm_type == "dpmsolver++":
-            # See https://arxiv.org/abs/2211.01095 for detailed derivations
+            # See https://huggingface.co/papers/2211.01095 for detailed derivations
             if self.config.solver_type == "midpoint":
                 x_t = (
                     (sigma_t / sigma_s0) * sample
@@ -548,7 +548,7 @@ class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1)
         D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1)
         if self.config.algorithm_type == "dpmsolver++":
-            # See https://arxiv.org/abs/2206.00927 for detailed derivations
+            # See https://huggingface.co/papers/2206.00927 for detailed derivations
             x_t = (
                 (sigma_t / sigma_s0) * sample
                 - (alpha_t * (torch.exp(-h) - 1.0)) * D0
@@ -703,5 +703,10 @@ class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         noisy_samples = original_samples + noise * sigma
         return noisy_samples
 
+    # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._get_conditioning_c_in
+    def _get_conditioning_c_in(self, sigma):
+        c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5)
+        return c_in
+
     def __len__(self):
         return self.config.num_train_timesteps

diffusers/schedulers/scheduling_edm_euler.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Katherine Crowson and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Katherine Crowson and 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.
@@ -51,7 +51,7 @@ class EDMEulerScheduler(SchedulerMixin, ConfigMixin):
     Implements the Euler scheduler in EDM formulation as presented in Karras et al. 2022 [1].
 
     [1] Karras, Tero, et al. "Elucidating the Design Space of Diffusion-Based Generative Models."
-    https://arxiv.org/abs/2206.00364
+    https://huggingface.co/papers/2206.00364
 
     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.
@@ -67,8 +67,8 @@ class EDMEulerScheduler(SchedulerMixin, ConfigMixin):
             The standard deviation of the data distribution. This is set to 0.5 in the EDM paper [1].
         sigma_schedule (`str`, *optional*, defaults to `karras`):
             Sigma schedule to compute the `sigmas`. By default, we the schedule introduced in the EDM paper
-            (https://arxiv.org/abs/2206.00364). Other acceptable value is "exponential". The exponential schedule was
-            incorporated in this model: https://huggingface.co/stabilityai/cosxl.
+            (https://huggingface.co/papers/2206.00364). Other acceptable value is "exponential". The exponential
+            schedule was incorporated in this model: https://huggingface.co/stabilityai/cosxl.
         num_train_timesteps (`int`, defaults to 1000):
             The number of diffusion steps to train the model.
         prediction_type (`str`, defaults to `epsilon`, *optional*):
@@ -103,11 +103,13 @@ class EDMEulerScheduler(SchedulerMixin, ConfigMixin):
         # setable values
         self.num_inference_steps = None
 
-        sigmas = torch.arange(num_train_timesteps + 1) / num_train_timesteps
+        sigmas_dtype = torch.float32 if torch.backends.mps.is_available() else torch.float64
+        sigmas = torch.arange(num_train_timesteps + 1, dtype=sigmas_dtype) / num_train_timesteps
         if sigma_schedule == "karras":
             sigmas = self._compute_karras_sigmas(sigmas)
         elif sigma_schedule == "exponential":
             sigmas = self._compute_exponential_sigmas(sigmas)
+        sigmas = sigmas.to(torch.float32)
 
         self.timesteps = self.precondition_noise(sigmas)
 
@@ -159,7 +161,7 @@ class EDMEulerScheduler(SchedulerMixin, ConfigMixin):
         self._begin_index = begin_index
 
     def precondition_inputs(self, sample, sigma):
-        c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5)
+        c_in = self._get_conditioning_c_in(sigma)
         scaled_sample = sample * c_in
         return scaled_sample
 
@@ -230,18 +232,19 @@ class EDMEulerScheduler(SchedulerMixin, ConfigMixin):
         """
         self.num_inference_steps = num_inference_steps
 
+        sigmas_dtype = torch.float32 if torch.backends.mps.is_available() else torch.float64
         if sigmas is None:
-            sigmas = torch.linspace(0, 1, self.num_inference_steps)
+            sigmas = torch.linspace(0, 1, self.num_inference_steps, dtype=sigmas_dtype)
         elif isinstance(sigmas, float):
-            sigmas = torch.tensor(sigmas, dtype=torch.float32)
+            sigmas = torch.tensor(sigmas, dtype=sigmas_dtype)
         else:
-            sigmas = sigmas
+            sigmas = sigmas.to(sigmas_dtype)
         if self.config.sigma_schedule == "karras":
             sigmas = self._compute_karras_sigmas(sigmas)
         elif self.config.sigma_schedule == "exponential":
             sigmas = self._compute_exponential_sigmas(sigmas)
-
         sigmas = sigmas.to(dtype=torch.float32, device=device)
+
         self.timesteps = self.precondition_noise(sigmas)
 
         if self.config.final_sigmas_type == "sigma_min":
@@ -315,6 +318,7 @@ class EDMEulerScheduler(SchedulerMixin, ConfigMixin):
         s_noise: float = 1.0,
         generator: Optional[torch.Generator] = None,
         return_dict: bool = True,
+        pred_original_sample: Optional[torch.Tensor] = None,
     ) -> Union[EDMEulerSchedulerOutput, Tuple]:
         """
         Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
@@ -378,7 +382,8 @@ class EDMEulerScheduler(SchedulerMixin, ConfigMixin):
             sample = sample + eps * (sigma_hat**2 - sigma**2) ** 0.5
 
         # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
-        pred_original_sample = self.precondition_outputs(sample, model_output, sigma_hat)
+        if pred_original_sample is None:
+            pred_original_sample = self.precondition_outputs(sample, model_output, sigma_hat)
 
         # 2. Convert to an ODE derivative
         derivative = (sample - pred_original_sample) / sigma_hat
@@ -435,5 +440,9 @@ class EDMEulerScheduler(SchedulerMixin, ConfigMixin):
         noisy_samples = original_samples + noise * sigma
         return noisy_samples
 
+    def _get_conditioning_c_in(self, sigma):
+        c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5)
+        return c_in
+
     def __len__(self):
         return self.config.num_train_timesteps

diffusers/schedulers/scheduling_euler_ancestral_discrete.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Katherine Crowson and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Katherine Crowson and 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.
@@ -95,7 +95,7 @@ def betas_for_alpha_bar(
 # Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
 def rescale_zero_terminal_snr(betas):
     """
-    Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+    Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1)
 
 
     Args:
@@ -286,7 +286,7 @@ class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
         """
         self.num_inference_steps = num_inference_steps
 
-        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891
         if self.config.timestep_spacing == "linspace":
             timesteps = np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[
                 ::-1

diffusers/schedulers/scheduling_euler_discrete.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Katherine Crowson and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Katherine Crowson and 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.
@@ -98,7 +98,7 @@ def betas_for_alpha_bar(
 # Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
 def rescale_zero_terminal_snr(betas):
     """
-    Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+    Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1)
 
 
     Args:
@@ -376,7 +376,7 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
             if timesteps is not None:
                 timesteps = np.array(timesteps).astype(np.float32)
             else:
-                # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+                # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891
                 if self.config.timestep_spacing == "linspace":
                     timesteps = np.linspace(
                         0, self.config.num_train_timesteps - 1, num_inference_steps, dtype=np.float32

diffusers/schedulers/scheduling_euler_discrete_flax.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Katherine Crowson and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Katherine Crowson and 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.
@@ -52,8 +52,8 @@ class FlaxEulerDiscreteSchedulerOutput(FlaxSchedulerOutput):
 
 class FlaxEulerDiscreteScheduler(FlaxSchedulerMixin, ConfigMixin):
     """
-    Euler scheduler (Algorithm 2) from Karras et al. (2022) https://arxiv.org/abs/2206.00364. . Based on the original
-    k-diffusion implementation by Katherine Crowson:
+    Euler scheduler (Algorithm 2) from Karras et al. (2022) https://huggingface.co/papers/2206.00364. . Based on the
+    original k-diffusion implementation by Katherine Crowson:
     https://github.com/crowsonkb/k-diffusion/blob/481677d114f6ea445aa009cf5bd7a9cdee909e47/k_diffusion/sampling.py#L51
 
 

diffusers/schedulers/scheduling_flow_match_euler_discrete.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Stability AI, Katherine Crowson and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Stability AI, Katherine Crowson and 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.
@@ -80,6 +80,8 @@ class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
             Whether to use beta sigmas for step sizes in the noise schedule during sampling.
         time_shift_type (`str`, defaults to "exponential"):
             The type of dynamic resolution-dependent timestep shifting to apply. Either "exponential" or "linear".
+        stochastic_sampling (`bool`, defaults to False):
+            Whether to use stochastic sampling.
     """
 
     _compatibles = []
@@ -101,6 +103,7 @@ class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
         use_exponential_sigmas: Optional[bool] = False,
         use_beta_sigmas: Optional[bool] = False,
         time_shift_type: str = "exponential",
+        stochastic_sampling: bool = False,
     ):
         if self.config.use_beta_sigmas and not is_scipy_available():
             raise ImportError("Make sure to install scipy if you want to use beta sigmas.")
@@ -437,13 +440,25 @@ class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
             lower_mask = sigmas < per_token_sigmas[None] - 1e-6
             lower_sigmas = lower_mask * sigmas
             lower_sigmas, _ = lower_sigmas.max(dim=0)
-            dt = (per_token_sigmas - lower_sigmas)[..., None]
+
+            current_sigma = per_token_sigmas[..., None]
+            next_sigma = lower_sigmas[..., None]
+            dt = current_sigma - next_sigma
         else:
-            sigma = self.sigmas[self.step_index]
-            sigma_next = self.sigmas[self.step_index + 1]
+            sigma_idx = self.step_index
+            sigma = self.sigmas[sigma_idx]
+            sigma_next = self.sigmas[sigma_idx + 1]
+
+            current_sigma = sigma
+            next_sigma = sigma_next
             dt = sigma_next - sigma
 
-        prev_sample = sample + dt * model_output
+        if self.config.stochastic_sampling:
+            x0 = sample - current_sigma * model_output
+            noise = torch.randn_like(sample)
+            prev_sample = (1.0 - next_sigma) * x0 + next_sigma * noise
+        else:
+            prev_sample = sample + dt * model_output
 
         # upon completion increase step index by one
         self._step_index += 1

diffusers/schedulers/scheduling_flow_match_heun_discrete.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Stability AI, Katherine Crowson and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Stability AI, Katherine Crowson and 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.