diffusers 0.33.0__py3-none-any.whl → 0.34.0__py3-none-any.whl

diffusers/quantizers/quantization_config.py

@@ -75,7 +75,7 @@ class QuantizationConfigMixin:
         Args:
             config_dict (`Dict[str, Any]`):
                 Dictionary that will be used to instantiate the configuration object.
-            return_unused_kwargs (`bool`,*optional*, defaults to `False`):
+            return_unused_kwargs (`bool`, *optional*, defaults to `False`):
                 Whether or not to return a list of unused keyword arguments. Used for `from_pretrained` method in
                 `PreTrainedModel`.
             kwargs (`Dict[str, Any]`):
@@ -179,7 +179,7 @@ class BitsAndBytesConfig(QuantizationConfigMixin):
     This is a wrapper class about all possible attributes and features that you can play with a model that has been
     loaded using `bitsandbytes`.
 
-    This replaces `load_in_8bit` or `load_in_4bit`therefore both options are mutually exclusive.
+    This replaces `load_in_8bit` or `load_in_4bit` therefore both options are mutually exclusive.
 
     Currently only supports `LLM.int8()`, `FP4`, and `NF4` quantization. If more methods are added to `bitsandbytes`,
     then more arguments will be added to this class.
@@ -192,10 +192,10 @@ class BitsAndBytesConfig(QuantizationConfigMixin):
             `bitsandbytes`.
         llm_int8_threshold (`float`, *optional*, defaults to 6.0):
             This corresponds to the outlier threshold for outlier detection as described in `LLM.int8() : 8-bit Matrix
-            Multiplication for Transformers at Scale` paper: https://arxiv.org/abs/2208.07339 Any hidden states value
-            that is above this threshold will be considered an outlier and the operation on those values will be done
-            in fp16. Values are usually normally distributed, that is, most values are in the range [-3.5, 3.5], but
-            there are some exceptional systematic outliers that are very differently distributed for large models.
+            Multiplication for Transformers at Scale` paper: https://huggingface.co/papers/2208.07339 Any hidden states
+            value that is above this threshold will be considered an outlier and the operation on those values will be
+            done in fp16. Values are usually normally distributed, that is, most values are in the range [-3.5, 3.5],
+            but there are some exceptional systematic outliers that are very differently distributed for large models.
             These outliers are often in the interval [-60, -6] or [6, 60]. Int8 quantization works well for values of
             magnitude ~5, but beyond that, there is a significant performance penalty. A good default threshold is 6,
             but a lower threshold might be needed for more unstable models (small models, fine-tuning).
@@ -493,7 +493,7 @@ class TorchAoConfig(QuantizationConfigMixin):
         TORCHAO_QUANT_TYPE_METHODS = self._get_torchao_quant_type_to_method()
         if self.quant_type not in TORCHAO_QUANT_TYPE_METHODS.keys():
             is_floating_quant_type = self.quant_type.startswith("float") or self.quant_type.startswith("fp")
-            if is_floating_quant_type and not self._is_cuda_capability_atleast_8_9():
+            if is_floating_quant_type and not self._is_xpu_or_cuda_capability_atleast_8_9():
                 raise ValueError(
                     f"Requested quantization type: {self.quant_type} is not supported on GPUs with CUDA capability <= 8.9. You "
                     f"can check the CUDA capability of your GPU using `torch.cuda.get_device_capability()`."
@@ -645,7 +645,7 @@ class TorchAoConfig(QuantizationConfigMixin):
             QUANTIZATION_TYPES.update(INT8_QUANTIZATION_TYPES)
             QUANTIZATION_TYPES.update(UINTX_QUANTIZATION_DTYPES)
 
-            if cls._is_cuda_capability_atleast_8_9():
+            if cls._is_xpu_or_cuda_capability_atleast_8_9():
                 QUANTIZATION_TYPES.update(FLOATX_QUANTIZATION_TYPES)
 
             return QUANTIZATION_TYPES
@@ -655,14 +655,16 @@ class TorchAoConfig(QuantizationConfigMixin):
             )
 
     @staticmethod
-    def _is_cuda_capability_atleast_8_9() -> bool:
-        if not torch.cuda.is_available():
-            raise RuntimeError("TorchAO requires a CUDA compatible GPU and installation of PyTorch.")
-
-        major, minor = torch.cuda.get_device_capability()
-        if major == 8:
-            return minor >= 9
-        return major >= 9
+    def _is_xpu_or_cuda_capability_atleast_8_9() -> bool:
+        if torch.cuda.is_available():
+            major, minor = torch.cuda.get_device_capability()
+            if major == 8:
+                return minor >= 9
+            return major >= 9
+        elif torch.xpu.is_available():
+            return True
+        else:
+            raise RuntimeError("TorchAO requires a CUDA compatible GPU or Intel XPU and installation of PyTorch.")
 
     def get_apply_tensor_subclass(self):
         TORCHAO_QUANT_TYPE_METHODS = self._get_torchao_quant_type_to_method()

diffusers/quantizers/quanto/quanto_quantizer.py

@@ -175,3 +175,7 @@ class QuantoQuantizer(DiffusersQuantizer):
     @property
     def is_serializable(self):
         return True
+
+    @property
+    def is_compileable(self) -> bool:
+        return True

diffusers/quantizers/torchao/torchao_quantizer.py

@@ -262,7 +262,7 @@ class TorchAoHfQuantizer(DiffusersQuantizer):
         **kwargs,
     ):
         r"""
-        Each nn.Linear layer that needs to be quantized is processsed here. First, we set the value the weight tensor,
+        Each nn.Linear layer that needs to be quantized is processed here. First, we set the value the weight tensor,
         then we move it to the target device. Finally, we quantize the module.
         """
         module, tensor_name = get_module_from_name(model, param_name)
@@ -335,3 +335,7 @@ class TorchAoHfQuantizer(DiffusersQuantizer):
     @property
     def is_trainable(self):
         return self.quantization_config.quant_type.startswith("int8")
+
+    @property
+    def is_compileable(self) -> bool:
+        return True

diffusers/schedulers/__init__.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 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.
@@ -60,6 +60,7 @@ else:
     _import_structure["scheduling_euler_discrete"] = ["EulerDiscreteScheduler"]
     _import_structure["scheduling_flow_match_euler_discrete"] = ["FlowMatchEulerDiscreteScheduler"]
     _import_structure["scheduling_flow_match_heun_discrete"] = ["FlowMatchHeunDiscreteScheduler"]
+    _import_structure["scheduling_flow_match_lcm"] = ["FlowMatchLCMScheduler"]
     _import_structure["scheduling_heun_discrete"] = ["HeunDiscreteScheduler"]
     _import_structure["scheduling_ipndm"] = ["IPNDMScheduler"]
     _import_structure["scheduling_k_dpm_2_ancestral_discrete"] = ["KDPM2AncestralDiscreteScheduler"]
@@ -161,6 +162,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .scheduling_euler_discrete import EulerDiscreteScheduler
         from .scheduling_flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler
         from .scheduling_flow_match_heun_discrete import FlowMatchHeunDiscreteScheduler
+        from .scheduling_flow_match_lcm import FlowMatchLCMScheduler
         from .scheduling_heun_discrete import HeunDiscreteScheduler
         from .scheduling_ipndm import IPNDMScheduler
         from .scheduling_k_dpm_2_ancestral_discrete import KDPM2AncestralDiscreteScheduler

diffusers/schedulers/deprecated/scheduling_karras_ve.py

@@ -1,4 +1,4 @@
-# Copyright 2024 NVIDIA and The HuggingFace Team. All rights reserved.
+# Copyright 2025 NVIDIA 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.
@@ -55,8 +55,9 @@ class KarrasVeScheduler(SchedulerMixin, ConfigMixin):
 
     <Tip>
 
-    For more details on the parameters, see [Appendix E](https://arxiv.org/abs/2206.00364). The grid search values used
-    to find the optimal `{s_noise, s_churn, s_min, s_max}` for a specific model are described in Table 5 of the paper.
+    For more details on the parameters, see [Appendix E](https://huggingface.co/papers/2206.00364). The grid search
+    values used to find the optimal `{s_noise, s_churn, s_min, s_max}` for a specific model are described in Table 5 of
+    the paper.
 
     </Tip>
 

diffusers/schedulers/deprecated/scheduling_sde_vp.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Google Brain and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Google Brain 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.

diffusers/schedulers/scheduling_consistency_models.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 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.

diffusers/schedulers/scheduling_cosine_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.
@@ -30,7 +30,7 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
     Implements a variant of `DPMSolverMultistepScheduler` with cosine schedule, proposed by Nichol and Dhariwal (2021).
     This scheduler was used in Stable Audio Open [1].
 
-    [1] Evans, Parker, et al. "Stable Audio Open" https://arxiv.org/abs/2407.14358
+    [1] Evans, Parker, et al. "Stable Audio Open" https://huggingface.co/papers/2407.14358
 
     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.
@@ -44,8 +44,8 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
             The standard deviation of the data distribution. This is set to 1.0 in Stable Audio Open [1].
         sigma_schedule (`str`, *optional*, defaults to `exponential`):
             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):
@@ -144,7 +144,7 @@ class CosineDPMSolverMultistepScheduler(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
 
@@ -568,5 +568,10 @@ class CosineDPMSolverMultistepScheduler(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_ddim.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Stanford University Team and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Stanford University 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.
@@ -94,7 +94,7 @@ def betas_for_alpha_bar(
 
 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:
@@ -269,7 +269,7 @@ class DDIMScheduler(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
@@ -312,7 +312,7 @@ class DDIMScheduler(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)
@@ -387,7 +387,7 @@ class DDIMScheduler(SchedulerMixin, ConfigMixin):
                 "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
             )
 
-        # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
+        # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502
         # Ideally, read DDIM paper in-detail understanding
 
         # Notation (<variable name> -> <name in paper>
@@ -408,7 +408,7 @@ class DDIMScheduler(SchedulerMixin, ConfigMixin):
         beta_prod_t = 1 - alpha_prod_t
 
         # 3. compute predicted original sample from predicted noise also called
-        # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502
         if self.config.prediction_type == "epsilon":
             pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
             pred_epsilon = model_output
@@ -441,10 +441,10 @@ class DDIMScheduler(SchedulerMixin, ConfigMixin):
             # the pred_epsilon is always re-derived from the clipped x_0 in Glide
             pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
 
-        # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # 6. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502
         pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon
 
-        # 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # 7. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502
         prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
 
         if eta > 0:

diffusers/schedulers/scheduling_ddim_cogvideox.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
+# Copyright 2025 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
 # All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -94,7 +94,7 @@ def betas_for_alpha_bar(
 
 def rescale_zero_terminal_snr(alphas_cumprod):
     """
-    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:
@@ -275,7 +275,7 @@ class CogVideoXDDIMScheduler(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)
@@ -350,7 +350,7 @@ class CogVideoXDDIMScheduler(SchedulerMixin, ConfigMixin):
                 "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
             )
 
-        # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
+        # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502
         # Ideally, read DDIM paper in-detail understanding
 
         # Notation (<variable name> -> <name in paper>
@@ -371,7 +371,7 @@ class CogVideoXDDIMScheduler(SchedulerMixin, ConfigMixin):
         beta_prod_t = 1 - alpha_prod_t
 
         # 3. compute predicted original sample from predicted noise also called
-        # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502
         # To make style tests pass, commented out `pred_epsilon` as it is an unused variable
         if self.config.prediction_type == "epsilon":
             pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)

diffusers/schedulers/scheduling_ddim_flax.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Stanford University Team and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Stanford University 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.
@@ -73,7 +73,7 @@ class FlaxDDIMScheduler(FlaxSchedulerMixin, ConfigMixin):
     [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
     [`~SchedulerMixin.from_pretrained`] functions.
 
-    For more details, see the original paper: https://arxiv.org/abs/2010.02502
+    For more details, see the original paper: https://huggingface.co/papers/2010.02502
 
     Args:
         num_train_timesteps (`int`): number of diffusion steps used to train the model.
@@ -230,7 +230,7 @@ class FlaxDDIMScheduler(FlaxSchedulerMixin, ConfigMixin):
                 "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
             )
 
-        # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
+        # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502
         # Ideally, read DDIM paper in-detail understanding
 
         # Notation (<variable name> -> <name in paper>
@@ -254,7 +254,7 @@ class FlaxDDIMScheduler(FlaxSchedulerMixin, ConfigMixin):
         beta_prod_t = 1 - alpha_prod_t
 
         # 3. compute predicted original sample from predicted noise also called
-        # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502
         if self.config.prediction_type == "epsilon":
             pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
             pred_epsilon = model_output
@@ -281,10 +281,10 @@ class FlaxDDIMScheduler(FlaxSchedulerMixin, ConfigMixin):
         variance = self._get_variance(state, timestep, prev_timestep)
         std_dev_t = eta * variance ** (0.5)
 
-        # 5. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # 5. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502
         pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon
 
-        # 6. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # 6. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502
         prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
 
         if not return_dict:

diffusers/schedulers/scheduling_ddim_inverse.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 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.
@@ -93,7 +93,7 @@ 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:
@@ -266,7 +266,7 @@ class DDIMInverseScheduler(SchedulerMixin, ConfigMixin):
 
         self.num_inference_steps = num_inference_steps
 
-        # "leading" and "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+        # "leading" and "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891
         if self.config.timestep_spacing == "leading":
             step_ratio = self.config.num_train_timesteps // self.num_inference_steps
             # creates integer timesteps by multiplying by ratio
@@ -338,7 +338,7 @@ class DDIMInverseScheduler(SchedulerMixin, ConfigMixin):
         beta_prod_t = 1 - alpha_prod_t
 
         # 3. compute predicted original sample from predicted noise also called
-        # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502
         if self.config.prediction_type == "epsilon":
             pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
             pred_epsilon = model_output
@@ -360,10 +360,10 @@ class DDIMInverseScheduler(SchedulerMixin, ConfigMixin):
                 -self.config.clip_sample_range, self.config.clip_sample_range
             )
 
-        # 5. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # 5. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502
         pred_sample_direction = (1 - alpha_prod_t_prev) ** (0.5) * pred_epsilon
 
-        # 6. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # 6. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502
         prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
 
         if not return_dict:

diffusers/schedulers/scheduling_ddim_parallel.py

@@ -1,4 +1,4 @@
-# Copyright 2024 ParaDiGMS authors and The HuggingFace Team. All rights reserved.
+# Copyright 2025 ParaDiGMS authors 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:
@@ -139,7 +139,7 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
     [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
     [`~SchedulerMixin.from_pretrained`] functions.
 
-    For more details, see the original paper: https://arxiv.org/abs/2010.02502
+    For more details, see the original paper: https://huggingface.co/papers/2010.02502
 
     Args:
         num_train_timesteps (`int`): number of diffusion steps used to train the model.
@@ -165,21 +165,21 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
             process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4
             https://imagen.research.google/video/paper.pdf)
         thresholding (`bool`, default `False`):
-            whether to use the "dynamic thresholding" method (introduced by Imagen, https://arxiv.org/abs/2205.11487).
-            Note that the thresholding method is unsuitable for latent-space diffusion models (such as
-            stable-diffusion).
+            whether to use the "dynamic thresholding" method (introduced by Imagen,
+            https://huggingface.co/papers/2205.11487). Note that the thresholding method is unsuitable for latent-space
+            diffusion models (such as stable-diffusion).
         dynamic_thresholding_ratio (`float`, default `0.995`):
             the ratio for the dynamic thresholding method. Default is `0.995`, the same as Imagen
-            (https://arxiv.org/abs/2205.11487). Valid only when `thresholding=True`.
+            (https://huggingface.co/papers/2205.11487). Valid only when `thresholding=True`.
         sample_max_value (`float`, default `1.0`):
             the threshold value for dynamic thresholding. Valid only when `thresholding=True`.
         timestep_spacing (`str`, default `"leading"`):
             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 are Flawed](https://huggingface.co/papers/2305.08891) for more information.
         rescale_betas_zero_snr (`bool`, default `False`):
-            whether to rescale the betas to have zero terminal SNR (proposed by https://arxiv.org/pdf/2305.08891.pdf).
-            This can enable the model to generate very bright and dark samples instead of limiting it to samples with
-            medium brightness. Loosely related to
+            whether to rescale the betas to have zero terminal SNR (proposed by
+            https://huggingface.co/papers/2305.08891). This can enable the model to generate very bright and dark
+            samples instead of limiting it to samples with medium brightness. Loosely related to
             [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
     """
 
@@ -291,7 +291,7 @@ class DDIMParallelScheduler(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
@@ -335,7 +335,7 @@ class DDIMParallelScheduler(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)
@@ -390,7 +390,7 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
             generator: random number generator.
             variance_noise (`torch.Tensor`): instead of generating noise for the variance using `generator`, we
                 can directly provide the noise for the variance itself. This is useful for methods such as
-                CycleDiffusion. (https://arxiv.org/abs/2210.05559)
+                CycleDiffusion. (https://huggingface.co/papers/2210.05559)
             return_dict (`bool`): option for returning tuple rather than DDIMParallelSchedulerOutput class
 
         Returns:
@@ -404,7 +404,7 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
                 "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
             )
 
-        # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
+        # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502
         # Ideally, read DDIM paper in-detail understanding
 
         # Notation (<variable name> -> <name in paper>
@@ -425,7 +425,7 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
         beta_prod_t = 1 - alpha_prod_t
 
         # 3. compute predicted original sample from predicted noise also called
-        # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502
         if self.config.prediction_type == "epsilon":
             pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
             pred_epsilon = model_output
@@ -458,10 +458,10 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
             # the pred_epsilon is always re-derived from the clipped x_0 in Glide
             pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
 
-        # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # 6. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502
         pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon
 
-        # 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # 7. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502
         prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
 
         if eta > 0:
@@ -526,7 +526,7 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
 
         assert eta == 0.0
 
-        # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
+        # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502
         # Ideally, read DDIM paper in-detail understanding
 
         # Notation (<variable name> -> <name in paper>
@@ -554,7 +554,7 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
         beta_prod_t = 1 - alpha_prod_t
 
         # 3. compute predicted original sample from predicted noise also called
-        # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502
         if self.config.prediction_type == "epsilon":
             pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
             pred_epsilon = model_output
@@ -587,10 +587,10 @@ class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
             # the pred_epsilon is always re-derived from the clipped x_0 in Glide
             pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
 
-        # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # 6. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502
         pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon
 
-        # 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # 7. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502
         prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
 
         return prev_sample

diffusers/schedulers/scheduling_ddpm.py

@@ -1,4 +1,4 @@
-# Copyright 2024 UC Berkeley Team and The HuggingFace Team. All rights reserved.
+# Copyright 2025 UC Berkeley 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.
@@ -92,7 +92,7 @@ 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:
@@ -295,7 +295,7 @@ class DDPMScheduler(SchedulerMixin, ConfigMixin):
             self.num_inference_steps = num_inference_steps
             self.custom_timesteps = False
 
-            # "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)
@@ -329,7 +329,7 @@ class DDPMScheduler(SchedulerMixin, ConfigMixin):
         alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one
         current_beta_t = 1 - alpha_prod_t / alpha_prod_t_prev
 
-        # For t > 0, compute predicted variance βt (see formula (6) and (7) from https://arxiv.org/pdf/2006.11239.pdf)
+        # For t > 0, compute predicted variance βt (see formula (6) and (7) from https://huggingface.co/papers/2006.11239)
         # and sample from it to get previous sample
         # x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample
         variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * current_beta_t
@@ -343,7 +343,7 @@ class DDPMScheduler(SchedulerMixin, ConfigMixin):
         # hacks - were probably added for training stability
         if variance_type == "fixed_small":
             variance = variance
-        # for rl-diffuser https://arxiv.org/abs/2205.09991
+        # for rl-diffuser https://huggingface.co/papers/2205.09991
         elif variance_type == "fixed_small_log":
             variance = torch.log(variance)
             variance = torch.exp(0.5 * variance)
@@ -370,7 +370,7 @@ class DDPMScheduler(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
@@ -443,7 +443,7 @@ class DDPMScheduler(SchedulerMixin, ConfigMixin):
         current_beta_t = 1 - current_alpha_t
 
         # 2. compute predicted original sample from predicted noise also called
-        # "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
+        # "predicted x_0" of formula (15) from https://huggingface.co/papers/2006.11239
         if self.config.prediction_type == "epsilon":
             pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
         elif self.config.prediction_type == "sample":
@@ -465,12 +465,12 @@ class DDPMScheduler(SchedulerMixin, ConfigMixin):
             )
 
         # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t
-        # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
+        # See formula (7) from https://huggingface.co/papers/2006.11239
         pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * current_beta_t) / beta_prod_t
         current_sample_coeff = current_alpha_t ** (0.5) * beta_prod_t_prev / beta_prod_t
 
         # 5. Compute predicted previous sample µ_t
-        # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
+        # See formula (7) from https://huggingface.co/papers/2006.11239
         pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample
 
         # 6. Add noise