up

docs/source/api/schedulers.mdx

@@ -76,6 +76,33 @@ Original paper can be found [here](https://arxiv.org/abs/2206.00927) and the [im
 
 [[autodoc]] DPMSolverMultistepScheduler
 
+#### Heun scheduler inspired by Karras et. al paper
+
+Algorithm 1 of [Karras et. al](https://arxiv.org/abs/2206.00364).
+Scheduler ported from @crowsonkb's https://github.com/crowsonkb/k-diffusion library:
+
+All credit for making this scheduler work goes to [Katherine Crowson](https://github.com/crowsonkb/)
+
+[[autodoc]] HeunDiscreteScheduler
+
+#### DPM Discrete Scheduler inspired by Karras et. al paper
+
+Inspired by [Karras et. al](https://arxiv.org/abs/2206.00364).
+Scheduler ported from @crowsonkb's https://github.com/crowsonkb/k-diffusion library:
+
+All credit for making this scheduler work goes to [Katherine Crowson](https://github.com/crowsonkb/)
+
+[[autodoc]] KDPM2DiscreteScheduler
+
+#### DPM Discrete Scheduler with ancestral sampling inspired by Karras et. al paper
+
+Inspired by [Karras et. al](https://arxiv.org/abs/2206.00364).
+Scheduler ported from @crowsonkb's https://github.com/crowsonkb/k-diffusion library:
+
+All credit for making this scheduler work goes to [Katherine Crowson](https://github.com/crowsonkb/)
+
+[[autodoc]] KDPM2AncestralDiscreteScheduler
+
 #### Variance exploding, stochastic sampling from Karras et. al
 
 Original paper can be found [here](https://arxiv.org/abs/2006.11239).
@@ -86,7 +113,6 @@ Original paper can be found [here](https://arxiv.org/abs/2006.11239).
 
 Original implementation can be found [here](https://arxiv.org/abs/2206.00364).
 
-
 [[autodoc]] LMSDiscreteScheduler
 
 #### Pseudo numerical methods for diffusion models (PNDM)

scripts/convert_original_stable_diffusion_to_diffusers.py

@@ -666,17 +666,29 @@ if __name__ == "__main__":
 
     args = parser.parse_args()
 
-    if args.original_config_file is None:
-        os.system(
-            "wget https://raw.githubusercontent.com/CompVis/stable-diffusion/main/configs/stable-diffusion/v1-inference.yaml"
-        )
-        args.original_config_file = "./v1-inference.yaml"
-
-    original_config = OmegaConf.load(args.original_config_file)
-
     checkpoint = torch.load(args.checkpoint_path)
     checkpoint = checkpoint["state_dict"]
 
+    prediction_type = "epsilon"
+    if args.original_config_file is None:
+        key_name = "model.diffusion_model.input_blocks.2.1.transformer_blocks.0.attn2.to_k.weight"
+
+        if key_name in checkpoint and checkpoint[key_name].shape[-1] == 1024:
+            # model_type = "v2"
+            os.system(
+                "wget https://raw.githubusercontent.com/Stability-AI/stablediffusion/main/configs/stable-diffusion/v2-inference-v.yaml"
+            )
+            args.original_config_file = "./v2-inference-v.yaml"
+            prediction_type
+        else:
+            # model_type = "v2"
+            os.system(
+                "wget https://raw.githubusercontent.com/CompVis/stable-diffusion/main/configs/stable-diffusion/v1-inference.yaml"
+            )
+            args.original_config_file = "./v1-inference.yaml"
+
+    original_config = OmegaConf.load(args.original_config_file)
+
     num_train_timesteps = original_config.model.params.timesteps
     beta_start = original_config.model.params.linear_start
     beta_end = original_config.model.params.linear_end

scripts/v1-inference.yaml

@@ -0,0 +1,70 @@
+model:
+  base_learning_rate: 1.0e-04
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: "jpg"
+    cond_stage_key: "txt"
+    image_size: 64
+    channels: 4
+    cond_stage_trainable: false   # Note: different from the one we trained before
+    conditioning_key: crossattn
+    monitor: val/loss_simple_ema
+    scale_factor: 0.18215
+    use_ema: False
+
+    scheduler_config: # 10000 warmup steps
+      target: ldm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [ 10000 ]
+        cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
+        f_start: [ 1.e-6 ]
+        f_max: [ 1. ]
+        f_min: [ 1. ]
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 32 # unused
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [ 4, 2, 1 ]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 4, 4 ]
+        num_heads: 8
+        use_spatial_transformer: True
+        transformer_depth: 1
+        context_dim: 768
+        use_checkpoint: True
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

src/diffusers/__init__.py

@@ -49,6 +49,8 @@ if is_torch_available():
         HeunDiscreteScheduler,
         IPNDMScheduler,
         KarrasVeScheduler,
+        KDPM2AncestralDiscreteScheduler,
+        KDPM2DiscreteScheduler,
         PNDMScheduler,
         RePaintScheduler,
         SchedulerMixin,

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion.py

@@ -576,7 +576,7 @@ class AltDiffusionPipeline(DiffusionPipeline):
                     latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
 
                 # call the callback, if provided
-                if (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0:
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)

src/diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py

@@ -593,7 +593,7 @@ class AltDiffusionImg2ImgPipeline(DiffusionPipeline):
                 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
 
                 # call the callback, if provided
-                if (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0:
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)

src/diffusers/pipelines/stable_diffusion/pipeline_cycle_diffusion.py

@@ -681,7 +681,7 @@ class CycleDiffusionPipeline(DiffusionPipeline):
                 ).prev_sample
 
                 # call the callback, if provided
-                if (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0:
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -575,7 +575,7 @@ class StableDiffusionPipeline(DiffusionPipeline):
                     latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
 
                 # call the callback, if provided
-                if (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0:
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py

@@ -460,7 +460,7 @@ class StableDiffusionImageVariationPipeline(DiffusionPipeline):
                 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
 
                 # call the callback, if provided
-                if (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0:
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py

@@ -602,7 +602,7 @@ class StableDiffusionImg2ImgPipeline(DiffusionPipeline):
                 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
 
                 # call the callback, if provided
-                if (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0:
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py

@@ -721,7 +721,7 @@ class StableDiffusionInpaintPipeline(DiffusionPipeline):
                 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
 
                 # call the callback, if provided
-                if (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0:
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint_legacy.py

@@ -617,7 +617,7 @@ class StableDiffusionInpaintPipelineLegacy(DiffusionPipeline):
                 latents = (init_latents_proper * mask) + (latents * (1 - mask))
 
                 # call the callback, if provided
-                if (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0:
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py

@@ -535,7 +535,7 @@ class StableDiffusionUpscalePipeline(DiffusionPipeline):
                 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
 
                 # call the callback, if provided
-                if (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0:
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)

src/diffusers/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py

@@ -729,7 +729,7 @@ class StableDiffusionPipelineSafe(DiffusionPipeline):
                 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
 
                 # call the callback, if provided
-                if (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0:
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                     progress_bar.update()
                     if callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)

src/diffusers/schedulers/__init__.py

@@ -22,8 +22,10 @@ if is_torch_available():
     from .scheduling_dpmsolver_multistep import DPMSolverMultistepScheduler
     from .scheduling_euler_ancestral_discrete import EulerAncestralDiscreteScheduler
     from .scheduling_euler_discrete import EulerDiscreteScheduler
-    from .scheduling_heun import HeunDiscreteScheduler
+    from .scheduling_heun_discrete import HeunDiscreteScheduler
     from .scheduling_ipndm import IPNDMScheduler
+    from .scheduling_k_dpm_2_ancestral_discrete import KDPM2AncestralDiscreteScheduler
+    from .scheduling_k_dpm_2_discrete import KDPM2DiscreteScheduler
     from .scheduling_karras_ve import KarrasVeScheduler
     from .scheduling_pndm import PNDMScheduler
     from .scheduling_repaint import RePaintScheduler

src/diffusers/schedulers/scheduling_heun_discrete.py

@@ -24,14 +24,16 @@ from .scheduling_utils import SchedulerMixin, SchedulerOutput
 
 class HeunDiscreteScheduler(SchedulerMixin, ConfigMixin):
     """
-    Args:
     Implements Algorithm 2 (Heun steps) from Karras et al. (2022). for discrete beta schedules. Based on the original
     k-diffusion implementation by Katherine Crowson:
     https://github.com/crowsonkb/k-diffusion/blob/481677d114f6ea445aa009cf5bd7a9cdee909e47/k_diffusion/sampling.py#L90
+
     [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__`
     function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`.
-    [`~ConfigMixin`] also provides general loading and saving functionality via the [`~ConfigMixin.save_config`] and
-    [`~ConfigMixin.from_config`] functions.
+    [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
+    [`~SchedulerMixin.from_pretrained`] functions.
+
+    Args:
         num_train_timesteps (`int`): number of diffusion steps used to train the model. beta_start (`float`): the
         starting `beta` value of inference. beta_end (`float`): the final `beta` value. beta_schedule (`str`):
             the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
@@ -76,7 +78,7 @@ class HeunDiscreteScheduler(SchedulerMixin, ConfigMixin):
     def index_for_timestep(self, timestep):
         indices = (self.timesteps == timestep).nonzero()
         if self.state_in_first_order:
-            pos = 0 if indices.shape[0] < 2 else 1
+            pos = -1
         else:
             pos = 0
         return indices[pos].item()
@@ -131,7 +133,7 @@ class HeunDiscreteScheduler(SchedulerMixin, ConfigMixin):
         self.init_noise_sigma = self.sigmas.max()
 
         timesteps = torch.from_numpy(timesteps)
-        timesteps = torch.cat([timesteps[:1], timesteps[1:].repeat_interleave(2), timesteps[-1:]])
+        timesteps = torch.cat([timesteps[:1], timesteps[1:].repeat_interleave(2)])
 
         if str(device).startswith("mps"):
             # mps does not support float64

src/diffusers/schedulers/scheduling_k_dpm_2_ancestral_discrete.py

@@ -0,0 +1,268 @@
+# Copyright 2022 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.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import _COMPATIBLE_STABLE_DIFFUSION_SCHEDULERS
+from .scheduling_utils import SchedulerMixin, SchedulerOutput
+
+
+class KDPM2AncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
+    """
+    Scheduler created by @crowsonkb in [k_diffusion](https://github.com/crowsonkb/k-diffusion), see:
+    https://github.com/crowsonkb/k-diffusion/blob/5b3af030dd83e0297272d861c19477735d0317ec/k_diffusion/sampling.py#L188
+
+    Scheduler inspired by DPM-Solver-2 and Algorthim 2 from Karras et al. (2022).
+
+    [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__`
+    function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`.
+    [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
+    [`~SchedulerMixin.from_pretrained`] functions.
+
+    Args:
+        num_train_timesteps (`int`): number of diffusion steps used to train the model. beta_start (`float`): the
+        starting `beta` value of inference. beta_end (`float`): the final `beta` value. beta_schedule (`str`):
+            the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+            `linear` or `scaled_linear`.
+        trained_betas (`np.ndarray`, optional):
+            option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc.
+            options to clip the variance used when adding noise to the denoised sample. Choose from `fixed_small`,
+            `fixed_small_log`, `fixed_large`, `fixed_large_log`, `learned` or `learned_range`.
+        tensor_format (`str`): whether the scheduler expects pytorch or numpy arrays.
+    """
+
+    _compatibles = _COMPATIBLE_STABLE_DIFFUSION_SCHEDULERS.copy()
+    order = 2
+
+    @register_to_config
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        beta_start: float = 0.00085,  # sensible defaults
+        beta_end: float = 0.012,
+        beta_schedule: str = "linear",
+        trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+    ):
+        if trained_betas is not None:
+            self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+        elif beta_schedule == "linear":
+            self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+        elif beta_schedule == "scaled_linear":
+            # this schedule is very specific to the latent diffusion model.
+            self.betas = (
+                torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+            )
+        else:
+            raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+        self.alphas = 1.0 - self.betas
+        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+        #  set all values
+        self.set_timesteps(num_train_timesteps, None, num_train_timesteps)
+
+    def index_for_timestep(self, timestep):
+        indices = (self.timesteps == timestep).nonzero()
+        if self.state_in_first_order:
+            pos = -1
+        else:
+            pos = 0
+        return indices[pos].item()
+
+    def scale_model_input(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[float, torch.FloatTensor],
+    ) -> torch.FloatTensor:
+        """
+        Args:
+        Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+        current timestep.
+            sample (`torch.FloatTensor`): input sample timestep (`int`, optional): current timestep
+        Returns:
+            `torch.FloatTensor`: scaled input sample
+        """
+        step_index = self.index_for_timestep(timestep)
+
+        sigma = self.sigmas[step_index]
+        sample = sample / ((sigma**2 + 1) ** 0.5)
+        return sample
+
+    def set_timesteps(
+        self,
+        num_inference_steps: int,
+        device: Union[str, torch.device] = None,
+        num_train_timesteps: Optional[int] = None,
+    ):
+        """
+        Sets the timesteps used for the diffusion chain. Supporting function to be run before inference.
+
+        Args:
+            num_inference_steps (`int`):
+                the number of diffusion steps used when generating samples with a pre-trained model.
+            device (`str` or `torch.device`, optional):
+                the device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        """
+        self.num_inference_steps = num_inference_steps
+
+        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()
+
+        sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+        self.log_sigmas = torch.from_numpy(np.log(sigmas)).to(device)
+
+        sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
+        sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32)
+        sigmas = torch.from_numpy(sigmas).to(device=device)
+
+        # compute up and down sigmas
+        sigmas_next = sigmas.roll(-1)
+        sigmas_next[-1] = 0.0
+        sigmas_up = (sigmas_next**2 * (sigmas**2 - sigmas_next**2) / sigmas**2) ** 0.5
+        sigmas_down = (sigmas_next**2 - sigmas_up**2) ** 0.5
+        sigmas_down[-1] = 0.0
+
+        self.sigmas = torch.cat([sigmas[:1], sigmas[1:].repeat_interleave(2), sigmas[-1:]])
+        self.sigmas_up = torch.cat([sigmas_up[:1], sigmas_up[1:].repeat_interleave(2), sigmas_up[-1:]])
+        self.sigmas_down = torch.cat([sigmas_down[:1], sigmas_down[1:].repeat_interleave(2), sigmas_down[-1:]])
+
+        # standard deviation of the initial noise distribution
+        self.init_noise_sigma = self.sigmas.max()
+
+        timesteps = torch.from_numpy(timesteps)
+        timesteps = torch.cat([timesteps[:1], timesteps[1:].repeat_interleave(2)])
+
+        if str(device).startswith("mps"):
+            # mps does not support float64
+            self.timesteps = timesteps.to(device, dtype=torch.float32)
+        else:
+            self.timesteps = timesteps
+
+        self.sample = None
+
+    @property
+    def state_in_first_order(self):
+        return self.sample is None
+
+    def step(
+        self,
+        model_output: Union[torch.FloatTensor, np.ndarray],
+        timestep: Union[float, torch.FloatTensor],
+        sample: Union[torch.FloatTensor, np.ndarray],
+        generator: Optional[torch.Generator] = None,
+        return_dict: bool = True,
+    ) -> Union[SchedulerOutput, Tuple]:
+        """
+        Args:
+        Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+        process from the learned model outputs (most often the predicted noise).
+            model_output (`torch.FloatTensor` or `np.ndarray`): direct output from learned diffusion model. timestep
+            (`int`): current discrete timestep in the diffusion chain. sample (`torch.FloatTensor` or `np.ndarray`):
+                current instance of sample being created by diffusion process.
+            return_dict (`bool`): option for returning tuple rather than SchedulerOutput class
+        Returns:
+            [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+            [`~schedulers.scheduling_utils.SchedulerOutput`] if `return_dict` is True, otherwise a `tuple`. When
+            returning a tuple, the first element is the sample tensor.
+        """
+        step_index = self.index_for_timestep(timestep)
+
+        if self.state_in_first_order:
+            sigma = self.sigmas[step_index]
+            sigma_next = self.sigmas[step_index + 1]
+        else:
+            # 2nd order / KPDM2's method
+            sigma = self.sigmas[step_index - 1]
+            sigma_next = self.sigmas[step_index]
+            sigma_up = self.sigmas_up[step_index - 1]
+            sigma_down = self.sigmas_down[step_index - 1]
+
+        # currently only gamma=0 is supported. This usually works best anyways.
+        # We can support gamma in the future but then need to scale the timestep before
+        # passing it to the model which requires a change in API
+        gamma = 0
+        sigma_hat = sigma * (gamma + 1)  # Note: sigma_hat == sigma for now
+
+        device = model_output.device
+        if device.type == "mps":
+            # randn does not work reproducibly on mps
+            noise = torch.randn(model_output.shape, dtype=model_output.dtype, device="cpu", generator=generator).to(
+                device
+            )
+        else:
+            noise = torch.randn(model_output.shape, dtype=model_output.dtype, device=device, generator=generator).to(
+                device
+            )
+
+        # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
+        pred_original_sample = sample - sigma_hat * model_output
+
+        if self.state_in_first_order:
+            # 2. Convert to an ODE derivative
+            derivative = (sample - pred_original_sample) / sigma_hat
+            # 3. 1st order derivative
+            dt = sigma_next - sigma_hat
+
+            # store for 2nd order step
+            self.sample = sample
+            self.dt = dt
+            prev_sample = sample + derivative * dt
+        else:
+            # DPM-Solver-2
+            derivative = (sample - pred_original_sample) / sigma_hat
+            dt = sigma_down - sigma_hat
+
+            sample = self.sample
+            self.sample = None
+
+            prev_sample = sample + derivative * dt
+            prev_sample = prev_sample + noise * sigma_up
+
+        if not return_dict:
+            return (prev_sample,)
+
+        return SchedulerOutput(prev_sample=prev_sample)
+
+    def add_noise(
+        self,
+        original_samples: torch.FloatTensor,
+        noise: torch.FloatTensor,
+        timesteps: torch.FloatTensor,
+    ) -> torch.FloatTensor:
+        # Make sure sigmas and timesteps have the same device and dtype as original_samples
+        self.sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+        if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+            # mps does not support float64
+            self.timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+            timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+        else:
+            self.timesteps = self.timesteps.to(original_samples.device)
+            timesteps = timesteps.to(original_samples.device)
+
+        step_indices = [self.index_for_timestep(t) for t in timesteps]
+
+        sigma = self.sigmas[step_indices].flatten()
+        while len(sigma.shape) < len(original_samples.shape):
+            sigma = sigma.unsqueeze(-1)
+
+        noisy_samples = original_samples + noise * sigma
+        return noisy_samples
+
+    def __len__(self):
+        return self.config.num_train_timesteps

src/diffusers/schedulers/scheduling_k_dpm_2_discrete.py

@@ -0,0 +1,283 @@
+# Copyright 2022 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.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import _COMPATIBLE_STABLE_DIFFUSION_SCHEDULERS
+from .scheduling_utils import SchedulerMixin, SchedulerOutput
+
+
+class KDPM2DiscreteScheduler(SchedulerMixin, ConfigMixin):
+    """
+    Scheduler created by @crowsonkb in [k_diffusion](https://github.com/crowsonkb/k-diffusion), see:
+    https://github.com/crowsonkb/k-diffusion/blob/5b3af030dd83e0297272d861c19477735d0317ec/k_diffusion/sampling.py#L188
+
+    Scheduler inspired by DPM-Solver-2 and Algorthim 2 from Karras et al. (2022).
+
+    [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__`
+    function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`.
+    [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
+    [`~SchedulerMixin.from_pretrained`] functions.
+
+    Args:
+        num_train_timesteps (`int`): number of diffusion steps used to train the model. beta_start (`float`): the
+        starting `beta` value of inference. beta_end (`float`): the final `beta` value. beta_schedule (`str`):
+            the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+            `linear` or `scaled_linear`.
+        trained_betas (`np.ndarray`, optional):
+            option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc.
+            options to clip the variance used when adding noise to the denoised sample. Choose from `fixed_small`,
+            `fixed_small_log`, `fixed_large`, `fixed_large_log`, `learned` or `learned_range`.
+        tensor_format (`str`): whether the scheduler expects pytorch or numpy arrays.
+    """
+
+    _compatibles = _COMPATIBLE_STABLE_DIFFUSION_SCHEDULERS.copy()
+    order = 2
+
+    @register_to_config
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        beta_start: float = 0.00085,  # sensible defaults
+        beta_end: float = 0.012,
+        beta_schedule: str = "linear",
+        trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+    ):
+        if trained_betas is not None:
+            self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+        elif beta_schedule == "linear":
+            self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+        elif beta_schedule == "scaled_linear":
+            # this schedule is very specific to the latent diffusion model.
+            self.betas = (
+                torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+            )
+        else:
+            raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+        self.alphas = 1.0 - self.betas
+        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+        #  set all values
+        self.set_timesteps(num_train_timesteps, None, num_train_timesteps)
+
+    def index_for_timestep(self, timestep):
+        indices = (self.timesteps == timestep).nonzero()
+        if self.state_in_first_order:
+            pos = -1
+        else:
+            pos = 0
+        return indices[pos].item()
+
+    def scale_model_input(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[float, torch.FloatTensor],
+    ) -> torch.FloatTensor:
+        """
+        Args:
+        Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+        current timestep.
+            sample (`torch.FloatTensor`): input sample timestep (`int`, optional): current timestep
+        Returns:
+            `torch.FloatTensor`: scaled input sample
+        """
+        step_index = self.index_for_timestep(timestep)
+
+        if self.state_in_first_order:
+            sigma = self.sigmas[step_index]
+        else:
+            sigma = self.sigmas_interpol[step_index]
+
+        sample = sample / ((sigma**2 + 1) ** 0.5)
+        return sample
+
+    def set_timesteps(
+        self,
+        num_inference_steps: int,
+        device: Union[str, torch.device] = None,
+        num_train_timesteps: Optional[int] = None,
+    ):
+        """
+        Sets the timesteps used for the diffusion chain. Supporting function to be run before inference.
+
+        Args:
+            num_inference_steps (`int`):
+                the number of diffusion steps used when generating samples with a pre-trained model.
+            device (`str` or `torch.device`, optional):
+                the device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        """
+        self.num_inference_steps = num_inference_steps
+
+        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()
+
+        sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+        self.log_sigmas = torch.from_numpy(np.log(sigmas)).to(device)
+
+        sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
+        sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32)
+        sigmas = torch.from_numpy(sigmas).to(device=device)
+
+        # interpolate sigmas
+        sigmas_interpol = sigmas.log().lerp(sigmas.roll(1).log(), 0.5).exp()
+
+        self.sigmas = torch.cat([sigmas[:1], sigmas[1:].repeat_interleave(2), sigmas[-1:]])
+        self.sigmas_interpol = torch.cat(
+            [sigmas_interpol[:1], sigmas_interpol[1:].repeat_interleave(2), sigmas_interpol[-1:]]
+        )
+
+        # standard deviation of the initial noise distribution
+        self.init_noise_sigma = self.sigmas.max()
+
+        timesteps = torch.from_numpy(timesteps).to(device)
+
+        # interpolate timesteps
+        timesteps_interpol = self.sigma_to_t(sigmas_interpol).to(device)
+        interleaved_timesteps = torch.stack((timesteps_interpol[1:-1, None], timesteps[1:, None]), dim=-1).flatten()
+        timesteps = torch.cat([timesteps[:1], interleaved_timesteps])
+
+        if str(device).startswith("mps"):
+            # mps does not support float64
+            self.timesteps = timesteps.to(torch.float32)
+        else:
+            self.timesteps = timesteps
+
+        self.sample = None
+
+    def sigma_to_t(self, sigma):
+        # get log sigma
+        log_sigma = sigma.log()
+
+        # get distribution
+        dists = log_sigma - self.log_sigmas[:, None]
+
+        # get sigmas range
+        low_idx = dists.ge(0).cumsum(dim=0).argmax(dim=0).clamp(max=self.log_sigmas.shape[0] - 2)
+        high_idx = low_idx + 1
+
+        low = self.log_sigmas[low_idx]
+        high = self.log_sigmas[high_idx]
+
+        # interpolate sigmas
+        w = (low - log_sigma) / (low - high)
+        w = w.clamp(0, 1)
+
+        # transform interpolation to time range
+        t = (1 - w) * low_idx + w * high_idx
+        t = t.view(sigma.shape)
+        return t
+
+    @property
+    def state_in_first_order(self):
+        return self.sample is None
+
+    def step(
+        self,
+        model_output: Union[torch.FloatTensor, np.ndarray],
+        timestep: Union[float, torch.FloatTensor],
+        sample: Union[torch.FloatTensor, np.ndarray],
+        return_dict: bool = True,
+    ) -> Union[SchedulerOutput, Tuple]:
+        """
+        Args:
+        Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+        process from the learned model outputs (most often the predicted noise).
+            model_output (`torch.FloatTensor` or `np.ndarray`): direct output from learned diffusion model. timestep
+            (`int`): current discrete timestep in the diffusion chain. sample (`torch.FloatTensor` or `np.ndarray`):
+                current instance of sample being created by diffusion process.
+            return_dict (`bool`): option for returning tuple rather than SchedulerOutput class
+        Returns:
+            [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+            [`~schedulers.scheduling_utils.SchedulerOutput`] if `return_dict` is True, otherwise a `tuple`. When
+            returning a tuple, the first element is the sample tensor.
+        """
+        step_index = self.index_for_timestep(timestep)
+
+        if self.state_in_first_order:
+            sigma = self.sigmas[step_index]
+            sigma_interpol = self.sigmas_interpol[step_index + 1]
+            sigma_next = self.sigmas[step_index + 1]
+        else:
+            # 2nd order / KDPM2's method
+            sigma = self.sigmas[step_index - 1]
+            sigma_interpol = self.sigmas_interpol[step_index]
+            sigma_next = self.sigmas[step_index]
+
+        # currently only gamma=0 is supported. This usually works best anyways.
+        # We can support gamma in the future but then need to scale the timestep before
+        # passing it to the model which requires a change in API
+        gamma = 0
+        sigma_hat = sigma * (gamma + 1)  # Note: sigma_hat == sigma for now
+
+        # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
+
+        if self.state_in_first_order:
+            pred_original_sample = sample - sigma_hat * model_output
+            # 2. Convert to an ODE derivative
+            derivative = (sample - pred_original_sample) / sigma_hat
+            # 3. 1st order derivative
+            dt = sigma_interpol - sigma_hat
+
+            # store for 2nd order step
+            self.sample = sample
+        else:
+            # DPM-Solver-2
+            pred_original_sample = sample - sigma_interpol * model_output
+            derivative = (sample - pred_original_sample) / sigma_interpol
+
+            dt = sigma_next - sigma_hat
+
+            sample = self.sample
+            self.sample = None
+
+        prev_sample = sample + derivative * dt
+
+        if not return_dict:
+            return (prev_sample,)
+
+        return SchedulerOutput(prev_sample=prev_sample)
+
+    def add_noise(
+        self,
+        original_samples: torch.FloatTensor,
+        noise: torch.FloatTensor,
+        timesteps: torch.FloatTensor,
+    ) -> torch.FloatTensor:
+        # Make sure sigmas and timesteps have the same device and dtype as original_samples
+        self.sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+        if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+            # mps does not support float64
+            self.timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+            timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+        else:
+            self.timesteps = self.timesteps.to(original_samples.device)
+            timesteps = timesteps.to(original_samples.device)
+
+        step_indices = [self.index_for_timestep(t) for t in timesteps]
+
+        sigma = self.sigmas[step_indices].flatten()
+        while len(sigma.shape) < len(original_samples.shape):
+            sigma = sigma.unsqueeze(-1)
+
+        noisy_samples = original_samples + noise * sigma
+        return noisy_samples
+
+    def __len__(self):
+        return self.config.num_train_timesteps

src/diffusers/utils/dummy_pt_objects.py

@@ -407,6 +407,36 @@ class KarrasVeScheduler(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class KDPM2AncestralDiscreteScheduler(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
+class KDPM2DiscreteScheduler(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class PNDMScheduler(metaclass=DummyObject):
     _backends = ["torch"]
 

tests/test_scheduler.py

@@ -32,6 +32,8 @@ from diffusers import (
     EulerDiscreteScheduler,
     HeunDiscreteScheduler,
     IPNDMScheduler,
+    KDPM2AncestralDiscreteScheduler,
+    KDPM2DiscreteScheduler,
     LMSDiscreteScheduler,
     PNDMScheduler,
     ScoreSdeVeScheduler,
@@ -1994,3 +1996,194 @@ class HeunDiscreteSchedulerTest(SchedulerCommonTest):
             # CUDA
             assert abs(result_sum.item() - 0.1233) < 1e-2
             assert abs(result_mean.item() - 0.0002) < 1e-3
+
+
+class KDPM2DiscreteSchedulerTest(SchedulerCommonTest):
+    scheduler_classes = (KDPM2DiscreteScheduler,)
+    num_inference_steps = 10
+
+    def get_scheduler_config(self, **kwargs):
+        config = {
+            "num_train_timesteps": 1100,
+            "beta_start": 0.0001,
+            "beta_end": 0.02,
+            "beta_schedule": "linear",
+        }
+
+        config.update(**kwargs)
+        return config
+
+    def test_timesteps(self):
+        for timesteps in [10, 50, 100, 1000]:
+            self.check_over_configs(num_train_timesteps=timesteps)
+
+    def test_betas(self):
+        for beta_start, beta_end in zip([0.00001, 0.0001, 0.001], [0.0002, 0.002, 0.02]):
+            self.check_over_configs(beta_start=beta_start, beta_end=beta_end)
+
+    def test_schedules(self):
+        for schedule in ["linear", "scaled_linear"]:
+            self.check_over_configs(beta_schedule=schedule)
+
+    def test_full_loop_no_noise(self):
+        if torch_device == "mps":
+            return
+        scheduler_class = self.scheduler_classes[0]
+        scheduler_config = self.get_scheduler_config()
+        scheduler = scheduler_class(**scheduler_config)
+
+        scheduler.set_timesteps(self.num_inference_steps)
+
+        model = self.dummy_model()
+        sample = self.dummy_sample_deter * scheduler.init_noise_sigma
+        sample = sample.to(torch_device)
+
+        for i, t in enumerate(scheduler.timesteps):
+            sample = scheduler.scale_model_input(sample, t)
+
+            model_output = model(sample, t)
+
+            output = scheduler.step(model_output, t, sample)
+            sample = output.prev_sample
+
+        result_sum = torch.sum(torch.abs(sample))
+        result_mean = torch.mean(torch.abs(sample))
+
+        if torch_device in ["cpu", "mps"]:
+            assert abs(result_sum.item() - 20.4125) < 1e-2
+            assert abs(result_mean.item() - 0.0266) < 1e-3
+        else:
+            # CUDA
+            assert abs(result_sum.item() - 20.4125) < 1e-2
+            assert abs(result_mean.item() - 0.0266) < 1e-3
+
+    def test_full_loop_device(self):
+        if torch_device == "mps":
+            return
+        scheduler_class = self.scheduler_classes[0]
+        scheduler_config = self.get_scheduler_config()
+        scheduler = scheduler_class(**scheduler_config)
+
+        scheduler.set_timesteps(self.num_inference_steps, device=torch_device)
+
+        model = self.dummy_model()
+        sample = self.dummy_sample_deter.to(torch_device) * scheduler.init_noise_sigma
+
+        for t in scheduler.timesteps:
+            sample = scheduler.scale_model_input(sample, t)
+
+            model_output = model(sample, t)
+
+            output = scheduler.step(model_output, t, sample)
+            sample = output.prev_sample
+
+        result_sum = torch.sum(torch.abs(sample))
+        result_mean = torch.mean(torch.abs(sample))
+
+        if str(torch_device).startswith("cpu"):
+            # The following sum varies between 148 and 156 on mps. Why?
+            assert abs(result_sum.item() - 20.4125) < 1e-2
+            assert abs(result_mean.item() - 0.0266) < 1e-3
+        else:
+            # CUDA
+            assert abs(result_sum.item() - 20.4125) < 1e-2
+            assert abs(result_mean.item() - 0.0266) < 1e-3
+
+
+class KDPM2AncestralDiscreteSchedulerTest(SchedulerCommonTest):
+    scheduler_classes = (KDPM2AncestralDiscreteScheduler,)
+    num_inference_steps = 10
+
+    def get_scheduler_config(self, **kwargs):
+        config = {
+            "num_train_timesteps": 1100,
+            "beta_start": 0.0001,
+            "beta_end": 0.02,
+            "beta_schedule": "linear",
+        }
+
+        config.update(**kwargs)
+        return config
+
+    def test_timesteps(self):
+        for timesteps in [10, 50, 100, 1000]:
+            self.check_over_configs(num_train_timesteps=timesteps)
+
+    def test_betas(self):
+        for beta_start, beta_end in zip([0.00001, 0.0001, 0.001], [0.0002, 0.002, 0.02]):
+            self.check_over_configs(beta_start=beta_start, beta_end=beta_end)
+
+    def test_schedules(self):
+        for schedule in ["linear", "scaled_linear"]:
+            self.check_over_configs(beta_schedule=schedule)
+
+    def test_full_loop_no_noise(self):
+        if torch_device == "mps":
+            return
+        scheduler_class = self.scheduler_classes[0]
+        scheduler_config = self.get_scheduler_config()
+        scheduler = scheduler_class(**scheduler_config)
+
+        scheduler.set_timesteps(self.num_inference_steps)
+
+        generator = torch.Generator(device=torch_device).manual_seed(0)
+
+        model = self.dummy_model()
+        sample = self.dummy_sample_deter * scheduler.init_noise_sigma
+        sample = sample.to(torch_device)
+
+        for i, t in enumerate(scheduler.timesteps):
+            sample = scheduler.scale_model_input(sample, t)
+
+            model_output = model(sample, t)
+
+            output = scheduler.step(model_output, t, sample, generator=generator)
+            sample = output.prev_sample
+
+        result_sum = torch.sum(torch.abs(sample))
+        result_mean = torch.mean(torch.abs(sample))
+
+        if torch_device in ["cpu", "mps"]:
+            assert abs(result_sum.item() - 13849.3945) < 1e-2
+            assert abs(result_mean.item() - 18.0331) < 5e-3
+        else:
+            # CUDA
+            assert abs(result_sum.item() - 13913.0449) < 1e-2
+            assert abs(result_mean.item() - 18.1159) < 5e-3
+
+    def test_full_loop_device(self):
+        if torch_device == "mps":
+            return
+        scheduler_class = self.scheduler_classes[0]
+        scheduler_config = self.get_scheduler_config()
+        scheduler = scheduler_class(**scheduler_config)
+
+        scheduler.set_timesteps(self.num_inference_steps, device=torch_device)
+
+        if torch_device == "mps":
+            # device type MPS is not supported for torch.Generator() api.
+            generator = torch.manual_seed(0)
+        else:
+            generator = torch.Generator(device=torch_device).manual_seed(0)
+
+        model = self.dummy_model()
+        sample = self.dummy_sample_deter.to(torch_device) * scheduler.init_noise_sigma
+
+        for t in scheduler.timesteps:
+            sample = scheduler.scale_model_input(sample, t)
+
+            model_output = model(sample, t)
+
+            output = scheduler.step(model_output, t, sample, generator=generator)
+            sample = output.prev_sample
+
+        result_sum = torch.sum(torch.abs(sample))
+        result_mean = torch.mean(torch.abs(sample))
+
+        if str(torch_device).startswith("cpu"):
+            assert abs(result_sum.item() - 13849.3945) < 1e-2
+            assert abs(result_mean.item() - 18.0331) < 5e-3
+        else:
+            # CUDA
+            assert abs(result_sum.item() - 13913.0459) < 1e-2
+            assert abs(result_mean.item() - 18.1159) < 1e-3