scheduling_dpmsolver_multistep.py 59.8 KB
Newer Older
Aryan's avatar
Aryan committed
1
# Copyright 2025 TSAIL Team and The HuggingFace Team. All rights reserved.
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
#
# 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.

# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver

import math
18
from typing import List, Literal, Optional, Tuple, Union
19
20
21
22
23

import numpy as np
import torch

from ..configuration_utils import ConfigMixin, register_to_config
24
from ..utils import deprecate, is_scipy_available
Dhruv Nair's avatar
Dhruv Nair committed
25
from ..utils.torch_utils import randn_tensor
Kashif Rasul's avatar
Kashif Rasul committed
26
from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
27
28


29
30
31
32
if is_scipy_available():
    import scipy.stats


33
# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
YiYi Xu's avatar
YiYi Xu committed
34
def betas_for_alpha_bar(
35
36
37
38
    num_diffusion_timesteps: int,
    max_beta: float = 0.999,
    alpha_transform_type: Literal["cosine", "exp"] = "cosine",
) -> torch.Tensor:
39
40
41
42
43
44
45
46
    """
    Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
    (1-beta) over time from t = [0,1].

    Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
    to that part of the diffusion process.

    Args:
47
48
49
50
51
52
        num_diffusion_timesteps (`int`):
            The number of betas to produce.
        max_beta (`float`, defaults to `0.999`):
            The maximum beta to use; use values lower than 1 to avoid numerical instability.
        alpha_transform_type (`"cosine"` or `"exp"`, defaults to `"cosine"`):
            The type of noise schedule for `alpha_bar`. Choose from `cosine` or `exp`.
53
54

    Returns:
55
56
        `torch.Tensor`:
            The betas used by the scheduler to step the model outputs.
57
    """
YiYi Xu's avatar
YiYi Xu committed
58
    if alpha_transform_type == "cosine":
59

YiYi Xu's avatar
YiYi Xu committed
60
61
62
63
64
65
66
67
68
        def alpha_bar_fn(t):
            return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2

    elif alpha_transform_type == "exp":

        def alpha_bar_fn(t):
            return math.exp(t * -12.0)

    else:
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
69
        raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}")
70
71
72
73
74

    betas = []
    for i in range(num_diffusion_timesteps):
        t1 = i / num_diffusion_timesteps
        t2 = (i + 1) / num_diffusion_timesteps
YiYi Xu's avatar
YiYi Xu committed
75
        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
76
77
78
    return torch.tensor(betas, dtype=torch.float32)


79
80
81
# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
def rescale_zero_terminal_snr(betas):
    """
Quentin Gallouédec's avatar
Quentin Gallouédec committed
82
    Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1)
83
84

    Args:
85
        betas (`torch.Tensor`):
86
            The betas that the scheduler is being initialized with.
87
88

    Returns:
89
90
        `torch.Tensor`:
            Rescaled betas with zero terminal SNR.
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
    """
    # Convert betas to alphas_bar_sqrt
    alphas = 1.0 - betas
    alphas_cumprod = torch.cumprod(alphas, dim=0)
    alphas_bar_sqrt = alphas_cumprod.sqrt()

    # Store old values.
    alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
    alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()

    # Shift so the last timestep is zero.
    alphas_bar_sqrt -= alphas_bar_sqrt_T

    # Scale so the first timestep is back to the old value.
    alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)

    # Convert alphas_bar_sqrt to betas
    alphas_bar = alphas_bar_sqrt**2  # Revert sqrt
    alphas = alphas_bar[1:] / alphas_bar[:-1]  # Revert cumprod
    alphas = torch.cat([alphas_bar[0:1], alphas])
    betas = 1 - alphas

    return betas


116
117
class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
    """
118
    `DPMSolverMultistepScheduler` is a fast dedicated high-order solver for diffusion ODEs.
119

120
121
    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.
122
123

    Args:
124
125
126
127
128
129
        num_train_timesteps (`int`, defaults to 1000):
            The number of diffusion steps to train the model.
        beta_start (`float`, defaults to 0.0001):
            The starting `beta` value of inference.
        beta_end (`float`, defaults to 0.02):
            The final `beta` value.
130
131
        beta_schedule (`"linear"`, `"scaled_linear"`, or `"squaredcos_cap_v2"`, defaults to `"linear"`):
            The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model.
132
133
134
135
        trained_betas (`np.ndarray`, *optional*):
            Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
        solver_order (`int`, defaults to 2):
            The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided
136
            sampling, and `solver_order=3` for unconditional sampling.
137
138
139
140
        prediction_type (`"epsilon"`, `"sample"`, `"v_prediction"`, or `"flow_prediction"`, defaults to `"epsilon"`):
            Prediction type of the scheduler function. `epsilon` predicts the noise of the diffusion process, `sample`
            directly predicts the noisy sample, `v_prediction` predicts the velocity (see section 2.4 of [Imagen
            Video](https://huggingface.co/papers/2210.02303) paper), and `flow_prediction` predicts the flow.
141
142
143
144
145
146
147
148
        thresholding (`bool`, defaults to `False`):
            Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
            as Stable Diffusion.
        dynamic_thresholding_ratio (`float`, defaults to 0.995):
            The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
        sample_max_value (`float`, defaults to 1.0):
            The threshold value for dynamic thresholding. Valid only when `thresholding=True` and
            `algorithm_type="dpmsolver++"`.
149
150
151
152
153
154
155
156
        algorithm_type (`"dpmsolver"`, `"dpmsolver++"`, `"sde-dpmsolver"`, or `"sde-dpmsolver++"`, defaults to `"dpmsolver++"`):
            Algorithm type for the solver. The `dpmsolver` type implements the algorithms in the
            [DPMSolver](https://huggingface.co/papers/2206.00927) paper, and the `dpmsolver++` type implements the
            algorithms in the [DPMSolver++](https://huggingface.co/papers/2211.01095) paper. It is recommended to use
            `dpmsolver++` or `sde-dpmsolver++` with `solver_order=2` for guided sampling like in Stable Diffusion.
        solver_type (`"midpoint"` or `"heun"`, defaults to `"midpoint"`):
            Solver type for the second-order solver. The solver type slightly affects the sample quality, especially
            for a small number of steps. It is recommended to use `midpoint` solvers.
157
158
159
        lower_order_final (`bool`, defaults to `True`):
            Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can
            stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10.
160
161
162
163
        euler_at_final (`bool`, defaults to `False`):
            Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail
            richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference
            steps, but sometimes may result in blurring.
164
        use_karras_sigmas (`bool`, *optional*, defaults to `False`):
165
166
            Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
            the sigmas are determined according to a sequence of noise levels {σi}.
167
168
        use_exponential_sigmas (`bool`, *optional*, defaults to `False`):
            Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process.
169
170
171
        use_beta_sigmas (`bool`, *optional*, defaults to `False`):
            Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta
            Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information.
172
173
174
175
        use_lu_lambdas (`bool`, *optional*, defaults to `False`):
            Whether to use the uniform-logSNR for step sizes proposed by Lu's DPM-Solver in the noise schedule during
            the sampling process. If `True`, the sigmas and time steps are determined according to a sequence of
            `lambda(t)`.
176
177
178
179
        use_flow_sigmas (`bool`, *optional*, defaults to `False`):
            Whether to use flow sigmas for step sizes in the noise schedule during the sampling process.
        flow_shift (`float`, *optional*, defaults to 1.0):
            The shift value for the timestep schedule for flow matching.
180
        final_sigmas_type (`"zero"` or `"sigma_min"`, *optional*, defaults to `"zero"`):
181
            The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final
182
            sigma is the same as the last sigma in the training schedule. If `"zero"`, the final sigma is set to 0.
183
184
185
        lambda_min_clipped (`float`, defaults to `-inf`):
            Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the
            cosine (`squaredcos_cap_v2`) noise schedule.
186
187
188
189
        variance_type (`"learned"` or `"learned_range"`, *optional*):
            Set to `"learned"` or `"learned_range"` for diffusion models that predict variance. If set, the model's
            output contains the predicted Gaussian variance.
        timestep_spacing (`"linspace"`, `"leading"`, or `"trailing"`, defaults to `"linspace"`):
190
191
192
            The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
            Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
        steps_offset (`int`, defaults to 0):
193
            An offset added to the inference steps, as required by some model families.
194
195
196
197
        rescale_betas_zero_snr (`bool`, defaults to `False`):
            Whether to rescale the betas to have zero terminal SNR. This enables 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).
198
199
200
201
        use_dynamic_shifting (`bool`, defaults to `False`):
            Whether to use dynamic shifting for the timestep schedule.
        time_shift_type (`"exponential"`, defaults to `"exponential"`):
            The type of time shift to apply when using dynamic shifting.
202
203
    """

Kashif Rasul's avatar
Kashif Rasul committed
204
    _compatibles = [e.name for e in KarrasDiffusionSchedulers]
205
    order = 1
206
207
208
209
210
211
212

    @register_to_config
    def __init__(
        self,
        num_train_timesteps: int = 1000,
        beta_start: float = 0.0001,
        beta_end: float = 0.02,
213
        beta_schedule: Literal["linear", "scaled_linear", "squaredcos_cap_v2"] = "linear",
214
        trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
215
        solver_order: int = 2,
216
        prediction_type: Literal["epsilon", "sample", "v_prediction", "flow_prediction"] = "epsilon",
217
218
219
        thresholding: bool = False,
        dynamic_thresholding_ratio: float = 0.995,
        sample_max_value: float = 1.0,
220
221
        algorithm_type: Literal["dpmsolver", "dpmsolver++", "sde-dpmsolver", "sde-dpmsolver++"] = "dpmsolver++",
        solver_type: Literal["midpoint", "heun"] = "midpoint",
222
        lower_order_final: bool = True,
223
        euler_at_final: bool = False,
224
        use_karras_sigmas: Optional[bool] = False,
225
        use_exponential_sigmas: Optional[bool] = False,
226
        use_beta_sigmas: Optional[bool] = False,
227
        use_lu_lambdas: Optional[bool] = False,
228
229
        use_flow_sigmas: Optional[bool] = False,
        flow_shift: Optional[float] = 1.0,
230
        final_sigmas_type: Optional[Literal["zero", "sigma_min"]] = "zero",
231
        lambda_min_clipped: float = -float("inf"),
232
233
        variance_type: Optional[Literal["learned", "learned_range"]] = None,
        timestep_spacing: Literal["linspace", "leading", "trailing"] = "linspace",
234
        steps_offset: int = 0,
235
        rescale_betas_zero_snr: bool = False,
236
        use_dynamic_shifting: bool = False,
237
        time_shift_type: Literal["exponential"] = "exponential",
238
    ):
239
240
241
242
243
244
        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.")
        if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1:
            raise ValueError(
                "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used."
            )
245
246
247
248
        if algorithm_type in ["dpmsolver", "sde-dpmsolver"]:
            deprecation_message = f"algorithm_type {algorithm_type} is deprecated and will be removed in a future version. Choose from `dpmsolver++` or `sde-dpmsolver++` instead"
            deprecate("algorithm_types dpmsolver and sde-dpmsolver", "1.0.0", deprecation_message)

249
        if trained_betas is not None:
250
            self.betas = torch.tensor(trained_betas, dtype=torch.float32)
251
252
253
254
        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.
255
            self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
256
257
258
259
        elif beta_schedule == "squaredcos_cap_v2":
            # Glide cosine schedule
            self.betas = betas_for_alpha_bar(num_train_timesteps)
        else:
260
            raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}")
261

262
263
264
        if rescale_betas_zero_snr:
            self.betas = rescale_zero_terminal_snr(self.betas)

265
266
        self.alphas = 1.0 - self.betas
        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
267
268
269
270
271
272

        if rescale_betas_zero_snr:
            # Close to 0 without being 0 so first sigma is not inf
            # FP16 smallest positive subnormal works well here
            self.alphas_cumprod[-1] = 2**-24

273
274
275
276
        # Currently we only support VP-type noise schedule
        self.alpha_t = torch.sqrt(self.alphas_cumprod)
        self.sigma_t = torch.sqrt(1 - self.alphas_cumprod)
        self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t)
277
        self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5
278
279
280
281
282

        # standard deviation of the initial noise distribution
        self.init_noise_sigma = 1.0

        # settings for DPM-Solver
283
        if algorithm_type not in ["dpmsolver", "dpmsolver++", "sde-dpmsolver", "sde-dpmsolver++"]:
284
            if algorithm_type == "deis":
285
                self.register_to_config(algorithm_type="dpmsolver++")
286
            else:
287
                raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}")
288

289
        if solver_type not in ["midpoint", "heun"]:
290
            if solver_type in ["logrho", "bh1", "bh2"]:
291
                self.register_to_config(solver_type="midpoint")
292
            else:
293
                raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}")
294

295
296
297
298
299
        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 choose `sigma_min` instead."
            )

300
301
302
303
304
305
        # setable values
        self.num_inference_steps = None
        timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy()
        self.timesteps = torch.from_numpy(timesteps)
        self.model_outputs = [None] * solver_order
        self.lower_order_nums = 0
306
        self._step_index = None
307
        self._begin_index = None
308
        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
309
310
311
312

    @property
    def step_index(self):
        """
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
313
        The index counter for current timestep. It will increase 1 after each scheduler step.
314
315
        """
        return self._step_index
316

317
318
319
320
321
322
323
324
325
326
327
328
    @property
    def begin_index(self):
        """
        The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
        """
        return self._begin_index

    def set_begin_index(self, begin_index: int = 0):
        """
        Sets the begin index for the scheduler. This function should be run from pipeline before the inference.

        Args:
329
            begin_index (`int`, defaults to `0`):
330
331
332
333
                The begin index for the scheduler.
        """
        self._begin_index = begin_index

334
335
    def set_timesteps(
        self,
336
337
        num_inference_steps: Optional[int] = None,
        device: Optional[Union[str, torch.device]] = None,
338
        mu: Optional[float] = None,
339
        timesteps: Optional[List[int]] = None,
340
    ) -> None:
341
        """
342
        Sets the discrete timesteps used for the diffusion chain (to be run before inference).
343
344

        Args:
345
            num_inference_steps (`int`, *optional*):
346
347
348
                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.
349
350
351
            mu (`float`, *optional*):
                The mu parameter for dynamic shifting. If provided, requires `use_dynamic_shifting=True` and
                `time_shift_type="exponential"`.
352
353
354
355
            timesteps (`List[int]`, *optional*):
                Custom timesteps used to support arbitrary timesteps schedule. If `None`, timesteps will be generated
                based on the `timestep_spacing` attribute. If `timesteps` is passed, `num_inference_steps` and `sigmas`
                must be `None`, and `timestep_spacing` attribute will be ignored.
356
        """
357
358
359
        if mu is not None:
            assert self.config.use_dynamic_shifting and self.config.time_shift_type == "exponential"
            self.config.flow_shift = np.exp(mu)
360
361
362
363
364
365
366
367
        if num_inference_steps is None and timesteps is None:
            raise ValueError("Must pass exactly one of `num_inference_steps` or `timesteps`.")
        if num_inference_steps is not None and timesteps is not None:
            raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.")
        if timesteps is not None and self.config.use_karras_sigmas:
            raise ValueError("Cannot use `timesteps` with `config.use_karras_sigmas = True`")
        if timesteps is not None and self.config.use_lu_lambdas:
            raise ValueError("Cannot use `timesteps` with `config.use_lu_lambdas = True`")
368
369
        if timesteps is not None and self.config.use_exponential_sigmas:
            raise ValueError("Cannot set `timesteps` with `config.use_exponential_sigmas = True`.")
370
371
        if timesteps is not None and self.config.use_beta_sigmas:
            raise ValueError("Cannot set `timesteps` with `config.use_beta_sigmas = True`.")
372
373
374

        if timesteps is not None:
            timesteps = np.array(timesteps).astype(np.int64)
375
        else:
376
377
378
379
380
            # Clipping the minimum of all lambda(t) for numerical stability.
            # This is critical for cosine (squaredcos_cap_v2) noise schedule.
            clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.config.lambda_min_clipped)
            last_timestep = ((self.config.num_train_timesteps - clipped_idx).numpy()).item()

Quentin Gallouédec's avatar
Quentin Gallouédec committed
381
            # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
            if self.config.timestep_spacing == "linspace":
                timesteps = (
                    np.linspace(0, last_timestep - 1, num_inference_steps + 1)
                    .round()[::-1][:-1]
                    .copy()
                    .astype(np.int64)
                )
            elif self.config.timestep_spacing == "leading":
                step_ratio = last_timestep // (num_inference_steps + 1)
                # creates integer timesteps by multiplying by ratio
                # casting to int to avoid issues when num_inference_step is power of 3
                timesteps = (
                    (np.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(np.int64)
                )
                timesteps += self.config.steps_offset
            elif self.config.timestep_spacing == "trailing":
                step_ratio = self.config.num_train_timesteps / num_inference_steps
                # creates integer timesteps by multiplying by ratio
                # casting to int to avoid issues when num_inference_step is power of 3
                timesteps = np.arange(last_timestep, 0, -step_ratio).round().copy().astype(np.int64)
                timesteps -= 1
            else:
                raise ValueError(
                    f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
                )
407

408
        sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
409
410
        log_sigmas = np.log(sigmas)

411
        if self.config.use_karras_sigmas:
412
            sigmas = np.flip(sigmas).copy()
413
            sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
414
415
416
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
            if self.config.beta_schedule != "squaredcos_cap_v2":
                timesteps = timesteps.round()
417
418
419
420
        elif self.config.use_lu_lambdas:
            lambdas = np.flip(log_sigmas.copy())
            lambdas = self._convert_to_lu(in_lambdas=lambdas, num_inference_steps=num_inference_steps)
            sigmas = np.exp(lambdas)
421
422
423
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
            if self.config.beta_schedule != "squaredcos_cap_v2":
                timesteps = timesteps.round()
424
        elif self.config.use_exponential_sigmas:
425
426
            sigmas = np.flip(sigmas).copy()
            sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
427
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
428
        elif self.config.use_beta_sigmas:
429
430
            sigmas = np.flip(sigmas).copy()
            sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
431
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
432
433
434
        elif self.config.use_flow_sigmas:
            alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1)
            sigmas = 1.0 - alphas
hlky's avatar
hlky committed
435
            sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1].copy()
436
            timesteps = (sigmas * self.config.num_train_timesteps).copy()
437
438
        else:
            sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
439
440

        if self.config.final_sigmas_type == "sigma_min":
441
            sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5
442
443
444
445
446
447
448
449
        elif self.config.final_sigmas_type == "zero":
            sigma_last = 0
        else:
            raise ValueError(
                f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}"
            )

        sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32)
450

451
452
        self.sigmas = torch.from_numpy(sigmas)
        self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64)
453
454
455

        self.num_inference_steps = len(timesteps)

456
457
458
459
460
        self.model_outputs = [
            None,
        ] * self.config.solver_order
        self.lower_order_nums = 0

461
462
        # add an index counter for schedulers that allow duplicated timesteps
        self._step_index = None
463
        self._begin_index = None
464
        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
465

466
    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
467
    def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor:
468
        """
469
470
        Apply dynamic thresholding to the predicted sample.

471
472
473
474
475
476
        "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
        prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
        s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
        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."

Quentin Gallouédec's avatar
Quentin Gallouédec committed
477
        https://huggingface.co/papers/2205.11487
478
479
480
481
482
483
484
485

        Args:
            sample (`torch.Tensor`):
                The predicted sample to be thresholded.

        Returns:
            `torch.Tensor`:
                The thresholded sample.
486
487
        """
        dtype = sample.dtype
488
        batch_size, channels, *remaining_dims = sample.shape
489
490
491
492
493

        if dtype not in (torch.float32, torch.float64):
            sample = sample.float()  # upcast for quantile calculation, and clamp not implemented for cpu half

        # Flatten sample for doing quantile calculation along each image
494
        sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
495
496
497
498
499
500
501
502
503
504

        abs_sample = sample.abs()  # "a certain percentile absolute pixel value"

        s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
        s = torch.clamp(
            s, min=1, max=self.config.sample_max_value
        )  # When clamped to min=1, equivalent to standard clipping to [-1, 1]
        s = s.unsqueeze(1)  # (batch_size, 1) because clamp will broadcast along dim=0
        sample = torch.clamp(sample, -s, s) / s  # "we threshold xt0 to the range [-s, s] and then divide by s"

505
        sample = sample.reshape(batch_size, channels, *remaining_dims)
506
507
508
        sample = sample.to(dtype)

        return sample
509

510
    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
511
    def _sigma_to_t(self, sigma: np.ndarray, log_sigmas: np.ndarray) -> np.ndarray:
512
513
514
515
516
517
518
519
520
521
522
523
524
        """
        Convert sigma values to corresponding timestep values through interpolation.

        Args:
            sigma (`np.ndarray`):
                The sigma value(s) to convert to timestep(s).
            log_sigmas (`np.ndarray`):
                The logarithm of the sigma schedule used for interpolation.

        Returns:
            `np.ndarray`:
                The interpolated timestep value(s) corresponding to the input sigma(s).
        """
525
        # get log sigma
526
        log_sigma = np.log(np.maximum(sigma, 1e-10))
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546

        # get distribution
        dists = log_sigma - log_sigmas[:, np.newaxis]

        # get sigmas range
        low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
        high_idx = low_idx + 1

        low = log_sigmas[low_idx]
        high = log_sigmas[high_idx]

        # interpolate sigmas
        w = (low - log_sigma) / (low - high)
        w = np.clip(w, 0, 1)

        # transform interpolation to time range
        t = (1 - w) * low_idx + w * high_idx
        t = t.reshape(sigma.shape)
        return t

547
548
549
550
551
552
553
554
555
556
557
558
    def _sigma_to_alpha_sigma_t(self, sigma: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Convert sigma values to alpha_t and sigma_t values.

        Args:
            sigma (`torch.Tensor`):
                The sigma value(s) to convert.

        Returns:
            `Tuple[torch.Tensor, torch.Tensor]`:
                A tuple containing (alpha_t, sigma_t) values.
        """
559
560
561
562
563
564
        if self.config.use_flow_sigmas:
            alpha_t = 1 - sigma
            sigma_t = sigma
        else:
            alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
            sigma_t = sigma * alpha_t
565
566
567

        return alpha_t, sigma_t

568
    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
569
    def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor:
570
571
572
573
574
575
576
577
578
579
580
581
582
583
        """
        Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative
        Models](https://huggingface.co/papers/2206.00364).

        Args:
            in_sigmas (`torch.Tensor`):
                The input sigma values to be converted.
            num_inference_steps (`int`):
                The number of inference steps to generate the noise schedule for.

        Returns:
            `torch.Tensor`:
                The converted sigma values following the Karras noise schedule.
        """
584

Suraj Patil's avatar
Suraj Patil committed
585
586
587
588
589
590
591
592
593
594
595
596
597
598
        # Hack to make sure that other schedulers which copy this function don't break
        # TODO: Add this logic to the other schedulers
        if hasattr(self.config, "sigma_min"):
            sigma_min = self.config.sigma_min
        else:
            sigma_min = None

        if hasattr(self.config, "sigma_max"):
            sigma_max = self.config.sigma_max
        else:
            sigma_max = None

        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
599
600
601
602
603
604
605
606

        rho = 7.0  # 7.0 is the value used in the paper
        ramp = np.linspace(0, 1, num_inference_steps)
        min_inv_rho = sigma_min ** (1 / rho)
        max_inv_rho = sigma_max ** (1 / rho)
        sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
        return sigmas

607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
    def _convert_to_lu(self, in_lambdas: torch.Tensor, num_inference_steps: int) -> torch.Tensor:
        """
        Construct the noise schedule as proposed in [DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model
        Sampling in Around 10 Steps](https://huggingface.co/papers/2206.00927) by Lu et al. (2022).

        Args:
            in_lambdas (`torch.Tensor`):
                The input lambda values to be converted.
            num_inference_steps (`int`):
                The number of inference steps to generate the noise schedule for.

        Returns:
            `torch.Tensor`:
                The converted lambda values following the Lu noise schedule.
        """
622
623
624
625
626
627
628
629
630
631
632

        lambda_min: float = in_lambdas[-1].item()
        lambda_max: float = in_lambdas[0].item()

        rho = 1.0  # 1.0 is the value used in the paper
        ramp = np.linspace(0, 1, num_inference_steps)
        min_inv_rho = lambda_min ** (1 / rho)
        max_inv_rho = lambda_max ** (1 / rho)
        lambdas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
        return lambdas

633
634
    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential
    def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor:
635
636
637
638
639
640
641
642
643
644
645
646
647
        """
        Construct an exponential noise schedule.

        Args:
            in_sigmas (`torch.Tensor`):
                The input sigma values to be converted.
            num_inference_steps (`int`):
                The number of inference steps to generate the noise schedule for.

        Returns:
            `torch.Tensor`:
                The converted sigma values following an exponential schedule.
        """
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663

        # Hack to make sure that other schedulers which copy this function don't break
        # TODO: Add this logic to the other schedulers
        if hasattr(self.config, "sigma_min"):
            sigma_min = self.config.sigma_min
        else:
            sigma_min = None

        if hasattr(self.config, "sigma_max"):
            sigma_max = self.config.sigma_max
        else:
            sigma_max = None

        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()

664
        sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps))
665
666
        return sigmas

667
668
669
670
    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta
    def _convert_to_beta(
        self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6
    ) -> torch.Tensor:
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
        """
        Construct a beta noise schedule as proposed in [Beta Sampling is All You
        Need](https://huggingface.co/papers/2407.12173).

        Args:
            in_sigmas (`torch.Tensor`):
                The input sigma values to be converted.
            num_inference_steps (`int`):
                The number of inference steps to generate the noise schedule for.
            alpha (`float`, *optional*, defaults to `0.6`):
                The alpha parameter for the beta distribution.
            beta (`float`, *optional*, defaults to `0.6`):
                The beta parameter for the beta distribution.

        Returns:
            `torch.Tensor`:
                The converted sigma values following a beta distribution schedule.
        """
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704

        # Hack to make sure that other schedulers which copy this function don't break
        # TODO: Add this logic to the other schedulers
        if hasattr(self.config, "sigma_min"):
            sigma_min = self.config.sigma_min
        else:
            sigma_min = None

        if hasattr(self.config, "sigma_max"):
            sigma_max = self.config.sigma_max
        else:
            sigma_max = None

        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()

705
        sigmas = np.array(
706
707
708
709
710
711
712
713
714
715
            [
                sigma_min + (ppf * (sigma_max - sigma_min))
                for ppf in [
                    scipy.stats.beta.ppf(timestep, alpha, beta)
                    for timestep in 1 - np.linspace(0, 1, num_inference_steps)
                ]
            ]
        )
        return sigmas

716
    def convert_model_output(
717
        self,
718
        model_output: torch.Tensor,
719
        *args,
720
        sample: torch.Tensor = None,
721
        **kwargs,
722
    ) -> torch.Tensor:
723
        """
724
725
726
        Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is
        designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an
        integral of the data prediction model.
727

Steven Liu's avatar
Steven Liu committed
728
729
        > [!TIP] > The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both
        noise > prediction and data prediction models.
730
731

        Args:
732
            model_output (`torch.Tensor`):
733
                The direct output from the learned diffusion model.
734
            sample (`torch.Tensor`):
735
                A current instance of a sample created by the diffusion process.
736
737

        Returns:
738
            `torch.Tensor`:
739
                The converted model output.
740
        """
741
742
743
744
745
        timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
        if sample is None:
            if len(args) > 1:
                sample = args[1]
            else:
746
                raise ValueError("missing `sample` as a required keyword argument")
747
748
749
750
751
752
        if timestep is not None:
            deprecate(
                "timesteps",
                "1.0.0",
                "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )
753

754
        # DPM-Solver++ needs to solve an integral of the data prediction model.
755
        if self.config.algorithm_type in ["dpmsolver++", "sde-dpmsolver++"]:
756
            if self.config.prediction_type == "epsilon":
757
                # DPM-Solver and DPM-Solver++ only need the "mean" output.
758
                if self.config.variance_type in ["learned", "learned_range"]:
759
                    model_output = model_output[:, :3]
760
761
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
762
                x0_pred = (sample - sigma_t * model_output) / alpha_t
763
            elif self.config.prediction_type == "sample":
764
                x0_pred = model_output
765
            elif self.config.prediction_type == "v_prediction":
766
767
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
768
                x0_pred = alpha_t * sample - sigma_t * model_output
769
770
771
            elif self.config.prediction_type == "flow_prediction":
                sigma_t = self.sigmas[self.step_index]
                x0_pred = sample - sigma_t * model_output
772
773
            else:
                raise ValueError(
774
775
                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, "
                    "`v_prediction`, or `flow_prediction` for the DPMSolverMultistepScheduler."
776
777
                )

778
            if self.config.thresholding:
779
780
                x0_pred = self._threshold_sample(x0_pred)

781
            return x0_pred
782

783
        # DPM-Solver needs to solve an integral of the noise prediction model.
784
        elif self.config.algorithm_type in ["dpmsolver", "sde-dpmsolver"]:
785
            if self.config.prediction_type == "epsilon":
786
                # DPM-Solver and DPM-Solver++ only need the "mean" output.
787
788
789
790
                if self.config.variance_type in ["learned", "learned_range"]:
                    epsilon = model_output[:, :3]
                else:
                    epsilon = model_output
791
            elif self.config.prediction_type == "sample":
792
793
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
794
                epsilon = (sample - alpha_t * model_output) / sigma_t
795
            elif self.config.prediction_type == "v_prediction":
796
797
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
798
                epsilon = alpha_t * model_output + sigma_t * sample
799
800
            else:
                raise ValueError(
801
802
                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
                    " `v_prediction` for the DPMSolverMultistepScheduler."
803
                )
804

805
            if self.config.thresholding:
806
807
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
808
809
810
811
812
813
                x0_pred = (sample - sigma_t * epsilon) / alpha_t
                x0_pred = self._threshold_sample(x0_pred)
                epsilon = (sample - alpha_t * x0_pred) / sigma_t

            return epsilon

814
815
    def dpm_solver_first_order_update(
        self,
816
        model_output: torch.Tensor,
817
        *args,
818
819
        sample: torch.Tensor = None,
        noise: Optional[torch.Tensor] = None,
820
        **kwargs,
821
    ) -> torch.Tensor:
822
        """
823
        One step for the first-order DPMSolver (equivalent to DDIM).
824
825

        Args:
826
            model_output (`torch.Tensor`):
827
                The direct output from the learned diffusion model.
828
            sample (`torch.Tensor`):
829
                A current instance of a sample created by the diffusion process.
830
831

        Returns:
832
            `torch.Tensor`:
833
                The sample tensor at the previous timestep.
834
        """
835
836
837
838
839
840
        timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
        if sample is None:
            if len(args) > 2:
                sample = args[2]
            else:
841
                raise ValueError("missing `sample` as a required keyword argument")
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
        if timestep is not None:
            deprecate(
                "timesteps",
                "1.0.0",
                "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        if prev_timestep is not None:
            deprecate(
                "prev_timestep",
                "1.0.0",
                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index]
        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
        alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s)
        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
        lambda_s = torch.log(alpha_s) - torch.log(sigma_s)

862
863
864
865
866
        h = lambda_t - lambda_s
        if self.config.algorithm_type == "dpmsolver++":
            x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output
        elif self.config.algorithm_type == "dpmsolver":
            x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output
867
868
869
870
871
872
873
874
875
876
877
878
879
880
        elif self.config.algorithm_type == "sde-dpmsolver++":
            assert noise is not None
            x_t = (
                (sigma_t / sigma_s * torch.exp(-h)) * sample
                + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output
                + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
            )
        elif self.config.algorithm_type == "sde-dpmsolver":
            assert noise is not None
            x_t = (
                (alpha_t / alpha_s) * sample
                - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * model_output
                + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
            )
881
882
883
884
        return x_t

    def multistep_dpm_solver_second_order_update(
        self,
885
        model_output_list: List[torch.Tensor],
886
        *args,
887
888
        sample: torch.Tensor = None,
        noise: Optional[torch.Tensor] = None,
889
        **kwargs,
890
    ) -> torch.Tensor:
891
        """
892
        One step for the second-order multistep DPMSolver.
893
894

        Args:
895
            model_output_list (`List[torch.Tensor]`):
896
                The direct outputs from learned diffusion model at current and latter timesteps.
897
            sample (`torch.Tensor`):
898
                A current instance of a sample created by the diffusion process.
899
900

        Returns:
901
            `torch.Tensor`:
902
                The sample tensor at the previous timestep.
903
        """
904
905
906
907
908
909
        timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
        if sample is None:
            if len(args) > 2:
                sample = args[2]
            else:
910
                raise ValueError("missing `sample` as a required keyword argument")
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
        if timestep_list is not None:
            deprecate(
                "timestep_list",
                "1.0.0",
                "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        if prev_timestep is not None:
            deprecate(
                "prev_timestep",
                "1.0.0",
                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        sigma_t, sigma_s0, sigma_s1 = (
            self.sigmas[self.step_index + 1],
            self.sigmas[self.step_index],
            self.sigmas[self.step_index - 1],
        )

        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
        alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
        alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)

        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
        lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
        lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)

939
        m0, m1 = model_output_list[-1], model_output_list[-2]
940

941
942
943
944
        h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1
        r0 = h_0 / h
        D0, D1 = m0, (1.0 / r0) * (m0 - m1)
        if self.config.algorithm_type == "dpmsolver++":
Quentin Gallouédec's avatar
Quentin Gallouédec committed
945
            # See https://huggingface.co/papers/2211.01095 for detailed derivations
946
947
948
949
950
951
952
953
954
955
956
957
958
            if self.config.solver_type == "midpoint":
                x_t = (
                    (sigma_t / sigma_s0) * sample
                    - (alpha_t * (torch.exp(-h) - 1.0)) * D0
                    - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (sigma_t / sigma_s0) * sample
                    - (alpha_t * (torch.exp(-h) - 1.0)) * D0
                    + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
                )
        elif self.config.algorithm_type == "dpmsolver":
Quentin Gallouédec's avatar
Quentin Gallouédec committed
959
            # See https://huggingface.co/papers/2206.00927 for detailed derivations
960
961
962
963
964
965
966
967
968
969
970
971
            if self.config.solver_type == "midpoint":
                x_t = (
                    (alpha_t / alpha_s0) * sample
                    - (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - 0.5 * (sigma_t * (torch.exp(h) - 1.0)) * D1
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (alpha_t / alpha_s0) * sample
                    - (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
                )
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
        elif self.config.algorithm_type == "sde-dpmsolver++":
            assert noise is not None
            if self.config.solver_type == "midpoint":
                x_t = (
                    (sigma_t / sigma_s0 * torch.exp(-h)) * sample
                    + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
                    + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1
                    + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (sigma_t / sigma_s0 * torch.exp(-h)) * sample
                    + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
                    + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1
                    + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
                )
        elif self.config.algorithm_type == "sde-dpmsolver":
            assert noise is not None
            if self.config.solver_type == "midpoint":
                x_t = (
                    (alpha_t / alpha_s0) * sample
                    - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - (sigma_t * (torch.exp(h) - 1.0)) * D1
                    + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (alpha_t / alpha_s0) * sample
                    - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - 2.0 * (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
                    + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
                )
1004
1005
1006
1007
        return x_t

    def multistep_dpm_solver_third_order_update(
        self,
1008
        model_output_list: List[torch.Tensor],
1009
        *args,
1010
        sample: torch.Tensor = None,
StAlKeR7779's avatar
StAlKeR7779 committed
1011
        noise: Optional[torch.Tensor] = None,
1012
        **kwargs,
1013
    ) -> torch.Tensor:
1014
        """
1015
        One step for the third-order multistep DPMSolver.
1016
1017

        Args:
1018
            model_output_list (`List[torch.Tensor]`):
1019
                The direct outputs from learned diffusion model at current and latter timesteps.
1020
            sample (`torch.Tensor`):
1021
                A current instance of a sample created by diffusion process.
1022
1023

        Returns:
1024
            `torch.Tensor`:
1025
                The sample tensor at the previous timestep.
1026
        """
1027
1028
1029
1030
1031
1032
1033

        timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
        if sample is None:
            if len(args) > 2:
                sample = args[2]
            else:
1034
                raise ValueError("missing `sample` as a required keyword argument")
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
        if timestep_list is not None:
            deprecate(
                "timestep_list",
                "1.0.0",
                "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        if prev_timestep is not None:
            deprecate(
                "prev_timestep",
                "1.0.0",
                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        sigma_t, sigma_s0, sigma_s1, sigma_s2 = (
            self.sigmas[self.step_index + 1],
            self.sigmas[self.step_index],
            self.sigmas[self.step_index - 1],
            self.sigmas[self.step_index - 2],
1054
        )
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067

        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
        alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
        alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)
        alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2)

        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
        lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
        lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)
        lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2)

        m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3]

1068
1069
1070
1071
1072
1073
1074
        h, h_0, h_1 = lambda_t - lambda_s0, lambda_s0 - lambda_s1, lambda_s1 - lambda_s2
        r0, r1 = h_0 / h, h_1 / h
        D0 = m0
        D1_0, D1_1 = (1.0 / r0) * (m0 - m1), (1.0 / r1) * (m1 - m2)
        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++":
Quentin Gallouédec's avatar
Quentin Gallouédec committed
1075
            # See https://huggingface.co/papers/2206.00927 for detailed derivations
1076
1077
1078
1079
1080
1081
1082
            x_t = (
                (sigma_t / sigma_s0) * sample
                - (alpha_t * (torch.exp(-h) - 1.0)) * D0
                + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
                - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2
            )
        elif self.config.algorithm_type == "dpmsolver":
Quentin Gallouédec's avatar
Quentin Gallouédec committed
1083
            # See https://huggingface.co/papers/2206.00927 for detailed derivations
1084
1085
1086
1087
1088
1089
            x_t = (
                (alpha_t / alpha_s0) * sample
                - (sigma_t * (torch.exp(h) - 1.0)) * D0
                - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
                - (sigma_t * ((torch.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2
            )
StAlKeR7779's avatar
StAlKeR7779 committed
1090
1091
1092
1093
1094
1095
1096
1097
1098
        elif self.config.algorithm_type == "sde-dpmsolver++":
            assert noise is not None
            x_t = (
                (sigma_t / sigma_s0 * torch.exp(-h)) * sample
                + (alpha_t * (1.0 - torch.exp(-2.0 * h))) * D0
                + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1
                + (alpha_t * ((1.0 - torch.exp(-2.0 * h) - 2.0 * h) / (2.0 * h) ** 2 - 0.5)) * D2
                + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
            )
1099
1100
        return x_t

1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
    def index_for_timestep(
        self, timestep: Union[int, torch.Tensor], schedule_timesteps: Optional[torch.Tensor] = None
    ) -> int:
        """
        Find the index for a given timestep in the schedule.

        Args:
            timestep (`int` or `torch.Tensor`):
                The timestep for which to find the index.
            schedule_timesteps (`torch.Tensor`, *optional*):
                The timestep schedule to search in. If `None`, uses `self.timesteps`.

        Returns:
            `int`:
                The index of the timestep in the schedule.
        """
1117
1118
        if schedule_timesteps is None:
            schedule_timesteps = self.timesteps
1119

1120
        index_candidates = (schedule_timesteps == timestep).nonzero()
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132

        if len(index_candidates) == 0:
            step_index = len(self.timesteps) - 1
        # The sigma index that is taken for the **very** first `step`
        # is always the second index (or the last index if there is only 1)
        # This way we can ensure we don't accidentally skip a sigma in
        # case we start in the middle of the denoising schedule (e.g. for image-to-image)
        elif len(index_candidates) > 1:
            step_index = index_candidates[1].item()
        else:
            step_index = index_candidates[0].item()

1133
1134
        return step_index

1135
    def _init_step_index(self, timestep: Union[int, torch.Tensor]) -> None:
1136
1137
        """
        Initialize the step_index counter for the scheduler.
1138
1139
1140
1141

        Args:
            timestep (`int` or `torch.Tensor`):
                The current timestep for which to initialize the step index.
1142
1143
1144
1145
1146
1147
1148
1149
        """

        if self.begin_index is None:
            if isinstance(timestep, torch.Tensor):
                timestep = timestep.to(self.timesteps.device)
            self._step_index = self.index_for_timestep(timestep)
        else:
            self._step_index = self._begin_index
1150

1151
1152
    def step(
        self,
1153
        model_output: torch.Tensor,
1154
        timestep: Union[int, torch.Tensor],
1155
        sample: torch.Tensor,
1156
        generator: Optional[torch.Generator] = None,
1157
        variance_noise: Optional[torch.Tensor] = None,
1158
1159
1160
        return_dict: bool = True,
    ) -> Union[SchedulerOutput, Tuple]:
        """
1161
1162
        Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
        the multistep DPMSolver.
1163
1164

        Args:
1165
            model_output (`torch.Tensor`):
1166
1167
                The direct output from the learned diffusion model.
            timestep (`int` or `torch.Tensor`):
1168
                The current discrete timestep in the diffusion chain.
1169
            sample (`torch.Tensor`):
1170
1171
1172
                A current instance of a sample created by the diffusion process.
            generator (`torch.Generator`, *optional*):
                A random number generator.
1173
            variance_noise (`torch.Tensor`, *optional*):
1174
1175
                Alternative to generating noise with `generator` by directly providing the noise for the variance
                itself. Useful for methods such as [`LEdits++`].
1176
            return_dict (`bool`, defaults to `True`):
1177
                Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
1178
1179

        Returns:
1180
            [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
1181
                If `return_dict` is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
1182
                tuple is returned where the first element is the sample tensor.
1183
1184
1185
1186
1187
1188
1189

        """
        if self.num_inference_steps is None:
            raise ValueError(
                "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
            )

1190
1191
1192
        if self.step_index is None:
            self._init_step_index(timestep)

1193
1194
        # Improve numerical stability for small number of steps
        lower_order_final = (self.step_index == len(self.timesteps) - 1) and (
1195
1196
1197
            self.config.euler_at_final
            or (self.config.lower_order_final and len(self.timesteps) < 15)
            or self.config.final_sigmas_type == "zero"
1198
1199
        )
        lower_order_second = (
1200
            (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15
1201
1202
        )

1203
        model_output = self.convert_model_output(model_output, sample=sample)
1204
1205
1206
1207
        for i in range(self.config.solver_order - 1):
            self.model_outputs[i] = self.model_outputs[i + 1]
        self.model_outputs[-1] = model_output

1208
1209
        # Upcast to avoid precision issues when computing prev_sample
        sample = sample.to(torch.float32)
1210
        if self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"] and variance_noise is None:
1211
            noise = randn_tensor(
1212
                model_output.shape, generator=generator, device=model_output.device, dtype=torch.float32
1213
            )
1214
1215
        elif self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"]:
            noise = variance_noise.to(device=model_output.device, dtype=torch.float32)
1216
1217
1218
        else:
            noise = None

1219
        if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final:
1220
            prev_sample = self.dpm_solver_first_order_update(model_output, sample=sample, noise=noise)
1221
        elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second:
1222
            prev_sample = self.multistep_dpm_solver_second_order_update(self.model_outputs, sample=sample, noise=noise)
1223
        else:
StAlKeR7779's avatar
StAlKeR7779 committed
1224
            prev_sample = self.multistep_dpm_solver_third_order_update(self.model_outputs, sample=sample, noise=noise)
1225
1226
1227
1228

        if self.lower_order_nums < self.config.solver_order:
            self.lower_order_nums += 1

1229
1230
1231
        # Cast sample back to expected dtype
        prev_sample = prev_sample.to(model_output.dtype)

1232
1233
1234
        # upon completion increase step index by one
        self._step_index += 1

1235
1236
1237
1238
1239
        if not return_dict:
            return (prev_sample,)

        return SchedulerOutput(prev_sample=prev_sample)

1240
    def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor:
1241
1242
1243
1244
1245
        """
        Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
        current timestep.

        Args:
1246
            sample (`torch.Tensor`):
1247
                The input sample.
1248
1249

        Returns:
1250
            `torch.Tensor`:
1251
                A scaled input sample.
1252
1253
1254
1255
1256
        """
        return sample

    def add_noise(
        self,
1257
1258
        original_samples: torch.Tensor,
        noise: torch.Tensor,
1259
        timesteps: torch.IntTensor,
1260
    ) -> torch.Tensor:
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
        """
        Add noise to the original samples according to the noise schedule at the specified timesteps.

        Args:
            original_samples (`torch.Tensor`):
                The original samples without noise.
            noise (`torch.Tensor`):
                The noise to add to the samples.
            timesteps (`torch.IntTensor`):
                The timesteps at which to add noise to the samples.

        Returns:
            `torch.Tensor`:
                The noisy samples.
        """
1276
1277
1278
1279
1280
1281
1282
1283
1284
        # Make sure sigmas and timesteps have the same device and dtype as original_samples
        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
            schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
            timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
        else:
            schedule_timesteps = self.timesteps.to(original_samples.device)
            timesteps = timesteps.to(original_samples.device)
1285

1286
        # begin_index is None when the scheduler is used for training or pipeline does not implement set_begin_index
1287
1288
        if self.begin_index is None:
            step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
1289
1290
1291
        elif self.step_index is not None:
            # add_noise is called after first denoising step (for inpainting)
            step_indices = [self.step_index] * timesteps.shape[0]
1292
        else:
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
1293
            # add noise is called before first denoising step to create initial latent(img2img)
1294
            step_indices = [self.begin_index] * timesteps.shape[0]
1295

1296
1297
1298
        sigma = sigmas[step_indices].flatten()
        while len(sigma.shape) < len(original_samples.shape):
            sigma = sigma.unsqueeze(-1)
1299

1300
1301
        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
        noisy_samples = alpha_t * original_samples + sigma_t * noise
1302
1303
1304
1305
        return noisy_samples

    def __len__(self):
        return self.config.num_train_timesteps