|
| 1 | +# Copyright 2025 The HuggingFace Team. All rights reserved. |
| 2 | +# |
| 3 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +# you may not use this file except in compliance with the License. |
| 5 | +# You may obtain a copy of the License at |
| 6 | +# |
| 7 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +# |
| 9 | +# Unless required by applicable law or agreed to in writing, software |
| 10 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | +# See the License for the specific language governing permissions and |
| 13 | +# limitations under the License. |
| 14 | + |
| 15 | +import math |
| 16 | +from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union |
| 17 | + |
| 18 | +import torch |
| 19 | + |
| 20 | +from ..configuration_utils import register_to_config |
| 21 | +from ..utils import is_kornia_available |
| 22 | +from .guider_utils import BaseGuidance, rescale_noise_cfg |
| 23 | + |
| 24 | + |
| 25 | +if TYPE_CHECKING: |
| 26 | + from ..modular_pipelines.modular_pipeline import BlockState |
| 27 | + |
| 28 | + |
| 29 | +_CAN_USE_KORNIA = is_kornia_available() |
| 30 | + |
| 31 | + |
| 32 | +if _CAN_USE_KORNIA: |
| 33 | + from kornia.geometry import pyrup as upsample_and_blur_func |
| 34 | + from kornia.geometry.transform import build_laplacian_pyramid as build_laplacian_pyramid_func |
| 35 | +else: |
| 36 | + upsample_and_blur_func = None |
| 37 | + build_laplacian_pyramid_func = None |
| 38 | + |
| 39 | + |
| 40 | +def project(v0: torch.Tensor, v1: torch.Tensor, upcast_to_double: bool = True) -> Tuple[torch.Tensor, torch.Tensor]: |
| 41 | + """ |
| 42 | + Project vector v0 onto vector v1, returning the parallel and orthogonal components of v0. Implementation from paper |
| 43 | + (Algorithm 2). |
| 44 | + """ |
| 45 | + # v0 shape: [B, ...] |
| 46 | + # v1 shape: [B, ...] |
| 47 | + # Assume first dim is a batch dim and all other dims are channel or "spatial" dims |
| 48 | + all_dims_but_first = list(range(1, len(v0.shape))) |
| 49 | + if upcast_to_double: |
| 50 | + dtype = v0.dtype |
| 51 | + v0, v1 = v0.double(), v1.double() |
| 52 | + v1 = torch.nn.functional.normalize(v1, dim=all_dims_but_first) |
| 53 | + v0_parallel = (v0 * v1).sum(dim=all_dims_but_first, keepdim=True) * v1 |
| 54 | + v0_orthogonal = v0 - v0_parallel |
| 55 | + if upcast_to_double: |
| 56 | + v0_parallel = v0_parallel.to(dtype) |
| 57 | + v0_orthogonal = v0_orthogonal.to(dtype) |
| 58 | + return v0_parallel, v0_orthogonal |
| 59 | + |
| 60 | + |
| 61 | +def build_image_from_pyramid(pyramid: List[torch.Tensor]) -> torch.Tensor: |
| 62 | + """ |
| 63 | + Recovers the data space latents from the Laplacian pyramid frequency space. Implementation from the paper |
| 64 | + (Algorihtm 2). |
| 65 | + """ |
| 66 | + # pyramid shapes: [[B, C, H, W], [B, C, H/2, W/2], ...] |
| 67 | + img = pyramid[-1] |
| 68 | + for i in range(len(pyramid) - 2, -1, -1): |
| 69 | + img = upsample_and_blur_func(img) + pyramid[i] |
| 70 | + return img |
| 71 | + |
| 72 | + |
| 73 | +class FrequencyDecoupledGuidance(BaseGuidance): |
| 74 | + """ |
| 75 | + Frequency-Decoupled Guidance (FDG): https://huggingface.co/papers/2506.19713 |
| 76 | + |
| 77 | + FDG is a technique similar to (and based on) classifier-free guidance (CFG) which is used to improve generation |
| 78 | + quality and condition-following in diffusion models. Like CFG, during training we jointly train the model on both |
| 79 | + conditional and unconditional data, and use a combination of the two during inference. (If you want more details on |
| 80 | + how CFG works, you can check out the CFG guider.) |
| 81 | + |
| 82 | + FDG differs from CFG in that the normal CFG prediction is instead decoupled into low- and high-frequency components |
| 83 | + using a frequency transform (such as a Laplacian pyramid). The CFG update is then performed in frequency space |
| 84 | + separately for the low- and high-frequency components with different guidance scales. Finally, the inverse |
| 85 | + frequency transform is used to map the CFG frequency predictions back to data space (e.g. pixel space for images) |
| 86 | + to form the final FDG prediction. |
| 87 | + |
| 88 | + For images, the FDG authors found that using low guidance scales for the low-frequency components retains sample |
| 89 | + diversity and realistic color composition, while using high guidance scales for high-frequency components enhances |
| 90 | + sample quality (such as better visual details). Therefore, they recommend using low guidance scales (low w_low) for |
| 91 | + the low-frequency components and high guidance scales (high w_high) for the high-frequency components. As an |
| 92 | + example, they suggest w_low = 5.0 and w_high = 10.0 for Stable Diffusion XL (see Table 8 in the paper). |
| 93 | + |
| 94 | + As with CFG, Diffusers implements the scaling and shifting on the unconditional prediction based on the [Imagen |
| 95 | + paper](https://huggingface.co/papers/2205.11487), which is equivalent to what the original CFG paper proposed in |
| 96 | + theory. [x_pred = x_uncond + scale * (x_cond - x_uncond)] |
| 97 | + |
| 98 | + The `use_original_formulation` argument can be set to `True` to use the original CFG formulation mentioned in the |
| 99 | + paper. By default, we use the diffusers-native implementation that has been in the codebase for a long time. |
| 100 | + |
| 101 | + Args: |
| 102 | + guidance_scales (`List[float]`, defaults to `[10.0, 5.0]`): |
| 103 | + The scale parameter for frequency-decoupled guidance for each frequency component, listed from highest |
| 104 | + frequency level to lowest. Higher values result in stronger conditioning on the text prompt, while lower |
| 105 | + values allow for more freedom in generation. Higher values may lead to saturation and deterioration of |
| 106 | + image quality. The FDG authors recommend using higher guidance scales for higher frequency components and |
| 107 | + lower guidance scales for lower frequency components (so `guidance_scales` should typically be sorted in |
| 108 | + descending order). |
| 109 | + guidance_rescale (`float` or `List[float]`, defaults to `0.0`): |
| 110 | + The rescale factor applied to the noise predictions. This is used to improve image quality and fix |
| 111 | + overexposure. Based on Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are |
| 112 | + Flawed](https://huggingface.co/papers/2305.08891). If a list is supplied, it should be the same length as |
| 113 | + `guidance_scales`. |
| 114 | + parallel_weights (`float` or `List[float]`, *optional*): |
| 115 | + Optional weights for the parallel component of each frequency component of the projected CFG shift. If not |
| 116 | + set, the weights will default to `1.0` for all components, which corresponds to using the normal CFG shift |
| 117 | + (that is, equal weights for the parallel and orthogonal components). If set, a value in `[0, 1]` is |
| 118 | + recommended. If a list is supplied, it should be the same length as `guidance_scales`. |
| 119 | + use_original_formulation (`bool`, defaults to `False`): |
| 120 | + Whether to use the original formulation of classifier-free guidance as proposed in the paper. By default, |
| 121 | + we use the diffusers-native implementation that has been in the codebase for a long time. See |
| 122 | + [~guiders.classifier_free_guidance.ClassifierFreeGuidance] for more details. |
| 123 | + start (`float` or `List[float]`, defaults to `0.0`): |
| 124 | + The fraction of the total number of denoising steps after which guidance starts. If a list is supplied, it |
| 125 | + should be the same length as `guidance_scales`. |
| 126 | + stop (`float` or `List[float]`, defaults to `1.0`): |
| 127 | + The fraction of the total number of denoising steps after which guidance stops. If a list is supplied, it |
| 128 | + should be the same length as `guidance_scales`. |
| 129 | + guidance_rescale_space (`str`, defaults to `"data"`): |
| 130 | + Whether to performance guidance rescaling in `"data"` space (after the full FDG update in data space) or in |
| 131 | + `"freq"` space (right after the CFG update, for each freq level). Note that frequency space rescaling is |
| 132 | + speculative and may not produce expected results. If `"data"` is set, the first `guidance_rescale` value |
| 133 | + will be used; otherwise, per-frequency-level guidance rescale values will be used if available. |
| 134 | + upcast_to_double (`bool`, defaults to `True`): |
| 135 | + Whether to upcast certain operations, such as the projection operation when using `parallel_weights`, to |
| 136 | + float64 when performing guidance. This may result in better performance at the cost of increased runtime. |
| 137 | + """ |
| 138 | + |
| 139 | + _input_predictions = ["pred_cond", "pred_uncond"] |
| 140 | + |
| 141 | + @register_to_config |
| 142 | + def __init__( |
| 143 | + self, |
| 144 | + guidance_scales: Union[List[float], Tuple[float]] = [10.0, 5.0], |
| 145 | + guidance_rescale: Union[float, List[float], Tuple[float]] = 0.0, |
| 146 | + parallel_weights: Optional[Union[float, List[float], Tuple[float]]] = None, |
| 147 | + use_original_formulation: bool = False, |
| 148 | + start: Union[float, List[float], Tuple[float]] = 0.0, |
| 149 | + stop: Union[float, List[float], Tuple[float]] = 1.0, |
| 150 | + guidance_rescale_space: str = "data", |
| 151 | + upcast_to_double: bool = True, |
| 152 | + ): |
| 153 | + if not _CAN_USE_KORNIA: |
| 154 | + raise ImportError( |
| 155 | + "The `FrequencyDecoupledGuidance` guider cannot be instantiated because the `kornia` library on which " |
| 156 | + "it depends is not available in the current environment. You can install `kornia` with `pip install " |
| 157 | + "kornia`." |
| 158 | + ) |
| 159 | + |
| 160 | + # Set start to earliest start for any freq component and stop to latest stop for any freq component |
| 161 | + min_start = start if isinstance(start, float) else min(start) |
| 162 | + max_stop = stop if isinstance(stop, float) else max(stop) |
| 163 | + super().__init__(min_start, max_stop) |
| 164 | + |
| 165 | + self.guidance_scales = guidance_scales |
| 166 | + self.levels = len(guidance_scales) |
| 167 | + |
| 168 | + if isinstance(guidance_rescale, float): |
| 169 | + self.guidance_rescale = [guidance_rescale] * self.levels |
| 170 | + elif len(guidance_rescale) == self.levels: |
| 171 | + self.guidance_rescale = guidance_rescale |
| 172 | + else: |
| 173 | + raise ValueError( |
| 174 | + f"`guidance_rescale` has length {len(guidance_rescale)} but should have the same length as " |
| 175 | + f"`guidance_scales` ({len(self.guidance_scales)})" |
| 176 | + ) |
| 177 | + # Whether to perform guidance rescaling in frequency space (right after the CFG update) or data space (after |
| 178 | + # transforming from frequency space back to data space) |
| 179 | + if guidance_rescale_space not in ["data", "freq"]: |
| 180 | + raise ValueError( |
| 181 | + f"Guidance rescale space is {guidance_rescale_space} but must be one of `data` or `freq`." |
| 182 | + ) |
| 183 | + self.guidance_rescale_space = guidance_rescale_space |
| 184 | + |
| 185 | + if parallel_weights is None: |
| 186 | + # Use normal CFG shift (equal weights for parallel and orthogonal components) |
| 187 | + self.parallel_weights = [1.0] * self.levels |
| 188 | + elif isinstance(parallel_weights, float): |
| 189 | + self.parallel_weights = [parallel_weights] * self.levels |
| 190 | + elif len(parallel_weights) == self.levels: |
| 191 | + self.parallel_weights = parallel_weights |
| 192 | + else: |
| 193 | + raise ValueError( |
| 194 | + f"`parallel_weights` has length {len(parallel_weights)} but should have the same length as " |
| 195 | + f"`guidance_scales` ({len(self.guidance_scales)})" |
| 196 | + ) |
| 197 | + |
| 198 | + self.use_original_formulation = use_original_formulation |
| 199 | + self.upcast_to_double = upcast_to_double |
| 200 | + |
| 201 | + if isinstance(start, float): |
| 202 | + self.guidance_start = [start] * self.levels |
| 203 | + elif len(start) == self.levels: |
| 204 | + self.guidance_start = start |
| 205 | + else: |
| 206 | + raise ValueError( |
| 207 | + f"`start` has length {len(start)} but should have the same length as `guidance_scales` " |
| 208 | + f"({len(self.guidance_scales)})" |
| 209 | + ) |
| 210 | + if isinstance(stop, float): |
| 211 | + self.guidance_stop = [stop] * self.levels |
| 212 | + elif len(stop) == self.levels: |
| 213 | + self.guidance_stop = stop |
| 214 | + else: |
| 215 | + raise ValueError( |
| 216 | + f"`stop` has length {len(stop)} but should have the same length as `guidance_scales` " |
| 217 | + f"({len(self.guidance_scales)})" |
| 218 | + ) |
| 219 | + |
| 220 | + def prepare_inputs( |
| 221 | + self, data: "BlockState", input_fields: Optional[Dict[str, Union[str, Tuple[str, str]]]] = None |
| 222 | + ) -> List["BlockState"]: |
| 223 | + if input_fields is None: |
| 224 | + input_fields = self._input_fields |
| 225 | + |
| 226 | + tuple_indices = [0] if self.num_conditions == 1 else [0, 1] |
| 227 | + data_batches = [] |
| 228 | + for i in range(self.num_conditions): |
| 229 | + data_batch = self._prepare_batch(input_fields, data, tuple_indices[i], self._input_predictions[i]) |
| 230 | + data_batches.append(data_batch) |
| 231 | + return data_batches |
| 232 | + |
| 233 | + def forward(self, pred_cond: torch.Tensor, pred_uncond: Optional[torch.Tensor] = None) -> torch.Tensor: |
| 234 | + pred = None |
| 235 | + |
| 236 | + if not self._is_fdg_enabled(): |
| 237 | + pred = pred_cond |
| 238 | + else: |
| 239 | + # Apply the frequency transform (e.g. Laplacian pyramid) to the conditional and unconditional predictions. |
| 240 | + pred_cond_pyramid = build_laplacian_pyramid_func(pred_cond, self.levels) |
| 241 | + pred_uncond_pyramid = build_laplacian_pyramid_func(pred_uncond, self.levels) |
| 242 | + |
| 243 | + # From high frequencies to low frequencies, following the paper implementation |
| 244 | + pred_guided_pyramid = [] |
| 245 | + parameters = zip(self.guidance_scales, self.parallel_weights, self.guidance_rescale) |
| 246 | + for level, (guidance_scale, parallel_weight, guidance_rescale) in enumerate(parameters): |
| 247 | + if self._is_fdg_enabled_for_level(level): |
| 248 | + # Get the cond/uncond preds (in freq space) at the current frequency level |
| 249 | + pred_cond_freq = pred_cond_pyramid[level] |
| 250 | + pred_uncond_freq = pred_uncond_pyramid[level] |
| 251 | + |
| 252 | + shift = pred_cond_freq - pred_uncond_freq |
| 253 | + |
| 254 | + # Apply parallel weights, if used (1.0 corresponds to using the normal CFG shift) |
| 255 | + if not math.isclose(parallel_weight, 1.0): |
| 256 | + shift_parallel, shift_orthogonal = project(shift, pred_cond_freq, self.upcast_to_double) |
| 257 | + shift = parallel_weight * shift_parallel + shift_orthogonal |
| 258 | + |
| 259 | + # Apply CFG update for the current frequency level |
| 260 | + pred = pred_cond_freq if self.use_original_formulation else pred_uncond_freq |
| 261 | + pred = pred + guidance_scale * shift |
| 262 | + |
| 263 | + if self.guidance_rescale_space == "freq" and guidance_rescale > 0.0: |
| 264 | + pred = rescale_noise_cfg(pred, pred_cond_freq, guidance_rescale) |
| 265 | + |
| 266 | + # Add the current FDG guided level to the FDG prediction pyramid |
| 267 | + pred_guided_pyramid.append(pred) |
| 268 | + else: |
| 269 | + # Add the current pred_cond_pyramid level as the "non-FDG" prediction |
| 270 | + pred_guided_pyramid.append(pred_cond_freq) |
| 271 | + |
| 272 | + # Convert from frequency space back to data (e.g. pixel) space by applying inverse freq transform |
| 273 | + pred = build_image_from_pyramid(pred_guided_pyramid) |
| 274 | + |
| 275 | + # If rescaling in data space, use the first elem of self.guidance_rescale as the "global" rescale value |
| 276 | + # across all freq levels |
| 277 | + if self.guidance_rescale_space == "data" and self.guidance_rescale[0] > 0.0: |
| 278 | + pred = rescale_noise_cfg(pred, pred_cond, self.guidance_rescale[0]) |
| 279 | + |
| 280 | + return pred, {} |
| 281 | + |
| 282 | + @property |
| 283 | + def is_conditional(self) -> bool: |
| 284 | + return self._count_prepared == 1 |
| 285 | + |
| 286 | + @property |
| 287 | + def num_conditions(self) -> int: |
| 288 | + num_conditions = 1 |
| 289 | + if self._is_fdg_enabled(): |
| 290 | + num_conditions += 1 |
| 291 | + return num_conditions |
| 292 | + |
| 293 | + def _is_fdg_enabled(self) -> bool: |
| 294 | + if not self._enabled: |
| 295 | + return False |
| 296 | + |
| 297 | + is_within_range = True |
| 298 | + if self._num_inference_steps is not None: |
| 299 | + skip_start_step = int(self._start * self._num_inference_steps) |
| 300 | + skip_stop_step = int(self._stop * self._num_inference_steps) |
| 301 | + is_within_range = skip_start_step <= self._step < skip_stop_step |
| 302 | + |
| 303 | + is_close = False |
| 304 | + if self.use_original_formulation: |
| 305 | + is_close = all(math.isclose(guidance_scale, 0.0) for guidance_scale in self.guidance_scales) |
| 306 | + else: |
| 307 | + is_close = all(math.isclose(guidance_scale, 1.0) for guidance_scale in self.guidance_scales) |
| 308 | + |
| 309 | + return is_within_range and not is_close |
| 310 | + |
| 311 | + def _is_fdg_enabled_for_level(self, level: int) -> bool: |
| 312 | + if not self._enabled: |
| 313 | + return False |
| 314 | + |
| 315 | + is_within_range = True |
| 316 | + if self._num_inference_steps is not None: |
| 317 | + skip_start_step = int(self.guidance_start[level] * self._num_inference_steps) |
| 318 | + skip_stop_step = int(self.guidance_stop[level] * self._num_inference_steps) |
| 319 | + is_within_range = skip_start_step <= self._step < skip_stop_step |
| 320 | + |
| 321 | + is_close = False |
| 322 | + if self.use_original_formulation: |
| 323 | + is_close = math.isclose(self.guidance_scales[level], 0.0) |
| 324 | + else: |
| 325 | + is_close = math.isclose(self.guidance_scales[level], 1.0) |
| 326 | + |
| 327 | + return is_within_range and not is_close |
0 commit comments