|
| 1 | +# Copyright 2024 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 | +from typing import TYPE_CHECKING, Any, Dict, List, Tuple, Union |
| 16 | + |
| 17 | +import torch |
| 18 | + |
| 19 | +from ..utils import get_logger |
| 20 | + |
| 21 | + |
| 22 | +if TYPE_CHECKING: |
| 23 | + from ..models.attention_processor import AttentionProcessor |
| 24 | + |
| 25 | + |
| 26 | +logger = get_logger(__name__) # pylint: disable=invalid-name |
| 27 | + |
| 28 | + |
| 29 | +class BaseGuidance: |
| 30 | + r"""Base class providing the skeleton for implementing guidance techniques.""" |
| 31 | + |
| 32 | + _input_predictions = None |
| 33 | + |
| 34 | + def __init__(self, start: float = 0.0, stop: float = 1.0): |
| 35 | + self._start = start |
| 36 | + self._stop = stop |
| 37 | + self._step: int = None |
| 38 | + self._num_inference_steps: int = None |
| 39 | + self._timestep: torch.LongTensor = None |
| 40 | + self._preds: Dict[str, torch.Tensor] = {} |
| 41 | + self._num_outputs_prepared: int = 0 |
| 42 | + |
| 43 | + if not (0.0 <= start < 1.0): |
| 44 | + raise ValueError( |
| 45 | + f"Expected `start` to be between 0.0 and 1.0, but got {start}." |
| 46 | + ) |
| 47 | + if not (start <= stop <= 1.0): |
| 48 | + raise ValueError( |
| 49 | + f"Expected `stop` to be between {start} and 1.0, but got {stop}." |
| 50 | + ) |
| 51 | + |
| 52 | + if self._input_predictions is None or not isinstance(self._input_predictions, list): |
| 53 | + raise ValueError( |
| 54 | + "`_input_predictions` must be a list of required prediction names for the guidance technique." |
| 55 | + ) |
| 56 | + |
| 57 | + def set_state(self, step: int, num_inference_steps: int, timestep: torch.LongTensor) -> None: |
| 58 | + self._step = step |
| 59 | + self._num_inference_steps = num_inference_steps |
| 60 | + self._timestep = timestep |
| 61 | + self._preds = {} |
| 62 | + self._num_outputs_prepared = 0 |
| 63 | + |
| 64 | + def prepare_inputs(self, denoiser: torch.nn.Module, *args: Union[Tuple[torch.Tensor], List[torch.Tensor]]) -> Tuple[List[torch.Tensor], ...]: |
| 65 | + raise NotImplementedError("GuidanceMixin::prepare_inputs must be implemented in subclasses.") |
| 66 | + |
| 67 | + def prepare_outputs(self, denoiser: torch.nn.Module, pred: torch.Tensor) -> None: |
| 68 | + raise NotImplementedError("GuidanceMixin::prepare_outputs must be implemented in subclasses.") |
| 69 | + |
| 70 | + def __call__(self, **kwargs) -> Any: |
| 71 | + if len(kwargs) != self.num_conditions: |
| 72 | + raise ValueError( |
| 73 | + f"Expected {self.num_conditions} arguments, but got {len(kwargs)}. Please provide the correct number of arguments." |
| 74 | + ) |
| 75 | + return self.forward(**kwargs) |
| 76 | + |
| 77 | + def forward(self, *args, **kwargs) -> Any: |
| 78 | + raise NotImplementedError("GuidanceMixin::forward must be implemented in subclasses.") |
| 79 | + |
| 80 | + @property |
| 81 | + def num_conditions(self) -> int: |
| 82 | + raise NotImplementedError("GuidanceMixin::num_conditions must be implemented in subclasses.") |
| 83 | + |
| 84 | + @property |
| 85 | + def outputs(self) -> Dict[str, torch.Tensor]: |
| 86 | + return self._preds |
| 87 | + |
| 88 | + |
| 89 | +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): |
| 90 | + r""" |
| 91 | + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on |
| 92 | + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are |
| 93 | + Flawed](https://arxiv.org/pdf/2305.08891.pdf). |
| 94 | + Args: |
| 95 | + noise_cfg (`torch.Tensor`): |
| 96 | + The predicted noise tensor for the guided diffusion process. |
| 97 | + noise_pred_text (`torch.Tensor`): |
| 98 | + The predicted noise tensor for the text-guided diffusion process. |
| 99 | + guidance_rescale (`float`, *optional*, defaults to 0.0): |
| 100 | + A rescale factor applied to the noise predictions. |
| 101 | + Returns: |
| 102 | + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. |
| 103 | + """ |
| 104 | + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) |
| 105 | + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) |
| 106 | + # rescale the results from guidance (fixes overexposure) |
| 107 | + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) |
| 108 | + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images |
| 109 | + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg |
| 110 | + return noise_cfg |
| 111 | + |
| 112 | + |
| 113 | +def _default_prepare_inputs(denoiser: torch.nn.Module, num_conditions: int, *args: Union[Tuple[torch.Tensor], List[torch.Tensor]]) -> Tuple[List[torch.Tensor], ...]: |
| 114 | + """ |
| 115 | + Prepares the inputs for the denoiser by ensuring that the conditional and unconditional inputs are correctly |
| 116 | + prepared based on required number of conditions. This function is used in the `prepare_inputs` method of the |
| 117 | + `GuidanceMixin` class. |
| 118 | +
|
| 119 | + Either tensors or tuples/lists of tensors can be provided. If a tuple/list is provided, it should contain two elements: |
| 120 | + - The first element is the conditional input. |
| 121 | + - The second element is the unconditional input or None. |
| 122 | + |
| 123 | + If only the conditional input is provided, it will be repeated for all batches. |
| 124 | + |
| 125 | + If both conditional and unconditional inputs are provided, they are alternated as batches of data. |
| 126 | + """ |
| 127 | + list_of_inputs = [] |
| 128 | + for arg in args: |
| 129 | + if arg is None or isinstance(arg, torch.Tensor): |
| 130 | + list_of_inputs.append([arg] * num_conditions) |
| 131 | + elif isinstance(arg, (tuple, list)): |
| 132 | + if len(arg) != 2: |
| 133 | + raise ValueError( |
| 134 | + f"Expected a tuple or list of length 2, but got {len(arg)} for argument {arg}. Please provide a tuple/list of length 2 " |
| 135 | + f"with the first element being the conditional input and the second element being the unconditional input or None." |
| 136 | + ) |
| 137 | + if arg[1] is None: |
| 138 | + # Only conditioning inputs for all batches |
| 139 | + list_of_inputs.append([arg[0]] * num_conditions) |
| 140 | + else: |
| 141 | + # Alternating conditional and unconditional inputs as batches |
| 142 | + inputs = [arg[i % 2] for i in range(num_conditions)] |
| 143 | + list_of_inputs.append(inputs) |
| 144 | + else: |
| 145 | + raise ValueError( |
| 146 | + f"Expected a tensor, tuple, or list, but got {type(arg)} for argument {arg}. Please provide a tensor, tuple, or list." |
| 147 | + ) |
| 148 | + return tuple(list_of_inputs) |
0 commit comments