auto_pipeline.py 57.5 KB
Newer Older
YiYi Xu's avatar
YiYi Xu committed
1
# coding=utf-8
2
# Copyright 2025 The HuggingFace Inc. team.
YiYi Xu's avatar
YiYi Xu committed
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.

from collections import OrderedDict

18
19
from huggingface_hub.utils import validate_hf_hub_args

YiYi Xu's avatar
YiYi Xu committed
20
from ..configuration_utils import ConfigMixin
21
from ..models.controlnets import ControlNetUnionModel
22
from ..utils import is_sentencepiece_available
23
from .aura_flow import AuraFlowPipeline
Edna's avatar
Edna committed
24
from .chroma import ChromaPipeline
Yuxuan.Zhang's avatar
Yuxuan.Zhang committed
25
from .cogview3 import CogView3PlusPipeline
Yuxuan Zhang's avatar
Yuxuan Zhang committed
26
from .cogview4 import CogView4ControlPipeline, CogView4Pipeline
YiYi Xu's avatar
YiYi Xu committed
27
28
29
30
from .controlnet import (
    StableDiffusionControlNetImg2ImgPipeline,
    StableDiffusionControlNetInpaintPipeline,
    StableDiffusionControlNetPipeline,
31
    StableDiffusionXLControlNetImg2ImgPipeline,
32
    StableDiffusionXLControlNetInpaintPipeline,
YiYi Xu's avatar
YiYi Xu committed
33
    StableDiffusionXLControlNetPipeline,
34
35
36
    StableDiffusionXLControlNetUnionImg2ImgPipeline,
    StableDiffusionXLControlNetUnionInpaintPipeline,
    StableDiffusionXLControlNetUnionPipeline,
YiYi Xu's avatar
YiYi Xu committed
37
)
38
39
40
41
from .controlnet_sd3 import (
    StableDiffusion3ControlNetInpaintingPipeline,
    StableDiffusion3ControlNetPipeline,
)
YiYi Xu's avatar
YiYi Xu committed
42
from .deepfloyd_if import IFImg2ImgPipeline, IFInpaintingPipeline, IFPipeline
43
from .flux import (
44
45
    FluxControlImg2ImgPipeline,
    FluxControlInpaintPipeline,
46
47
48
    FluxControlNetImg2ImgPipeline,
    FluxControlNetInpaintPipeline,
    FluxControlNetPipeline,
49
    FluxControlPipeline,
50
51
    FluxImg2ImgPipeline,
    FluxInpaintPipeline,
jlonge4's avatar
jlonge4 committed
52
    FluxKontextPipeline,
53
54
    FluxPipeline,
)
55
from .hunyuandit import HunyuanDiTPipeline
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
from .kandinsky import (
    KandinskyCombinedPipeline,
    KandinskyImg2ImgCombinedPipeline,
    KandinskyImg2ImgPipeline,
    KandinskyInpaintCombinedPipeline,
    KandinskyInpaintPipeline,
    KandinskyPipeline,
)
from .kandinsky2_2 import (
    KandinskyV22CombinedPipeline,
    KandinskyV22Img2ImgCombinedPipeline,
    KandinskyV22Img2ImgPipeline,
    KandinskyV22InpaintCombinedPipeline,
    KandinskyV22InpaintPipeline,
    KandinskyV22Pipeline,
)
72
from .kandinsky3 import Kandinsky3Img2ImgPipeline, Kandinsky3Pipeline
73
from .latent_consistency_models import LatentConsistencyModelImg2ImgPipeline, LatentConsistencyModelPipeline
74
75
from .lumina import LuminaPipeline
from .lumina2 import Lumina2Pipeline
YiYi Xu's avatar
YiYi Xu committed
76
from .pag import (
77
    HunyuanDiTPAGPipeline,
78
    PixArtSigmaPAGPipeline,
79
    SanaPAGPipeline,
80
    StableDiffusion3PAGImg2ImgPipeline,
81
    StableDiffusion3PAGPipeline,
82
    StableDiffusionControlNetPAGInpaintPipeline,
83
    StableDiffusionControlNetPAGPipeline,
84
    StableDiffusionPAGImg2ImgPipeline,
85
    StableDiffusionPAGInpaintPipeline,
86
    StableDiffusionPAGPipeline,
87
    StableDiffusionXLControlNetPAGImg2ImgPipeline,
YiYi Xu's avatar
YiYi Xu committed
88
89
90
91
92
    StableDiffusionXLControlNetPAGPipeline,
    StableDiffusionXLPAGImg2ImgPipeline,
    StableDiffusionXLPAGInpaintPipeline,
    StableDiffusionXLPAGPipeline,
)
93
from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline
YiYi Xu's avatar
YiYi Xu committed
94
95
96
97
98
99
100
101
from .qwenimage import (
    QwenImageControlNetPipeline,
    QwenImageEditInpaintPipeline,
    QwenImageEditPipeline,
    QwenImageImg2ImgPipeline,
    QwenImageInpaintPipeline,
    QwenImagePipeline,
)
102
from .sana import SanaPipeline
103
from .stable_cascade import StableCascadeCombinedPipeline, StableCascadeDecoderPipeline
YiYi Xu's avatar
YiYi Xu committed
104
105
106
107
108
from .stable_diffusion import (
    StableDiffusionImg2ImgPipeline,
    StableDiffusionInpaintPipeline,
    StableDiffusionPipeline,
)
109
110
from .stable_diffusion_3 import (
    StableDiffusion3Img2ImgPipeline,
111
    StableDiffusion3InpaintPipeline,
112
113
    StableDiffusion3Pipeline,
)
YiYi Xu's avatar
YiYi Xu committed
114
115
116
117
118
from .stable_diffusion_xl import (
    StableDiffusionXLImg2ImgPipeline,
    StableDiffusionXLInpaintPipeline,
    StableDiffusionXLPipeline,
)
Kashif Rasul's avatar
Kashif Rasul committed
119
from .wuerstchen import WuerstchenCombinedPipeline, WuerstchenDecoderPipeline
YiYi Xu's avatar
YiYi Xu committed
120
121
122
123
124
125


AUTO_TEXT2IMAGE_PIPELINES_MAPPING = OrderedDict(
    [
        ("stable-diffusion", StableDiffusionPipeline),
        ("stable-diffusion-xl", StableDiffusionXLPipeline),
126
        ("stable-diffusion-3", StableDiffusion3Pipeline),
127
        ("stable-diffusion-3-pag", StableDiffusion3PAGPipeline),
YiYi Xu's avatar
YiYi Xu committed
128
        ("if", IFPipeline),
129
        ("hunyuan", HunyuanDiTPipeline),
130
        ("hunyuan-pag", HunyuanDiTPAGPipeline),
131
132
        ("kandinsky", KandinskyCombinedPipeline),
        ("kandinsky22", KandinskyV22CombinedPipeline),
133
        ("kandinsky3", Kandinsky3Pipeline),
YiYi Xu's avatar
YiYi Xu committed
134
135
        ("stable-diffusion-controlnet", StableDiffusionControlNetPipeline),
        ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetPipeline),
136
        ("stable-diffusion-xl-controlnet-union", StableDiffusionXLControlNetUnionPipeline),
137
        ("stable-diffusion-3-controlnet", StableDiffusion3ControlNetPipeline),
Kashif Rasul's avatar
Kashif Rasul committed
138
        ("wuerstchen", WuerstchenCombinedPipeline),
139
        ("cascade", StableCascadeCombinedPipeline),
140
        ("lcm", LatentConsistencyModelPipeline),
141
142
        ("pixart-alpha", PixArtAlphaPipeline),
        ("pixart-sigma", PixArtSigmaPipeline),
143
144
        ("sana", SanaPipeline),
        ("sana-pag", SanaPAGPipeline),
145
        ("stable-diffusion-pag", StableDiffusionPAGPipeline),
146
        ("stable-diffusion-controlnet-pag", StableDiffusionControlNetPAGPipeline),
YiYi Xu's avatar
YiYi Xu committed
147
148
        ("stable-diffusion-xl-pag", StableDiffusionXLPAGPipeline),
        ("stable-diffusion-xl-controlnet-pag", StableDiffusionXLControlNetPAGPipeline),
149
        ("pixart-sigma-pag", PixArtSigmaPAGPipeline),
150
        ("auraflow", AuraFlowPipeline),
Sayak Paul's avatar
Sayak Paul committed
151
        ("flux", FluxPipeline),
152
        ("flux-control", FluxControlPipeline),
153
        ("flux-controlnet", FluxControlNetPipeline),
jlonge4's avatar
jlonge4 committed
154
        ("flux-kontext", FluxKontextPipeline),
155
156
        ("lumina", LuminaPipeline),
        ("lumina2", Lumina2Pipeline),
Edna's avatar
Edna committed
157
        ("chroma", ChromaPipeline),
Yuxuan.Zhang's avatar
Yuxuan.Zhang committed
158
        ("cogview3", CogView3PlusPipeline),
159
        ("cogview4", CogView4Pipeline),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
160
        ("cogview4-control", CogView4ControlPipeline),
YiYi Xu's avatar
YiYi Xu committed
161
162
        ("qwenimage", QwenImagePipeline),
        ("qwenimage-controlnet", QwenImageControlNetPipeline),
YiYi Xu's avatar
YiYi Xu committed
163
164
165
166
167
168
169
    ]
)

AUTO_IMAGE2IMAGE_PIPELINES_MAPPING = OrderedDict(
    [
        ("stable-diffusion", StableDiffusionImg2ImgPipeline),
        ("stable-diffusion-xl", StableDiffusionXLImg2ImgPipeline),
170
        ("stable-diffusion-3", StableDiffusion3Img2ImgPipeline),
171
        ("stable-diffusion-3-pag", StableDiffusion3PAGImg2ImgPipeline),
YiYi Xu's avatar
YiYi Xu committed
172
        ("if", IFImg2ImgPipeline),
173
174
        ("kandinsky", KandinskyImg2ImgCombinedPipeline),
        ("kandinsky22", KandinskyV22Img2ImgCombinedPipeline),
175
        ("kandinsky3", Kandinsky3Img2ImgPipeline),
YiYi Xu's avatar
YiYi Xu committed
176
        ("stable-diffusion-controlnet", StableDiffusionControlNetImg2ImgPipeline),
177
        ("stable-diffusion-pag", StableDiffusionPAGImg2ImgPipeline),
178
        ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetImg2ImgPipeline),
179
        ("stable-diffusion-xl-controlnet-union", StableDiffusionXLControlNetUnionImg2ImgPipeline),
YiYi Xu's avatar
YiYi Xu committed
180
        ("stable-diffusion-xl-pag", StableDiffusionXLPAGImg2ImgPipeline),
181
        ("stable-diffusion-xl-controlnet-pag", StableDiffusionXLControlNetPAGImg2ImgPipeline),
182
        ("lcm", LatentConsistencyModelImg2ImgPipeline),
183
        ("flux", FluxImg2ImgPipeline),
184
        ("flux-controlnet", FluxControlNetImg2ImgPipeline),
185
        ("flux-control", FluxControlImg2ImgPipeline),
jlonge4's avatar
jlonge4 committed
186
        ("flux-kontext", FluxKontextPipeline),
YiYi Xu's avatar
YiYi Xu committed
187
188
        ("qwenimage", QwenImageImg2ImgPipeline),
        ("qwenimage-edit", QwenImageEditPipeline),
YiYi Xu's avatar
YiYi Xu committed
189
190
191
192
193
194
195
    ]
)

AUTO_INPAINT_PIPELINES_MAPPING = OrderedDict(
    [
        ("stable-diffusion", StableDiffusionInpaintPipeline),
        ("stable-diffusion-xl", StableDiffusionXLInpaintPipeline),
196
        ("stable-diffusion-3", StableDiffusion3InpaintPipeline),
YiYi Xu's avatar
YiYi Xu committed
197
        ("if", IFInpaintingPipeline),
198
199
200
        ("kandinsky", KandinskyInpaintCombinedPipeline),
        ("kandinsky22", KandinskyV22InpaintCombinedPipeline),
        ("stable-diffusion-controlnet", StableDiffusionControlNetInpaintPipeline),
201
        ("stable-diffusion-controlnet-pag", StableDiffusionControlNetPAGInpaintPipeline),
202
        ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetInpaintPipeline),
203
        ("stable-diffusion-xl-controlnet-union", StableDiffusionXLControlNetUnionInpaintPipeline),
204
        ("stable-diffusion-3-controlnet", StableDiffusion3ControlNetInpaintingPipeline),
YiYi Xu's avatar
YiYi Xu committed
205
        ("stable-diffusion-xl-pag", StableDiffusionXLPAGInpaintPipeline),
206
        ("flux", FluxInpaintPipeline),
207
        ("flux-controlnet", FluxControlNetInpaintPipeline),
208
        ("flux-control", FluxControlInpaintPipeline),
209
        ("stable-diffusion-pag", StableDiffusionPAGInpaintPipeline),
YiYi Xu's avatar
YiYi Xu committed
210
211
        ("qwenimage", QwenImageInpaintPipeline),
        ("qwenimage-edit", QwenImageEditInpaintPipeline),
212
213
214
215
216
217
218
    ]
)

_AUTO_TEXT2IMAGE_DECODER_PIPELINES_MAPPING = OrderedDict(
    [
        ("kandinsky", KandinskyPipeline),
        ("kandinsky22", KandinskyV22Pipeline),
Kashif Rasul's avatar
Kashif Rasul committed
219
        ("wuerstchen", WuerstchenDecoderPipeline),
220
        ("cascade", StableCascadeDecoderPipeline),
221
222
223
224
225
226
227
228
229
230
    ]
)
_AUTO_IMAGE2IMAGE_DECODER_PIPELINES_MAPPING = OrderedDict(
    [
        ("kandinsky", KandinskyImg2ImgPipeline),
        ("kandinsky22", KandinskyV22Img2ImgPipeline),
    ]
)
_AUTO_INPAINT_DECODER_PIPELINES_MAPPING = OrderedDict(
    [
YiYi Xu's avatar
YiYi Xu committed
231
232
233
234
235
        ("kandinsky", KandinskyInpaintPipeline),
        ("kandinsky22", KandinskyV22InpaintPipeline),
    ]
)

236
if is_sentencepiece_available():
237
    from .kolors import KolorsImg2ImgPipeline, KolorsPipeline
Álvaro Somoza's avatar
Álvaro Somoza committed
238
    from .pag import KolorsPAGPipeline
239
240

    AUTO_TEXT2IMAGE_PIPELINES_MAPPING["kolors"] = KolorsPipeline
Álvaro Somoza's avatar
Álvaro Somoza committed
241
    AUTO_TEXT2IMAGE_PIPELINES_MAPPING["kolors-pag"] = KolorsPAGPipeline
242
    AUTO_IMAGE2IMAGE_PIPELINES_MAPPING["kolors"] = KolorsImg2ImgPipeline
243

YiYi Xu's avatar
YiYi Xu committed
244
245
246
247
SUPPORTED_TASKS_MAPPINGS = [
    AUTO_TEXT2IMAGE_PIPELINES_MAPPING,
    AUTO_IMAGE2IMAGE_PIPELINES_MAPPING,
    AUTO_INPAINT_PIPELINES_MAPPING,
248
249
250
    _AUTO_TEXT2IMAGE_DECODER_PIPELINES_MAPPING,
    _AUTO_IMAGE2IMAGE_DECODER_PIPELINES_MAPPING,
    _AUTO_INPAINT_DECODER_PIPELINES_MAPPING,
YiYi Xu's avatar
YiYi Xu committed
251
252
253
]


254
255
256
257
258
259
260
261
262
263
264
265
266
267
def _get_connected_pipeline(pipeline_cls):
    # for now connected pipelines can only be loaded from decoder pipelines, such as kandinsky-community/kandinsky-2-2-decoder
    if pipeline_cls in _AUTO_TEXT2IMAGE_DECODER_PIPELINES_MAPPING.values():
        return _get_task_class(
            AUTO_TEXT2IMAGE_PIPELINES_MAPPING, pipeline_cls.__name__, throw_error_if_not_exist=False
        )
    if pipeline_cls in _AUTO_IMAGE2IMAGE_DECODER_PIPELINES_MAPPING.values():
        return _get_task_class(
            AUTO_IMAGE2IMAGE_PIPELINES_MAPPING, pipeline_cls.__name__, throw_error_if_not_exist=False
        )
    if pipeline_cls in _AUTO_INPAINT_DECODER_PIPELINES_MAPPING.values():
        return _get_task_class(AUTO_INPAINT_PIPELINES_MAPPING, pipeline_cls.__name__, throw_error_if_not_exist=False)


YiYi Xu's avatar
YiYi Xu committed
268
269
270
271
272
273
def _get_model(pipeline_class_name):
    for task_mapping in SUPPORTED_TASKS_MAPPINGS:
        for model_name, pipeline in task_mapping.items():
            if pipeline.__name__ == pipeline_class_name:
                return model_name

YiYi Xu's avatar
YiYi Xu committed
274

YiYi Xu's avatar
YiYi Xu committed
275
276
def _get_task_class(mapping, pipeline_class_name, throw_error_if_not_exist: bool = True):
    model_name = _get_model(pipeline_class_name)
YiYi Xu's avatar
YiYi Xu committed
277
278
279
280
281

    if model_name is not None:
        task_class = mapping.get(model_name, None)
        if task_class is not None:
            return task_class
282
283
284

    if throw_error_if_not_exist:
        raise ValueError(f"AutoPipeline can't find a pipeline linked to {pipeline_class_name} for {model_name}")
YiYi Xu's avatar
YiYi Xu committed
285
286
287
288
289


class AutoPipelineForText2Image(ConfigMixin):
    r"""

290
291
292
    [`AutoPipelineForText2Image`] is a generic pipeline class that instantiates a text-to-image pipeline class. The
    specific underlying pipeline class is automatically selected from either the
    [`~AutoPipelineForText2Image.from_pretrained`] or [`~AutoPipelineForText2Image.from_pipe`] methods.
YiYi Xu's avatar
YiYi Xu committed
293

294
    This class cannot be instantiated using `__init__()` (throws an error).
YiYi Xu's avatar
YiYi Xu committed
295
296
297
298
299
300
301

    Class attributes:

        - **config_name** (`str`) -- The configuration filename that stores the class and module names of all the
          diffusion pipeline's components.

    """
302

YiYi Xu's avatar
YiYi Xu committed
303
304
305
306
307
308
309
310
311
312
    config_name = "model_index.json"

    def __init__(self, *args, **kwargs):
        raise EnvironmentError(
            f"{self.__class__.__name__} is designed to be instantiated "
            f"using the `{self.__class__.__name__}.from_pretrained(pretrained_model_name_or_path)` or "
            f"`{self.__class__.__name__}.from_pipe(pipeline)` methods."
        )

    @classmethod
313
    @validate_hf_hub_args
YiYi Xu's avatar
YiYi Xu committed
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
    def from_pretrained(cls, pretrained_model_or_path, **kwargs):
        r"""
        Instantiates a text-to-image Pytorch diffusion pipeline from pretrained pipeline weight.

        The from_pretrained() method takes care of returning the correct pipeline class instance by:
            1. Detect the pipeline class of the pretrained_model_or_path based on the _class_name property of its
               config object
            2. Find the text-to-image pipeline linked to the pipeline class using pattern matching on pipeline class
               name.

        If a `controlnet` argument is passed, it will instantiate a [`StableDiffusionControlNetPipeline`] object.

        The pipeline is set in evaluation mode (`model.eval()`) by default.

        If you get the error message below, you need to finetune the weights for your downstream task:

        ```
331
        Some weights of UNet2DConditionModel were not initialized from the model checkpoint at stable-diffusion-v1-5/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
YiYi Xu's avatar
YiYi Xu committed
332
333
334
335
336
        - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
        You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
        ```

        Parameters:
337
            pretrained_model_or_path (`str` or `os.PathLike`, *optional*):
YiYi Xu's avatar
YiYi Xu committed
338
339
340
341
342
343
344
                Can be either:

                    - A string, the *repo id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline
                      hosted on the Hub.
                    - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights
                      saved using
                    [`~DiffusionPipeline.save_pretrained`].
345
346
            torch_dtype (`torch.dtype`, *optional*):
                Override the default `torch.dtype` and load the model with another dtype.
YiYi Xu's avatar
YiYi Xu committed
347
348
349
350
351
352
            force_download (`bool`, *optional*, defaults to `False`):
                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                cached versions if they exist.
            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
353

YiYi Xu's avatar
YiYi Xu committed
354
355
356
357
358
359
360
361
            proxies (`Dict[str, str]`, *optional*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            output_loading_info(`bool`, *optional*, defaults to `False`):
                Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
            local_files_only (`bool`, *optional*, defaults to `False`):
                Whether to only load local model weights and configuration files or not. If set to `True`, the model
                won't be downloaded from the Hub.
362
            token (`str` or *bool*, *optional*):
YiYi Xu's avatar
YiYi Xu committed
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
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
407
408
409
                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                `diffusers-cli login` (stored in `~/.huggingface`) is used.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            custom_revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, or a commit id similar to
                `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a
                custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub.
            mirror (`str`, *optional*):
                Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not
                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
                information.
            device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
                A map that specifies where each submodule should go. It doesn’t need to be defined for each
                parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the
                same device.

                Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
                more information about each option see [designing a device
                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
            max_memory (`Dict`, *optional*):
                A dictionary device identifier for the maximum memory. Will default to the maximum memory available for
                each GPU and the available CPU RAM if unset.
            offload_folder (`str` or `os.PathLike`, *optional*):
                The path to offload weights if device_map contains the value `"disk"`.
            offload_state_dict (`bool`, *optional*):
                If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if
                the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True`
                when there is some disk offload.
            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
                argument to `True` will raise an error.
            use_safetensors (`bool`, *optional*, defaults to `None`):
                If set to `None`, the safetensors weights are downloaded if they're available **and** if the
                safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
                weights. If set to `False`, safetensors weights are not loaded.
            kwargs (remaining dictionary of keyword arguments, *optional*):
                Can be used to overwrite load and saveable variables (the pipeline components of the specific pipeline
                class). The overwritten components are passed directly to the pipelines `__init__` method. See example
                below for more information.
            variant (`str`, *optional*):
                Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when
                loading `from_flax`.

Steven Liu's avatar
Steven Liu committed
410
411
        > [!TIP] > To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in
        with `hf > auth login`.
YiYi Xu's avatar
YiYi Xu committed
412
413
414
415

        Examples:

        ```py
416
        >>> from diffusers import AutoPipelineForText2Image
YiYi Xu's avatar
YiYi Xu committed
417

418
        >>> pipeline = AutoPipelineForText2Image.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5")
419
        >>> image = pipeline(prompt).images[0]
YiYi Xu's avatar
YiYi Xu committed
420
421
        ```
        """
422
        cache_dir = kwargs.pop("cache_dir", None)
423
424
        force_download = kwargs.pop("force_download", False)
        proxies = kwargs.pop("proxies", None)
425
        token = kwargs.pop("token", None)
426
427
428
429
430
431
432
        local_files_only = kwargs.pop("local_files_only", False)
        revision = kwargs.pop("revision", None)

        load_config_kwargs = {
            "cache_dir": cache_dir,
            "force_download": force_download,
            "proxies": proxies,
433
            "token": token,
434
435
436
437
438
            "local_files_only": local_files_only,
            "revision": revision,
        }

        config = cls.load_config(pretrained_model_or_path, **load_config_kwargs)
YiYi Xu's avatar
YiYi Xu committed
439
        orig_class_name = config["_class_name"]
440
441
442
443
        if "ControlPipeline" in orig_class_name:
            to_replace = "ControlPipeline"
        else:
            to_replace = "Pipeline"
YiYi Xu's avatar
YiYi Xu committed
444
445

        if "controlnet" in kwargs:
446
            if isinstance(kwargs["controlnet"], ControlNetUnionModel):
447
                orig_class_name = config["_class_name"].replace(to_replace, "ControlNetUnionPipeline")
448
            else:
449
                orig_class_name = config["_class_name"].replace(to_replace, "ControlNetPipeline")
YiYi Xu's avatar
YiYi Xu committed
450
451
452
        if "enable_pag" in kwargs:
            enable_pag = kwargs.pop("enable_pag")
            if enable_pag:
453
                orig_class_name = orig_class_name.replace(to_replace, "PAGPipeline")
YiYi Xu's avatar
YiYi Xu committed
454
455
456

        text_2_image_cls = _get_task_class(AUTO_TEXT2IMAGE_PIPELINES_MAPPING, orig_class_name)

457
        kwargs = {**load_config_kwargs, **kwargs}
YiYi Xu's avatar
YiYi Xu committed
458
459
460
461
462
463
464
465
466
467
468
        return text_2_image_cls.from_pretrained(pretrained_model_or_path, **kwargs)

    @classmethod
    def from_pipe(cls, pipeline, **kwargs):
        r"""
        Instantiates a text-to-image Pytorch diffusion pipeline from another instantiated diffusion pipeline class.

        The from_pipe() method takes care of returning the correct pipeline class instance by finding the text-to-image
        pipeline linked to the pipeline class using pattern matching on pipeline class name.

        All the modules the pipeline contains will be used to initialize the new pipeline without reallocating
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
469
        additional memory.
YiYi Xu's avatar
YiYi Xu committed
470
471
472
473
474
475
476
477

        The pipeline is set in evaluation mode (`model.eval()`) by default.

        Parameters:
            pipeline (`DiffusionPipeline`):
                an instantiated `DiffusionPipeline` object

        ```py
478
        >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForImage2Image
YiYi Xu's avatar
YiYi Xu committed
479
480

        >>> pipe_i2i = AutoPipelineForImage2Image.from_pretrained(
481
        ...     "stable-diffusion-v1-5/stable-diffusion-v1-5", requires_safety_checker=False
YiYi Xu's avatar
YiYi Xu committed
482
483
        ... )

484
485
        >>> pipe_t2i = AutoPipelineForText2Image.from_pipe(pipe_i2i)
        >>> image = pipe_t2i(prompt).images[0]
YiYi Xu's avatar
YiYi Xu committed
486
487
488
489
490
491
492
493
494
        ```
        """

        original_config = dict(pipeline.config)
        original_cls_name = pipeline.__class__.__name__

        # derive the pipeline class to instantiate
        text_2_image_cls = _get_task_class(AUTO_TEXT2IMAGE_PIPELINES_MAPPING, original_cls_name)

495
496
        if "controlnet" in kwargs:
            if kwargs["controlnet"] is not None:
YiYi Xu's avatar
YiYi Xu committed
497
                to_replace = "PAGPipeline" if "PAG" in text_2_image_cls.__name__ else "Pipeline"
498
499
                text_2_image_cls = _get_task_class(
                    AUTO_TEXT2IMAGE_PIPELINES_MAPPING,
YiYi Xu's avatar
YiYi Xu committed
500
                    text_2_image_cls.__name__.replace("ControlNet", "").replace(to_replace, "ControlNet" + to_replace),
501
502
503
504
                )
            else:
                text_2_image_cls = _get_task_class(
                    AUTO_TEXT2IMAGE_PIPELINES_MAPPING,
YiYi Xu's avatar
YiYi Xu committed
505
506
507
508
509
510
511
512
513
514
515
516
517
518
                    text_2_image_cls.__name__.replace("ControlNet", ""),
                )

        if "enable_pag" in kwargs:
            enable_pag = kwargs.pop("enable_pag")
            if enable_pag:
                text_2_image_cls = _get_task_class(
                    AUTO_TEXT2IMAGE_PIPELINES_MAPPING,
                    text_2_image_cls.__name__.replace("PAG", "").replace("Pipeline", "PAGPipeline"),
                )
            else:
                text_2_image_cls = _get_task_class(
                    AUTO_TEXT2IMAGE_PIPELINES_MAPPING,
                    text_2_image_cls.__name__.replace("PAG", ""),
519
520
                )

YiYi Xu's avatar
YiYi Xu committed
521
        # define expected module and optional kwargs given the pipeline signature
522
        expected_modules, optional_kwargs = text_2_image_cls._get_signature_keys(text_2_image_cls)
YiYi Xu's avatar
YiYi Xu committed
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560

        pretrained_model_name_or_path = original_config.pop("_name_or_path", None)

        # allow users pass modules in `kwargs` to override the original pipeline's components
        passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs}
        original_class_obj = {
            k: pipeline.components[k]
            for k, v in pipeline.components.items()
            if k in expected_modules and k not in passed_class_obj
        }

        # allow users pass optional kwargs to override the original pipelines config attribute
        passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs}
        original_pipe_kwargs = {
            k: original_config[k]
            for k, v in original_config.items()
            if k in optional_kwargs and k not in passed_pipe_kwargs
        }

        # config that were not expected by original pipeline is stored as private attribute
        # we will pass them as optional arguments if they can be accepted by the pipeline
        additional_pipe_kwargs = [
            k[1:]
            for k in original_config.keys()
            if k.startswith("_") and k[1:] in optional_kwargs and k[1:] not in passed_pipe_kwargs
        ]
        for k in additional_pipe_kwargs:
            original_pipe_kwargs[k] = original_config.pop(f"_{k}")

        text_2_image_kwargs = {**passed_class_obj, **original_class_obj, **passed_pipe_kwargs, **original_pipe_kwargs}

        # store unused config as private attribute
        unused_original_config = {
            f"{'' if k.startswith('_') else '_'}{k}": original_config[k]
            for k, v in original_config.items()
            if k not in text_2_image_kwargs
        }

561
562
563
        missing_modules = (
            set(expected_modules) - set(text_2_image_cls._optional_components) - set(text_2_image_kwargs.keys())
        )
YiYi Xu's avatar
YiYi Xu committed
564
565
566

        if len(missing_modules) > 0:
            raise ValueError(
567
                f"Pipeline {text_2_image_cls} expected {expected_modules}, but only {set(list(passed_class_obj.keys()) + list(original_class_obj.keys()))} were passed"
YiYi Xu's avatar
YiYi Xu committed
568
569
570
571
572
573
574
575
576
577
578
579
            )

        model = text_2_image_cls(**text_2_image_kwargs)
        model.register_to_config(_name_or_path=pretrained_model_name_or_path)
        model.register_to_config(**unused_original_config)

        return model


class AutoPipelineForImage2Image(ConfigMixin):
    r"""

580
581
582
    [`AutoPipelineForImage2Image`] is a generic pipeline class that instantiates an image-to-image pipeline class. The
    specific underlying pipeline class is automatically selected from either the
    [`~AutoPipelineForImage2Image.from_pretrained`] or [`~AutoPipelineForImage2Image.from_pipe`] methods.
YiYi Xu's avatar
YiYi Xu committed
583

584
    This class cannot be instantiated using `__init__()` (throws an error).
YiYi Xu's avatar
YiYi Xu committed
585
586
587
588
589
590
591

    Class attributes:

        - **config_name** (`str`) -- The configuration filename that stores the class and module names of all the
          diffusion pipeline's components.

    """
592

YiYi Xu's avatar
YiYi Xu committed
593
594
595
596
597
598
599
600
601
602
    config_name = "model_index.json"

    def __init__(self, *args, **kwargs):
        raise EnvironmentError(
            f"{self.__class__.__name__} is designed to be instantiated "
            f"using the `{self.__class__.__name__}.from_pretrained(pretrained_model_name_or_path)` or "
            f"`{self.__class__.__name__}.from_pipe(pipeline)` methods."
        )

    @classmethod
603
    @validate_hf_hub_args
YiYi Xu's avatar
YiYi Xu committed
604
605
606
607
608
609
610
611
612
613
    def from_pretrained(cls, pretrained_model_or_path, **kwargs):
        r"""
        Instantiates a image-to-image Pytorch diffusion pipeline from pretrained pipeline weight.

        The from_pretrained() method takes care of returning the correct pipeline class instance by:
            1. Detect the pipeline class of the pretrained_model_or_path based on the _class_name property of its
               config object
            2. Find the image-to-image pipeline linked to the pipeline class using pattern matching on pipeline class
               name.

614
615
        If a `controlnet` argument is passed, it will instantiate a [`StableDiffusionControlNetImg2ImgPipeline`]
        object.
YiYi Xu's avatar
YiYi Xu committed
616
617
618
619
620
621

        The pipeline is set in evaluation mode (`model.eval()`) by default.

        If you get the error message below, you need to finetune the weights for your downstream task:

        ```
622
        Some weights of UNet2DConditionModel were not initialized from the model checkpoint at stable-diffusion-v1-5/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
YiYi Xu's avatar
YiYi Xu committed
623
624
625
626
627
        - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
        You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
        ```

        Parameters:
628
            pretrained_model_or_path (`str` or `os.PathLike`, *optional*):
YiYi Xu's avatar
YiYi Xu committed
629
630
631
632
633
634
635
636
                Can be either:

                    - A string, the *repo id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline
                      hosted on the Hub.
                    - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights
                      saved using
                    [`~DiffusionPipeline.save_pretrained`].
            torch_dtype (`str` or `torch.dtype`, *optional*):
637
                Override the default `torch.dtype` and load the model with another dtype.
YiYi Xu's avatar
YiYi Xu committed
638
639
640
641
642
643
            force_download (`bool`, *optional*, defaults to `False`):
                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                cached versions if they exist.
            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
644

YiYi Xu's avatar
YiYi Xu committed
645
646
647
648
649
650
651
652
            proxies (`Dict[str, str]`, *optional*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            output_loading_info(`bool`, *optional*, defaults to `False`):
                Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
            local_files_only (`bool`, *optional*, defaults to `False`):
                Whether to only load local model weights and configuration files or not. If set to `True`, the model
                won't be downloaded from the Hub.
653
            token (`str` or *bool*, *optional*):
YiYi Xu's avatar
YiYi Xu committed
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                `diffusers-cli login` (stored in `~/.huggingface`) is used.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            custom_revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, or a commit id similar to
                `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a
                custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub.
            mirror (`str`, *optional*):
                Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not
                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
                information.
            device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
                A map that specifies where each submodule should go. It doesn’t need to be defined for each
                parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the
                same device.

                Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
                more information about each option see [designing a device
                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
            max_memory (`Dict`, *optional*):
                A dictionary device identifier for the maximum memory. Will default to the maximum memory available for
                each GPU and the available CPU RAM if unset.
            offload_folder (`str` or `os.PathLike`, *optional*):
                The path to offload weights if device_map contains the value `"disk"`.
            offload_state_dict (`bool`, *optional*):
                If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if
                the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True`
                when there is some disk offload.
            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
                argument to `True` will raise an error.
            use_safetensors (`bool`, *optional*, defaults to `None`):
                If set to `None`, the safetensors weights are downloaded if they're available **and** if the
                safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
                weights. If set to `False`, safetensors weights are not loaded.
            kwargs (remaining dictionary of keyword arguments, *optional*):
                Can be used to overwrite load and saveable variables (the pipeline components of the specific pipeline
                class). The overwritten components are passed directly to the pipelines `__init__` method. See example
                below for more information.
            variant (`str`, *optional*):
                Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when
                loading `from_flax`.

Steven Liu's avatar
Steven Liu committed
701
702
        > [!TIP] > To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in
        with `hf > auth login`.
YiYi Xu's avatar
YiYi Xu committed
703
704
705
706

        Examples:

        ```py
707
        >>> from diffusers import AutoPipelineForImage2Image
YiYi Xu's avatar
YiYi Xu committed
708

709
        >>> pipeline = AutoPipelineForImage2Image.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5")
710
        >>> image = pipeline(prompt, image).images[0]
YiYi Xu's avatar
YiYi Xu committed
711
712
        ```
        """
713
        cache_dir = kwargs.pop("cache_dir", None)
714
715
        force_download = kwargs.pop("force_download", False)
        proxies = kwargs.pop("proxies", None)
716
        token = kwargs.pop("token", None)
717
718
719
720
721
722
723
        local_files_only = kwargs.pop("local_files_only", False)
        revision = kwargs.pop("revision", None)

        load_config_kwargs = {
            "cache_dir": cache_dir,
            "force_download": force_download,
            "proxies": proxies,
724
            "token": token,
725
726
727
728
729
            "local_files_only": local_files_only,
            "revision": revision,
        }

        config = cls.load_config(pretrained_model_or_path, **load_config_kwargs)
YiYi Xu's avatar
YiYi Xu committed
730
731
        orig_class_name = config["_class_name"]

732
733
        # the `orig_class_name` can be:
        # `- *Pipeline` (for regular text-to-image checkpoint)
734
        #  - `*ControlPipeline` (for Flux tools specific checkpoint)
735
        # `- *Img2ImgPipeline` (for refiner checkpoint)
736
737
738
739
740
741
        if "Img2Img" in orig_class_name:
            to_replace = "Img2ImgPipeline"
        elif "ControlPipeline" in orig_class_name:
            to_replace = "ControlPipeline"
        else:
            to_replace = "Pipeline"
742

YiYi Xu's avatar
YiYi Xu committed
743
        if "controlnet" in kwargs:
744
745
746
747
            if isinstance(kwargs["controlnet"], ControlNetUnionModel):
                orig_class_name = orig_class_name.replace(to_replace, "ControlNetUnion" + to_replace)
            else:
                orig_class_name = orig_class_name.replace(to_replace, "ControlNet" + to_replace)
YiYi Xu's avatar
YiYi Xu committed
748
749
750
        if "enable_pag" in kwargs:
            enable_pag = kwargs.pop("enable_pag")
            if enable_pag:
751
                orig_class_name = orig_class_name.replace(to_replace, "PAG" + to_replace)
YiYi Xu's avatar
YiYi Xu committed
752

753
754
755
        if to_replace == "ControlPipeline":
            orig_class_name = orig_class_name.replace(to_replace, "ControlImg2ImgPipeline")

YiYi Xu's avatar
YiYi Xu committed
756
757
        image_2_image_cls = _get_task_class(AUTO_IMAGE2IMAGE_PIPELINES_MAPPING, orig_class_name)

758
        kwargs = {**load_config_kwargs, **kwargs}
YiYi Xu's avatar
YiYi Xu committed
759
760
761
762
763
764
765
766
767
768
769
        return image_2_image_cls.from_pretrained(pretrained_model_or_path, **kwargs)

    @classmethod
    def from_pipe(cls, pipeline, **kwargs):
        r"""
        Instantiates a image-to-image Pytorch diffusion pipeline from another instantiated diffusion pipeline class.

        The from_pipe() method takes care of returning the correct pipeline class instance by finding the
        image-to-image pipeline linked to the pipeline class using pattern matching on pipeline class name.

        All the modules the pipeline contains will be used to initialize the new pipeline without reallocating
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
770
        additional memory.
YiYi Xu's avatar
YiYi Xu committed
771
772
773
774
775
776
777
778
779
780

        The pipeline is set in evaluation mode (`model.eval()`) by default.

        Parameters:
            pipeline (`DiffusionPipeline`):
                an instantiated `DiffusionPipeline` object

        Examples:

        ```py
781
        >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForImage2Image
YiYi Xu's avatar
YiYi Xu committed
782
783

        >>> pipe_t2i = AutoPipelineForText2Image.from_pretrained(
784
        ...     "stable-diffusion-v1-5/stable-diffusion-v1-5", requires_safety_checker=False
YiYi Xu's avatar
YiYi Xu committed
785
786
        ... )

787
788
        >>> pipe_i2i = AutoPipelineForImage2Image.from_pipe(pipe_t2i)
        >>> image = pipe_i2i(prompt, image).images[0]
YiYi Xu's avatar
YiYi Xu committed
789
790
791
792
793
794
795
796
797
        ```
        """

        original_config = dict(pipeline.config)
        original_cls_name = pipeline.__class__.__name__

        # derive the pipeline class to instantiate
        image_2_image_cls = _get_task_class(AUTO_IMAGE2IMAGE_PIPELINES_MAPPING, original_cls_name)

798
799
        if "controlnet" in kwargs:
            if kwargs["controlnet"] is not None:
YiYi Xu's avatar
YiYi Xu committed
800
801
802
                to_replace = "Img2ImgPipeline"
                if "PAG" in image_2_image_cls.__name__:
                    to_replace = "PAG" + to_replace
803
804
                image_2_image_cls = _get_task_class(
                    AUTO_IMAGE2IMAGE_PIPELINES_MAPPING,
805
                    image_2_image_cls.__name__.replace("ControlNet", "").replace(
YiYi Xu's avatar
YiYi Xu committed
806
                        to_replace, "ControlNet" + to_replace
807
                    ),
808
809
810
811
                )
            else:
                image_2_image_cls = _get_task_class(
                    AUTO_IMAGE2IMAGE_PIPELINES_MAPPING,
YiYi Xu's avatar
YiYi Xu committed
812
813
814
815
816
817
818
819
820
821
822
823
824
825
                    image_2_image_cls.__name__.replace("ControlNet", ""),
                )

        if "enable_pag" in kwargs:
            enable_pag = kwargs.pop("enable_pag")
            if enable_pag:
                image_2_image_cls = _get_task_class(
                    AUTO_IMAGE2IMAGE_PIPELINES_MAPPING,
                    image_2_image_cls.__name__.replace("PAG", "").replace("Img2ImgPipeline", "PAGImg2ImgPipeline"),
                )
            else:
                image_2_image_cls = _get_task_class(
                    AUTO_IMAGE2IMAGE_PIPELINES_MAPPING,
                    image_2_image_cls.__name__.replace("PAG", ""),
826
827
                )

YiYi Xu's avatar
YiYi Xu committed
828
        # define expected module and optional kwargs given the pipeline signature
829
        expected_modules, optional_kwargs = image_2_image_cls._get_signature_keys(image_2_image_cls)
YiYi Xu's avatar
YiYi Xu committed
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867

        pretrained_model_name_or_path = original_config.pop("_name_or_path", None)

        # allow users pass modules in `kwargs` to override the original pipeline's components
        passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs}
        original_class_obj = {
            k: pipeline.components[k]
            for k, v in pipeline.components.items()
            if k in expected_modules and k not in passed_class_obj
        }

        # allow users pass optional kwargs to override the original pipelines config attribute
        passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs}
        original_pipe_kwargs = {
            k: original_config[k]
            for k, v in original_config.items()
            if k in optional_kwargs and k not in passed_pipe_kwargs
        }

        # config attribute that were not expected by original pipeline is stored as its private attribute
        # we will pass them as optional arguments if they can be accepted by the pipeline
        additional_pipe_kwargs = [
            k[1:]
            for k in original_config.keys()
            if k.startswith("_") and k[1:] in optional_kwargs and k[1:] not in passed_pipe_kwargs
        ]
        for k in additional_pipe_kwargs:
            original_pipe_kwargs[k] = original_config.pop(f"_{k}")

        image_2_image_kwargs = {**passed_class_obj, **original_class_obj, **passed_pipe_kwargs, **original_pipe_kwargs}

        # store unused config as private attribute
        unused_original_config = {
            f"{'' if k.startswith('_') else '_'}{k}": original_config[k]
            for k, v in original_config.items()
            if k not in image_2_image_kwargs
        }

868
869
870
        missing_modules = (
            set(expected_modules) - set(image_2_image_cls._optional_components) - set(image_2_image_kwargs.keys())
        )
YiYi Xu's avatar
YiYi Xu committed
871
872
873

        if len(missing_modules) > 0:
            raise ValueError(
874
                f"Pipeline {image_2_image_cls} expected {expected_modules}, but only {set(list(passed_class_obj.keys()) + list(original_class_obj.keys()))} were passed"
YiYi Xu's avatar
YiYi Xu committed
875
876
877
878
879
880
881
882
883
884
885
886
            )

        model = image_2_image_cls(**image_2_image_kwargs)
        model.register_to_config(_name_or_path=pretrained_model_name_or_path)
        model.register_to_config(**unused_original_config)

        return model


class AutoPipelineForInpainting(ConfigMixin):
    r"""

887
888
889
    [`AutoPipelineForInpainting`] is a generic pipeline class that instantiates an inpainting pipeline class. The
    specific underlying pipeline class is automatically selected from either the
    [`~AutoPipelineForInpainting.from_pretrained`] or [`~AutoPipelineForInpainting.from_pipe`] methods.
YiYi Xu's avatar
YiYi Xu committed
890

891
    This class cannot be instantiated using `__init__()` (throws an error).
YiYi Xu's avatar
YiYi Xu committed
892
893
894
895
896
897
898

    Class attributes:

        - **config_name** (`str`) -- The configuration filename that stores the class and module names of all the
          diffusion pipeline's components.

    """
899

YiYi Xu's avatar
YiYi Xu committed
900
901
902
903
904
905
906
907
908
909
    config_name = "model_index.json"

    def __init__(self, *args, **kwargs):
        raise EnvironmentError(
            f"{self.__class__.__name__} is designed to be instantiated "
            f"using the `{self.__class__.__name__}.from_pretrained(pretrained_model_name_or_path)` or "
            f"`{self.__class__.__name__}.from_pipe(pipeline)` methods."
        )

    @classmethod
910
    @validate_hf_hub_args
YiYi Xu's avatar
YiYi Xu committed
911
912
913
914
915
916
917
918
919
    def from_pretrained(cls, pretrained_model_or_path, **kwargs):
        r"""
        Instantiates a inpainting Pytorch diffusion pipeline from pretrained pipeline weight.

        The from_pretrained() method takes care of returning the correct pipeline class instance by:
            1. Detect the pipeline class of the pretrained_model_or_path based on the _class_name property of its
               config object
            2. Find the inpainting pipeline linked to the pipeline class using pattern matching on pipeline class name.

920
921
        If a `controlnet` argument is passed, it will instantiate a [`StableDiffusionControlNetInpaintPipeline`]
        object.
YiYi Xu's avatar
YiYi Xu committed
922
923
924
925
926
927

        The pipeline is set in evaluation mode (`model.eval()`) by default.

        If you get the error message below, you need to finetune the weights for your downstream task:

        ```
928
        Some weights of UNet2DConditionModel were not initialized from the model checkpoint at stable-diffusion-v1-5/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
YiYi Xu's avatar
YiYi Xu committed
929
930
931
932
933
        - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
        You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
        ```

        Parameters:
934
            pretrained_model_or_path (`str` or `os.PathLike`, *optional*):
YiYi Xu's avatar
YiYi Xu committed
935
936
937
938
939
940
941
942
                Can be either:

                    - A string, the *repo id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline
                      hosted on the Hub.
                    - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights
                      saved using
                    [`~DiffusionPipeline.save_pretrained`].
            torch_dtype (`str` or `torch.dtype`, *optional*):
943
                Override the default `torch.dtype` and load the model with another dtype.
YiYi Xu's avatar
YiYi Xu committed
944
945
946
947
948
949
            force_download (`bool`, *optional*, defaults to `False`):
                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                cached versions if they exist.
            cache_dir (`Union[str, os.PathLike]`, *optional*):
                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
                is not used.
950

YiYi Xu's avatar
YiYi Xu committed
951
952
953
954
955
956
957
958
            proxies (`Dict[str, str]`, *optional*):
                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
            output_loading_info(`bool`, *optional*, defaults to `False`):
                Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
            local_files_only (`bool`, *optional*, defaults to `False`):
                Whether to only load local model weights and configuration files or not. If set to `True`, the model
                won't be downloaded from the Hub.
959
            token (`str` or *bool*, *optional*):
YiYi Xu's avatar
YiYi Xu committed
960
961
962
963
964
965
966
967
968
969
970
971
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
1004
1005
1006
                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
                `diffusers-cli login` (stored in `~/.huggingface`) is used.
            revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                allowed by Git.
            custom_revision (`str`, *optional*, defaults to `"main"`):
                The specific model version to use. It can be a branch name, a tag name, or a commit id similar to
                `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a
                custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub.
            mirror (`str`, *optional*):
                Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not
                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
                information.
            device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
                A map that specifies where each submodule should go. It doesn’t need to be defined for each
                parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the
                same device.

                Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
                more information about each option see [designing a device
                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
            max_memory (`Dict`, *optional*):
                A dictionary device identifier for the maximum memory. Will default to the maximum memory available for
                each GPU and the available CPU RAM if unset.
            offload_folder (`str` or `os.PathLike`, *optional*):
                The path to offload weights if device_map contains the value `"disk"`.
            offload_state_dict (`bool`, *optional*):
                If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if
                the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True`
                when there is some disk offload.
            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
                argument to `True` will raise an error.
            use_safetensors (`bool`, *optional*, defaults to `None`):
                If set to `None`, the safetensors weights are downloaded if they're available **and** if the
                safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
                weights. If set to `False`, safetensors weights are not loaded.
            kwargs (remaining dictionary of keyword arguments, *optional*):
                Can be used to overwrite load and saveable variables (the pipeline components of the specific pipeline
                class). The overwritten components are passed directly to the pipelines `__init__` method. See example
                below for more information.
            variant (`str`, *optional*):
                Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when
                loading `from_flax`.

Steven Liu's avatar
Steven Liu committed
1007
1008
        > [!TIP] > To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in
        with `hf > auth login`.
YiYi Xu's avatar
YiYi Xu committed
1009
1010
1011
1012

        Examples:

        ```py
1013
        >>> from diffusers import AutoPipelineForInpainting
YiYi Xu's avatar
YiYi Xu committed
1014

1015
        >>> pipeline = AutoPipelineForInpainting.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5")
1016
        >>> image = pipeline(prompt, image=init_image, mask_image=mask_image).images[0]
YiYi Xu's avatar
YiYi Xu committed
1017
1018
        ```
        """
1019
        cache_dir = kwargs.pop("cache_dir", None)
1020
1021
        force_download = kwargs.pop("force_download", False)
        proxies = kwargs.pop("proxies", None)
1022
        token = kwargs.pop("token", None)
1023
1024
1025
1026
1027
1028
1029
        local_files_only = kwargs.pop("local_files_only", False)
        revision = kwargs.pop("revision", None)

        load_config_kwargs = {
            "cache_dir": cache_dir,
            "force_download": force_download,
            "proxies": proxies,
1030
            "token": token,
1031
1032
1033
1034
1035
            "local_files_only": local_files_only,
            "revision": revision,
        }

        config = cls.load_config(pretrained_model_or_path, **load_config_kwargs)
YiYi Xu's avatar
YiYi Xu committed
1036
1037
        orig_class_name = config["_class_name"]

1038
1039
        # The `orig_class_name`` can be:
        # `- *InpaintPipeline` (for inpaint-specific checkpoint)
1040
        #  - `*ControlPipeline` (for Flux tools specific checkpoint)
1041
        #  - or *Pipeline (for regular text-to-image checkpoint)
1042
1043
1044
1045
1046
1047
        if "Inpaint" in orig_class_name:
            to_replace = "InpaintPipeline"
        elif "ControlPipeline" in orig_class_name:
            to_replace = "ControlPipeline"
        else:
            to_replace = "Pipeline"
1048

YiYi Xu's avatar
YiYi Xu committed
1049
        if "controlnet" in kwargs:
1050
1051
1052
1053
            if isinstance(kwargs["controlnet"], ControlNetUnionModel):
                orig_class_name = orig_class_name.replace(to_replace, "ControlNetUnion" + to_replace)
            else:
                orig_class_name = orig_class_name.replace(to_replace, "ControlNet" + to_replace)
YiYi Xu's avatar
YiYi Xu committed
1054
1055
1056
        if "enable_pag" in kwargs:
            enable_pag = kwargs.pop("enable_pag")
            if enable_pag:
1057
                orig_class_name = orig_class_name.replace(to_replace, "PAG" + to_replace)
1058
1059
        if to_replace == "ControlPipeline":
            orig_class_name = orig_class_name.replace(to_replace, "ControlInpaintPipeline")
YiYi Xu's avatar
YiYi Xu committed
1060
1061
        inpainting_cls = _get_task_class(AUTO_INPAINT_PIPELINES_MAPPING, orig_class_name)

1062
        kwargs = {**load_config_kwargs, **kwargs}
YiYi Xu's avatar
YiYi Xu committed
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
        return inpainting_cls.from_pretrained(pretrained_model_or_path, **kwargs)

    @classmethod
    def from_pipe(cls, pipeline, **kwargs):
        r"""
        Instantiates a inpainting Pytorch diffusion pipeline from another instantiated diffusion pipeline class.

        The from_pipe() method takes care of returning the correct pipeline class instance by finding the inpainting
        pipeline linked to the pipeline class using pattern matching on pipeline class name.

        All the modules the pipeline class contain will be used to initialize the new pipeline without reallocating
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
1074
        additional memory.
YiYi Xu's avatar
YiYi Xu committed
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084

        The pipeline is set in evaluation mode (`model.eval()`) by default.

        Parameters:
            pipeline (`DiffusionPipeline`):
                an instantiated `DiffusionPipeline` object

        Examples:

        ```py
1085
        >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForInpainting
YiYi Xu's avatar
YiYi Xu committed
1086
1087
1088
1089
1090
1091

        >>> pipe_t2i = AutoPipelineForText2Image.from_pretrained(
        ...     "DeepFloyd/IF-I-XL-v1.0", requires_safety_checker=False
        ... )

        >>> pipe_inpaint = AutoPipelineForInpainting.from_pipe(pipe_t2i)
1092
        >>> image = pipe_inpaint(prompt, image=init_image, mask_image=mask_image).images[0]
YiYi Xu's avatar
YiYi Xu committed
1093
1094
1095
1096
1097
1098
1099
1100
        ```
        """
        original_config = dict(pipeline.config)
        original_cls_name = pipeline.__class__.__name__

        # derive the pipeline class to instantiate
        inpainting_cls = _get_task_class(AUTO_INPAINT_PIPELINES_MAPPING, original_cls_name)

1101
1102
1103
1104
        if "controlnet" in kwargs:
            if kwargs["controlnet"] is not None:
                inpainting_cls = _get_task_class(
                    AUTO_INPAINT_PIPELINES_MAPPING,
1105
1106
1107
                    inpainting_cls.__name__.replace("ControlNet", "").replace(
                        "InpaintPipeline", "ControlNetInpaintPipeline"
                    ),
1108
1109
1110
1111
1112
1113
1114
                )
            else:
                inpainting_cls = _get_task_class(
                    AUTO_INPAINT_PIPELINES_MAPPING,
                    inpainting_cls.__name__.replace("ControlNetInpaintPipeline", "InpaintPipeline"),
                )

YiYi Xu's avatar
YiYi Xu committed
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
        if "enable_pag" in kwargs:
            enable_pag = kwargs.pop("enable_pag")
            if enable_pag:
                inpainting_cls = _get_task_class(
                    AUTO_INPAINT_PIPELINES_MAPPING,
                    inpainting_cls.__name__.replace("PAG", "").replace("InpaintPipeline", "PAGInpaintPipeline"),
                )
            else:
                inpainting_cls = _get_task_class(
                    AUTO_INPAINT_PIPELINES_MAPPING,
                    inpainting_cls.__name__.replace("PAGInpaintPipeline", "InpaintPipeline"),
                )

YiYi Xu's avatar
YiYi Xu committed
1128
        # define expected module and optional kwargs given the pipeline signature
1129
        expected_modules, optional_kwargs = inpainting_cls._get_signature_keys(inpainting_cls)
YiYi Xu's avatar
YiYi Xu committed
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167

        pretrained_model_name_or_path = original_config.pop("_name_or_path", None)

        # allow users pass modules in `kwargs` to override the original pipeline's components
        passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs}
        original_class_obj = {
            k: pipeline.components[k]
            for k, v in pipeline.components.items()
            if k in expected_modules and k not in passed_class_obj
        }

        # allow users pass optional kwargs to override the original pipelines config attribute
        passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs}
        original_pipe_kwargs = {
            k: original_config[k]
            for k, v in original_config.items()
            if k in optional_kwargs and k not in passed_pipe_kwargs
        }

        # config that were not expected by original pipeline is stored as private attribute
        # we will pass them as optional arguments if they can be accepted by the pipeline
        additional_pipe_kwargs = [
            k[1:]
            for k in original_config.keys()
            if k.startswith("_") and k[1:] in optional_kwargs and k[1:] not in passed_pipe_kwargs
        ]
        for k in additional_pipe_kwargs:
            original_pipe_kwargs[k] = original_config.pop(f"_{k}")

        inpainting_kwargs = {**passed_class_obj, **original_class_obj, **passed_pipe_kwargs, **original_pipe_kwargs}

        # store unused config as private attribute
        unused_original_config = {
            f"{'' if k.startswith('_') else '_'}{k}": original_config[k]
            for k, v in original_config.items()
            if k not in inpainting_kwargs
        }

1168
1169
1170
        missing_modules = (
            set(expected_modules) - set(inpainting_cls._optional_components) - set(inpainting_kwargs.keys())
        )
YiYi Xu's avatar
YiYi Xu committed
1171
1172
1173

        if len(missing_modules) > 0:
            raise ValueError(
1174
                f"Pipeline {inpainting_cls} expected {expected_modules}, but only {set(list(passed_class_obj.keys()) + list(original_class_obj.keys()))} were passed"
YiYi Xu's avatar
YiYi Xu committed
1175
1176
1177
1178
1179
1180
1181
            )

        model = inpainting_cls(**inpainting_kwargs)
        model.register_to_config(_name_or_path=pretrained_model_name_or_path)
        model.register_to_config(**unused_original_config)

        return model