auto_pipeline.py 57 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
94
from .sana import SanaPipeline
95
from .stable_cascade import StableCascadeCombinedPipeline, StableCascadeDecoderPipeline
YiYi Xu's avatar
YiYi Xu committed
96
97
98
99
100
from .stable_diffusion import (
    StableDiffusionImg2ImgPipeline,
    StableDiffusionInpaintPipeline,
    StableDiffusionPipeline,
)
101
102
from .stable_diffusion_3 import (
    StableDiffusion3Img2ImgPipeline,
103
    StableDiffusion3InpaintPipeline,
104
105
    StableDiffusion3Pipeline,
)
YiYi Xu's avatar
YiYi Xu committed
106
107
108
109
110
from .stable_diffusion_xl import (
    StableDiffusionXLImg2ImgPipeline,
    StableDiffusionXLInpaintPipeline,
    StableDiffusionXLPipeline,
)
Kashif Rasul's avatar
Kashif Rasul committed
111
from .wuerstchen import WuerstchenCombinedPipeline, WuerstchenDecoderPipeline
YiYi Xu's avatar
YiYi Xu committed
112
113
114
115
116
117


AUTO_TEXT2IMAGE_PIPELINES_MAPPING = OrderedDict(
    [
        ("stable-diffusion", StableDiffusionPipeline),
        ("stable-diffusion-xl", StableDiffusionXLPipeline),
118
        ("stable-diffusion-3", StableDiffusion3Pipeline),
119
        ("stable-diffusion-3-pag", StableDiffusion3PAGPipeline),
YiYi Xu's avatar
YiYi Xu committed
120
        ("if", IFPipeline),
121
        ("hunyuan", HunyuanDiTPipeline),
122
        ("hunyuan-pag", HunyuanDiTPAGPipeline),
123
124
        ("kandinsky", KandinskyCombinedPipeline),
        ("kandinsky22", KandinskyV22CombinedPipeline),
125
        ("kandinsky3", Kandinsky3Pipeline),
YiYi Xu's avatar
YiYi Xu committed
126
127
        ("stable-diffusion-controlnet", StableDiffusionControlNetPipeline),
        ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetPipeline),
128
        ("stable-diffusion-xl-controlnet-union", StableDiffusionXLControlNetUnionPipeline),
129
        ("stable-diffusion-3-controlnet", StableDiffusion3ControlNetPipeline),
Kashif Rasul's avatar
Kashif Rasul committed
130
        ("wuerstchen", WuerstchenCombinedPipeline),
131
        ("cascade", StableCascadeCombinedPipeline),
132
        ("lcm", LatentConsistencyModelPipeline),
133
134
        ("pixart-alpha", PixArtAlphaPipeline),
        ("pixart-sigma", PixArtSigmaPipeline),
135
136
        ("sana", SanaPipeline),
        ("sana-pag", SanaPAGPipeline),
137
        ("stable-diffusion-pag", StableDiffusionPAGPipeline),
138
        ("stable-diffusion-controlnet-pag", StableDiffusionControlNetPAGPipeline),
YiYi Xu's avatar
YiYi Xu committed
139
140
        ("stable-diffusion-xl-pag", StableDiffusionXLPAGPipeline),
        ("stable-diffusion-xl-controlnet-pag", StableDiffusionXLControlNetPAGPipeline),
141
        ("pixart-sigma-pag", PixArtSigmaPAGPipeline),
142
        ("auraflow", AuraFlowPipeline),
Sayak Paul's avatar
Sayak Paul committed
143
        ("flux", FluxPipeline),
144
        ("flux-control", FluxControlPipeline),
145
        ("flux-controlnet", FluxControlNetPipeline),
jlonge4's avatar
jlonge4 committed
146
        ("flux-kontext", FluxKontextPipeline),
147
148
        ("lumina", LuminaPipeline),
        ("lumina2", Lumina2Pipeline),
Edna's avatar
Edna committed
149
        ("chroma", ChromaPipeline),
Yuxuan.Zhang's avatar
Yuxuan.Zhang committed
150
        ("cogview3", CogView3PlusPipeline),
151
        ("cogview4", CogView4Pipeline),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
152
        ("cogview4-control", CogView4ControlPipeline),
YiYi Xu's avatar
YiYi Xu committed
153
154
155
156
157
158
159
    ]
)

AUTO_IMAGE2IMAGE_PIPELINES_MAPPING = OrderedDict(
    [
        ("stable-diffusion", StableDiffusionImg2ImgPipeline),
        ("stable-diffusion-xl", StableDiffusionXLImg2ImgPipeline),
160
        ("stable-diffusion-3", StableDiffusion3Img2ImgPipeline),
161
        ("stable-diffusion-3-pag", StableDiffusion3PAGImg2ImgPipeline),
YiYi Xu's avatar
YiYi Xu committed
162
        ("if", IFImg2ImgPipeline),
163
164
        ("kandinsky", KandinskyImg2ImgCombinedPipeline),
        ("kandinsky22", KandinskyV22Img2ImgCombinedPipeline),
165
        ("kandinsky3", Kandinsky3Img2ImgPipeline),
YiYi Xu's avatar
YiYi Xu committed
166
        ("stable-diffusion-controlnet", StableDiffusionControlNetImg2ImgPipeline),
167
        ("stable-diffusion-pag", StableDiffusionPAGImg2ImgPipeline),
168
        ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetImg2ImgPipeline),
169
        ("stable-diffusion-xl-controlnet-union", StableDiffusionXLControlNetUnionImg2ImgPipeline),
YiYi Xu's avatar
YiYi Xu committed
170
        ("stable-diffusion-xl-pag", StableDiffusionXLPAGImg2ImgPipeline),
171
        ("stable-diffusion-xl-controlnet-pag", StableDiffusionXLControlNetPAGImg2ImgPipeline),
172
        ("lcm", LatentConsistencyModelImg2ImgPipeline),
173
        ("flux", FluxImg2ImgPipeline),
174
        ("flux-controlnet", FluxControlNetImg2ImgPipeline),
175
        ("flux-control", FluxControlImg2ImgPipeline),
jlonge4's avatar
jlonge4 committed
176
        ("flux-kontext", FluxKontextPipeline),
YiYi Xu's avatar
YiYi Xu committed
177
178
179
180
181
182
183
    ]
)

AUTO_INPAINT_PIPELINES_MAPPING = OrderedDict(
    [
        ("stable-diffusion", StableDiffusionInpaintPipeline),
        ("stable-diffusion-xl", StableDiffusionXLInpaintPipeline),
184
        ("stable-diffusion-3", StableDiffusion3InpaintPipeline),
YiYi Xu's avatar
YiYi Xu committed
185
        ("if", IFInpaintingPipeline),
186
187
188
        ("kandinsky", KandinskyInpaintCombinedPipeline),
        ("kandinsky22", KandinskyV22InpaintCombinedPipeline),
        ("stable-diffusion-controlnet", StableDiffusionControlNetInpaintPipeline),
189
        ("stable-diffusion-controlnet-pag", StableDiffusionControlNetPAGInpaintPipeline),
190
        ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetInpaintPipeline),
191
        ("stable-diffusion-xl-controlnet-union", StableDiffusionXLControlNetUnionInpaintPipeline),
192
        ("stable-diffusion-3-controlnet", StableDiffusion3ControlNetInpaintingPipeline),
YiYi Xu's avatar
YiYi Xu committed
193
        ("stable-diffusion-xl-pag", StableDiffusionXLPAGInpaintPipeline),
194
        ("flux", FluxInpaintPipeline),
195
        ("flux-controlnet", FluxControlNetInpaintPipeline),
196
        ("flux-control", FluxControlInpaintPipeline),
197
        ("stable-diffusion-pag", StableDiffusionPAGInpaintPipeline),
198
199
200
201
202
203
204
    ]
)

_AUTO_TEXT2IMAGE_DECODER_PIPELINES_MAPPING = OrderedDict(
    [
        ("kandinsky", KandinskyPipeline),
        ("kandinsky22", KandinskyV22Pipeline),
Kashif Rasul's avatar
Kashif Rasul committed
205
        ("wuerstchen", WuerstchenDecoderPipeline),
206
        ("cascade", StableCascadeDecoderPipeline),
207
208
209
210
211
212
213
214
215
216
    ]
)
_AUTO_IMAGE2IMAGE_DECODER_PIPELINES_MAPPING = OrderedDict(
    [
        ("kandinsky", KandinskyImg2ImgPipeline),
        ("kandinsky22", KandinskyV22Img2ImgPipeline),
    ]
)
_AUTO_INPAINT_DECODER_PIPELINES_MAPPING = OrderedDict(
    [
YiYi Xu's avatar
YiYi Xu committed
217
218
219
220
221
        ("kandinsky", KandinskyInpaintPipeline),
        ("kandinsky22", KandinskyV22InpaintPipeline),
    ]
)

222
if is_sentencepiece_available():
223
    from .kolors import KolorsImg2ImgPipeline, KolorsPipeline
Álvaro Somoza's avatar
Álvaro Somoza committed
224
    from .pag import KolorsPAGPipeline
225
226

    AUTO_TEXT2IMAGE_PIPELINES_MAPPING["kolors"] = KolorsPipeline
Álvaro Somoza's avatar
Álvaro Somoza committed
227
    AUTO_TEXT2IMAGE_PIPELINES_MAPPING["kolors-pag"] = KolorsPAGPipeline
228
    AUTO_IMAGE2IMAGE_PIPELINES_MAPPING["kolors"] = KolorsImg2ImgPipeline
229

YiYi Xu's avatar
YiYi Xu committed
230
231
232
233
SUPPORTED_TASKS_MAPPINGS = [
    AUTO_TEXT2IMAGE_PIPELINES_MAPPING,
    AUTO_IMAGE2IMAGE_PIPELINES_MAPPING,
    AUTO_INPAINT_PIPELINES_MAPPING,
234
235
236
    _AUTO_TEXT2IMAGE_DECODER_PIPELINES_MAPPING,
    _AUTO_IMAGE2IMAGE_DECODER_PIPELINES_MAPPING,
    _AUTO_INPAINT_DECODER_PIPELINES_MAPPING,
YiYi Xu's avatar
YiYi Xu committed
237
238
239
]


240
241
242
243
244
245
246
247
248
249
250
251
252
253
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
254
255
256
257
258
259
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
260

YiYi Xu's avatar
YiYi Xu committed
261
262
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
263
264
265
266
267

    if model_name is not None:
        task_class = mapping.get(model_name, None)
        if task_class is not None:
            return task_class
268
269
270

    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
271
272
273
274
275


class AutoPipelineForText2Image(ConfigMixin):
    r"""

276
277
278
    [`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
279

280
    This class cannot be instantiated using `__init__()` (throws an error).
YiYi Xu's avatar
YiYi Xu committed
281
282
283
284
285
286
287

    Class attributes:

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

    """
288

YiYi Xu's avatar
YiYi Xu committed
289
290
291
292
293
294
295
296
297
298
    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
299
    @validate_hf_hub_args
YiYi Xu's avatar
YiYi Xu committed
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
    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:

        ```
317
        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
318
319
320
321
322
        - 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:
323
            pretrained_model_or_path (`str` or `os.PathLike`, *optional*):
YiYi Xu's avatar
YiYi Xu committed
324
325
326
327
328
329
330
                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`].
331
332
            torch_dtype (`torch.dtype`, *optional*):
                Override the default `torch.dtype` and load the model with another dtype.
YiYi Xu's avatar
YiYi Xu committed
333
334
335
336
337
338
            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.
339

YiYi Xu's avatar
YiYi Xu committed
340
341
342
343
344
345
346
347
            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.
348
            token (`str` or *bool*, *optional*):
YiYi Xu's avatar
YiYi Xu committed
349
350
351
352
353
354
355
356
357
358
359
360
361
362
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
                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`.

        <Tip>

398
399
        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
400
401
402
403
404
405

        </Tip>

        Examples:

        ```py
406
        >>> from diffusers import AutoPipelineForText2Image
YiYi Xu's avatar
YiYi Xu committed
407

408
        >>> pipeline = AutoPipelineForText2Image.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5")
409
        >>> image = pipeline(prompt).images[0]
YiYi Xu's avatar
YiYi Xu committed
410
411
        ```
        """
412
        cache_dir = kwargs.pop("cache_dir", None)
413
414
        force_download = kwargs.pop("force_download", False)
        proxies = kwargs.pop("proxies", None)
415
        token = kwargs.pop("token", None)
416
417
418
419
420
421
422
        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,
423
            "token": token,
424
425
426
427
428
            "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
429
        orig_class_name = config["_class_name"]
430
431
432
433
        if "ControlPipeline" in orig_class_name:
            to_replace = "ControlPipeline"
        else:
            to_replace = "Pipeline"
YiYi Xu's avatar
YiYi Xu committed
434
435

        if "controlnet" in kwargs:
436
            if isinstance(kwargs["controlnet"], ControlNetUnionModel):
437
                orig_class_name = config["_class_name"].replace(to_replace, "ControlNetUnionPipeline")
438
            else:
439
                orig_class_name = config["_class_name"].replace(to_replace, "ControlNetPipeline")
YiYi Xu's avatar
YiYi Xu committed
440
441
442
        if "enable_pag" in kwargs:
            enable_pag = kwargs.pop("enable_pag")
            if enable_pag:
443
                orig_class_name = orig_class_name.replace(to_replace, "PAGPipeline")
YiYi Xu's avatar
YiYi Xu committed
444
445
446

        text_2_image_cls = _get_task_class(AUTO_TEXT2IMAGE_PIPELINES_MAPPING, orig_class_name)

447
        kwargs = {**load_config_kwargs, **kwargs}
YiYi Xu's avatar
YiYi Xu committed
448
449
450
451
452
453
454
455
456
457
458
        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
459
        additional memory.
YiYi Xu's avatar
YiYi Xu committed
460
461
462
463
464
465
466
467

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

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

        ```py
468
        >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForImage2Image
YiYi Xu's avatar
YiYi Xu committed
469
470

        >>> pipe_i2i = AutoPipelineForImage2Image.from_pretrained(
471
        ...     "stable-diffusion-v1-5/stable-diffusion-v1-5", requires_safety_checker=False
YiYi Xu's avatar
YiYi Xu committed
472
473
        ... )

474
475
        >>> pipe_t2i = AutoPipelineForText2Image.from_pipe(pipe_i2i)
        >>> image = pipe_t2i(prompt).images[0]
YiYi Xu's avatar
YiYi Xu committed
476
477
478
479
480
481
482
483
484
        ```
        """

        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)

485
486
        if "controlnet" in kwargs:
            if kwargs["controlnet"] is not None:
YiYi Xu's avatar
YiYi Xu committed
487
                to_replace = "PAGPipeline" if "PAG" in text_2_image_cls.__name__ else "Pipeline"
488
489
                text_2_image_cls = _get_task_class(
                    AUTO_TEXT2IMAGE_PIPELINES_MAPPING,
YiYi Xu's avatar
YiYi Xu committed
490
                    text_2_image_cls.__name__.replace("ControlNet", "").replace(to_replace, "ControlNet" + to_replace),
491
492
493
494
                )
            else:
                text_2_image_cls = _get_task_class(
                    AUTO_TEXT2IMAGE_PIPELINES_MAPPING,
YiYi Xu's avatar
YiYi Xu committed
495
496
497
498
499
500
501
502
503
504
505
506
507
508
                    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", ""),
509
510
                )

YiYi Xu's avatar
YiYi Xu committed
511
        # define expected module and optional kwargs given the pipeline signature
512
        expected_modules, optional_kwargs = text_2_image_cls._get_signature_keys(text_2_image_cls)
YiYi Xu's avatar
YiYi Xu committed
513
514
515
516
517
518
519
520
521
522
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

        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
        }

551
552
553
        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
554
555
556

        if len(missing_modules) > 0:
            raise ValueError(
557
                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
558
559
560
561
562
563
564
565
566
567
568
569
            )

        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"""

570
571
572
    [`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
573

574
    This class cannot be instantiated using `__init__()` (throws an error).
YiYi Xu's avatar
YiYi Xu committed
575
576
577
578
579
580
581

    Class attributes:

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

    """
582

YiYi Xu's avatar
YiYi Xu committed
583
584
585
586
587
588
589
590
591
592
    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
593
    @validate_hf_hub_args
YiYi Xu's avatar
YiYi Xu committed
594
595
596
597
598
599
600
601
602
603
    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.

604
605
        If a `controlnet` argument is passed, it will instantiate a [`StableDiffusionControlNetImg2ImgPipeline`]
        object.
YiYi Xu's avatar
YiYi Xu committed
606
607
608
609
610
611

        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:

        ```
612
        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
613
614
615
616
617
        - 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:
618
            pretrained_model_or_path (`str` or `os.PathLike`, *optional*):
YiYi Xu's avatar
YiYi Xu committed
619
620
621
622
623
624
625
626
                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*):
627
                Override the default `torch.dtype` and load the model with another dtype.
YiYi Xu's avatar
YiYi Xu committed
628
629
630
631
632
633
            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.
634

YiYi Xu's avatar
YiYi Xu committed
635
636
637
638
639
640
641
642
            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.
643
            token (`str` or *bool*, *optional*):
YiYi Xu's avatar
YiYi Xu committed
644
645
646
647
648
649
650
651
652
653
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
                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`.

        <Tip>

693
694
        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
695
696
697
698
699
700

        </Tip>

        Examples:

        ```py
701
        >>> from diffusers import AutoPipelineForImage2Image
YiYi Xu's avatar
YiYi Xu committed
702

703
        >>> pipeline = AutoPipelineForImage2Image.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5")
704
        >>> image = pipeline(prompt, image).images[0]
YiYi Xu's avatar
YiYi Xu committed
705
706
        ```
        """
707
        cache_dir = kwargs.pop("cache_dir", None)
708
709
        force_download = kwargs.pop("force_download", False)
        proxies = kwargs.pop("proxies", None)
710
        token = kwargs.pop("token", None)
711
712
713
714
715
716
717
        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,
718
            "token": token,
719
720
721
722
723
            "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
724
725
        orig_class_name = config["_class_name"]

726
727
        # the `orig_class_name` can be:
        # `- *Pipeline` (for regular text-to-image checkpoint)
728
        #  - `*ControlPipeline` (for Flux tools specific checkpoint)
729
        # `- *Img2ImgPipeline` (for refiner checkpoint)
730
731
732
733
734
735
        if "Img2Img" in orig_class_name:
            to_replace = "Img2ImgPipeline"
        elif "ControlPipeline" in orig_class_name:
            to_replace = "ControlPipeline"
        else:
            to_replace = "Pipeline"
736

YiYi Xu's avatar
YiYi Xu committed
737
        if "controlnet" in kwargs:
738
739
740
741
            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
742
743
744
        if "enable_pag" in kwargs:
            enable_pag = kwargs.pop("enable_pag")
            if enable_pag:
745
                orig_class_name = orig_class_name.replace(to_replace, "PAG" + to_replace)
YiYi Xu's avatar
YiYi Xu committed
746

747
748
749
        if to_replace == "ControlPipeline":
            orig_class_name = orig_class_name.replace(to_replace, "ControlImg2ImgPipeline")

YiYi Xu's avatar
YiYi Xu committed
750
751
        image_2_image_cls = _get_task_class(AUTO_IMAGE2IMAGE_PIPELINES_MAPPING, orig_class_name)

752
        kwargs = {**load_config_kwargs, **kwargs}
YiYi Xu's avatar
YiYi Xu committed
753
754
755
756
757
758
759
760
761
762
763
        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
764
        additional memory.
YiYi Xu's avatar
YiYi Xu committed
765
766
767
768
769
770
771
772
773
774

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

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

        Examples:

        ```py
775
        >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForImage2Image
YiYi Xu's avatar
YiYi Xu committed
776
777

        >>> pipe_t2i = AutoPipelineForText2Image.from_pretrained(
778
        ...     "stable-diffusion-v1-5/stable-diffusion-v1-5", requires_safety_checker=False
YiYi Xu's avatar
YiYi Xu committed
779
780
        ... )

781
782
        >>> pipe_i2i = AutoPipelineForImage2Image.from_pipe(pipe_t2i)
        >>> image = pipe_i2i(prompt, image).images[0]
YiYi Xu's avatar
YiYi Xu committed
783
784
785
786
787
788
789
790
791
        ```
        """

        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)

792
793
        if "controlnet" in kwargs:
            if kwargs["controlnet"] is not None:
YiYi Xu's avatar
YiYi Xu committed
794
795
796
                to_replace = "Img2ImgPipeline"
                if "PAG" in image_2_image_cls.__name__:
                    to_replace = "PAG" + to_replace
797
798
                image_2_image_cls = _get_task_class(
                    AUTO_IMAGE2IMAGE_PIPELINES_MAPPING,
799
                    image_2_image_cls.__name__.replace("ControlNet", "").replace(
YiYi Xu's avatar
YiYi Xu committed
800
                        to_replace, "ControlNet" + to_replace
801
                    ),
802
803
804
805
                )
            else:
                image_2_image_cls = _get_task_class(
                    AUTO_IMAGE2IMAGE_PIPELINES_MAPPING,
YiYi Xu's avatar
YiYi Xu committed
806
807
808
809
810
811
812
813
814
815
816
817
818
819
                    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", ""),
820
821
                )

YiYi Xu's avatar
YiYi Xu committed
822
        # define expected module and optional kwargs given the pipeline signature
823
        expected_modules, optional_kwargs = image_2_image_cls._get_signature_keys(image_2_image_cls)
YiYi Xu's avatar
YiYi Xu committed
824
825
826
827
828
829
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

        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
        }

862
863
864
        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
865
866
867

        if len(missing_modules) > 0:
            raise ValueError(
868
                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
869
870
871
872
873
874
875
876
877
878
879
880
            )

        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"""

881
882
883
    [`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
884

885
    This class cannot be instantiated using `__init__()` (throws an error).
YiYi Xu's avatar
YiYi Xu committed
886
887
888
889
890
891
892

    Class attributes:

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

    """
893

YiYi Xu's avatar
YiYi Xu committed
894
895
896
897
898
899
900
901
902
903
    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
904
    @validate_hf_hub_args
YiYi Xu's avatar
YiYi Xu committed
905
906
907
908
909
910
911
912
913
    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.

914
915
        If a `controlnet` argument is passed, it will instantiate a [`StableDiffusionControlNetInpaintPipeline`]
        object.
YiYi Xu's avatar
YiYi Xu committed
916
917
918
919
920
921

        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:

        ```
922
        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
923
924
925
926
927
        - 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:
928
            pretrained_model_or_path (`str` or `os.PathLike`, *optional*):
YiYi Xu's avatar
YiYi Xu committed
929
930
931
932
933
934
935
936
                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*):
937
                Override the default `torch.dtype` and load the model with another dtype.
YiYi Xu's avatar
YiYi Xu committed
938
939
940
941
942
943
            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.
944

YiYi Xu's avatar
YiYi Xu committed
945
946
947
948
949
950
951
952
            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.
953
            token (`str` or *bool*, *optional*):
YiYi Xu's avatar
YiYi Xu committed
954
955
956
957
958
959
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
                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`.

        <Tip>

1003
1004
        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
1005
1006
1007
1008
1009
1010

        </Tip>

        Examples:

        ```py
1011
        >>> from diffusers import AutoPipelineForInpainting
YiYi Xu's avatar
YiYi Xu committed
1012

1013
        >>> pipeline = AutoPipelineForInpainting.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5")
1014
        >>> image = pipeline(prompt, image=init_image, mask_image=mask_image).images[0]
YiYi Xu's avatar
YiYi Xu committed
1015
1016
        ```
        """
1017
        cache_dir = kwargs.pop("cache_dir", None)
1018
1019
        force_download = kwargs.pop("force_download", False)
        proxies = kwargs.pop("proxies", None)
1020
        token = kwargs.pop("token", None)
1021
1022
1023
1024
1025
1026
1027
        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,
1028
            "token": token,
1029
1030
1031
1032
1033
            "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
1034
1035
        orig_class_name = config["_class_name"]

1036
1037
        # The `orig_class_name`` can be:
        # `- *InpaintPipeline` (for inpaint-specific checkpoint)
1038
        #  - `*ControlPipeline` (for Flux tools specific checkpoint)
1039
        #  - or *Pipeline (for regular text-to-image checkpoint)
1040
1041
1042
1043
1044
1045
        if "Inpaint" in orig_class_name:
            to_replace = "InpaintPipeline"
        elif "ControlPipeline" in orig_class_name:
            to_replace = "ControlPipeline"
        else:
            to_replace = "Pipeline"
1046

YiYi Xu's avatar
YiYi Xu committed
1047
        if "controlnet" in kwargs:
1048
1049
1050
1051
            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
1052
1053
1054
        if "enable_pag" in kwargs:
            enable_pag = kwargs.pop("enable_pag")
            if enable_pag:
1055
                orig_class_name = orig_class_name.replace(to_replace, "PAG" + to_replace)
1056
1057
        if to_replace == "ControlPipeline":
            orig_class_name = orig_class_name.replace(to_replace, "ControlInpaintPipeline")
YiYi Xu's avatar
YiYi Xu committed
1058
1059
        inpainting_cls = _get_task_class(AUTO_INPAINT_PIPELINES_MAPPING, orig_class_name)

1060
        kwargs = {**load_config_kwargs, **kwargs}
YiYi Xu's avatar
YiYi Xu committed
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
        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
1072
        additional memory.
YiYi Xu's avatar
YiYi Xu committed
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082

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

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

        Examples:

        ```py
1083
        >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForInpainting
YiYi Xu's avatar
YiYi Xu committed
1084
1085
1086
1087
1088
1089

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

        >>> pipe_inpaint = AutoPipelineForInpainting.from_pipe(pipe_t2i)
1090
        >>> image = pipe_inpaint(prompt, image=init_image, mask_image=mask_image).images[0]
YiYi Xu's avatar
YiYi Xu committed
1091
1092
1093
1094
1095
1096
1097
1098
        ```
        """
        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)

1099
1100
1101
1102
        if "controlnet" in kwargs:
            if kwargs["controlnet"] is not None:
                inpainting_cls = _get_task_class(
                    AUTO_INPAINT_PIPELINES_MAPPING,
1103
1104
1105
                    inpainting_cls.__name__.replace("ControlNet", "").replace(
                        "InpaintPipeline", "ControlNetInpaintPipeline"
                    ),
1106
1107
1108
1109
1110
1111
1112
                )
            else:
                inpainting_cls = _get_task_class(
                    AUTO_INPAINT_PIPELINES_MAPPING,
                    inpainting_cls.__name__.replace("ControlNetInpaintPipeline", "InpaintPipeline"),
                )

YiYi Xu's avatar
YiYi Xu committed
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
        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
1126
        # define expected module and optional kwargs given the pipeline signature
1127
        expected_modules, optional_kwargs = inpainting_cls._get_signature_keys(inpainting_cls)
YiYi Xu's avatar
YiYi Xu committed
1128
1129
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

        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
        }

1166
1167
1168
        missing_modules = (
            set(expected_modules) - set(inpainting_cls._optional_components) - set(inpainting_kwargs.keys())
        )
YiYi Xu's avatar
YiYi Xu committed
1169
1170
1171

        if len(missing_modules) > 0:
            raise ValueError(
1172
                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
1173
1174
1175
1176
1177
1178
1179
            )

        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