auto_pipeline.py 56.9 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
Yuxuan.Zhang's avatar
Yuxuan.Zhang committed
24
from .cogview3 import CogView3PlusPipeline
Yuxuan Zhang's avatar
Yuxuan Zhang committed
25
from .cogview4 import CogView4ControlPipeline, CogView4Pipeline
YiYi Xu's avatar
YiYi Xu committed
26
27
28
29
from .controlnet import (
    StableDiffusionControlNetImg2ImgPipeline,
    StableDiffusionControlNetInpaintPipeline,
    StableDiffusionControlNetPipeline,
30
    StableDiffusionXLControlNetImg2ImgPipeline,
31
    StableDiffusionXLControlNetInpaintPipeline,
YiYi Xu's avatar
YiYi Xu committed
32
    StableDiffusionXLControlNetPipeline,
33
34
35
    StableDiffusionXLControlNetUnionImg2ImgPipeline,
    StableDiffusionXLControlNetUnionInpaintPipeline,
    StableDiffusionXLControlNetUnionPipeline,
YiYi Xu's avatar
YiYi Xu committed
36
)
37
38
39
40
from .controlnet_sd3 import (
    StableDiffusion3ControlNetInpaintingPipeline,
    StableDiffusion3ControlNetPipeline,
)
YiYi Xu's avatar
YiYi Xu committed
41
from .deepfloyd_if import IFImg2ImgPipeline, IFInpaintingPipeline, IFPipeline
42
from .flux import (
43
44
    FluxControlImg2ImgPipeline,
    FluxControlInpaintPipeline,
45
46
47
    FluxControlNetImg2ImgPipeline,
    FluxControlNetInpaintPipeline,
    FluxControlNetPipeline,
48
    FluxControlPipeline,
49
50
51
52
    FluxImg2ImgPipeline,
    FluxInpaintPipeline,
    FluxPipeline,
)
53
from .hunyuandit import HunyuanDiTPipeline
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
from .kandinsky import (
    KandinskyCombinedPipeline,
    KandinskyImg2ImgCombinedPipeline,
    KandinskyImg2ImgPipeline,
    KandinskyInpaintCombinedPipeline,
    KandinskyInpaintPipeline,
    KandinskyPipeline,
)
from .kandinsky2_2 import (
    KandinskyV22CombinedPipeline,
    KandinskyV22Img2ImgCombinedPipeline,
    KandinskyV22Img2ImgPipeline,
    KandinskyV22InpaintCombinedPipeline,
    KandinskyV22InpaintPipeline,
    KandinskyV22Pipeline,
)
70
from .kandinsky3 import Kandinsky3Img2ImgPipeline, Kandinsky3Pipeline
71
from .latent_consistency_models import LatentConsistencyModelImg2ImgPipeline, LatentConsistencyModelPipeline
72
73
from .lumina import LuminaPipeline
from .lumina2 import Lumina2Pipeline
YiYi Xu's avatar
YiYi Xu committed
74
from .pag import (
75
    HunyuanDiTPAGPipeline,
76
    PixArtSigmaPAGPipeline,
77
    SanaPAGPipeline,
78
    StableDiffusion3PAGImg2ImgPipeline,
79
    StableDiffusion3PAGPipeline,
80
    StableDiffusionControlNetPAGInpaintPipeline,
81
    StableDiffusionControlNetPAGPipeline,
82
    StableDiffusionPAGImg2ImgPipeline,
83
    StableDiffusionPAGInpaintPipeline,
84
    StableDiffusionPAGPipeline,
85
    StableDiffusionXLControlNetPAGImg2ImgPipeline,
YiYi Xu's avatar
YiYi Xu committed
86
87
88
89
90
    StableDiffusionXLControlNetPAGPipeline,
    StableDiffusionXLPAGImg2ImgPipeline,
    StableDiffusionXLPAGInpaintPipeline,
    StableDiffusionXLPAGPipeline,
)
91
from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline
92
from .sana import SanaPipeline
93
from .stable_cascade import StableCascadeCombinedPipeline, StableCascadeDecoderPipeline
YiYi Xu's avatar
YiYi Xu committed
94
95
96
97
98
from .stable_diffusion import (
    StableDiffusionImg2ImgPipeline,
    StableDiffusionInpaintPipeline,
    StableDiffusionPipeline,
)
99
100
from .stable_diffusion_3 import (
    StableDiffusion3Img2ImgPipeline,
101
    StableDiffusion3InpaintPipeline,
102
103
    StableDiffusion3Pipeline,
)
YiYi Xu's avatar
YiYi Xu committed
104
105
106
107
108
from .stable_diffusion_xl import (
    StableDiffusionXLImg2ImgPipeline,
    StableDiffusionXLInpaintPipeline,
    StableDiffusionXLPipeline,
)
Kashif Rasul's avatar
Kashif Rasul committed
109
from .wuerstchen import WuerstchenCombinedPipeline, WuerstchenDecoderPipeline
YiYi Xu's avatar
YiYi Xu committed
110
111
112
113
114
115


AUTO_TEXT2IMAGE_PIPELINES_MAPPING = OrderedDict(
    [
        ("stable-diffusion", StableDiffusionPipeline),
        ("stable-diffusion-xl", StableDiffusionXLPipeline),
116
        ("stable-diffusion-3", StableDiffusion3Pipeline),
117
        ("stable-diffusion-3-pag", StableDiffusion3PAGPipeline),
YiYi Xu's avatar
YiYi Xu committed
118
        ("if", IFPipeline),
119
        ("hunyuan", HunyuanDiTPipeline),
120
        ("hunyuan-pag", HunyuanDiTPAGPipeline),
121
122
        ("kandinsky", KandinskyCombinedPipeline),
        ("kandinsky22", KandinskyV22CombinedPipeline),
123
        ("kandinsky3", Kandinsky3Pipeline),
YiYi Xu's avatar
YiYi Xu committed
124
125
        ("stable-diffusion-controlnet", StableDiffusionControlNetPipeline),
        ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetPipeline),
126
        ("stable-diffusion-xl-controlnet-union", StableDiffusionXLControlNetUnionPipeline),
127
        ("stable-diffusion-3-controlnet", StableDiffusion3ControlNetPipeline),
Kashif Rasul's avatar
Kashif Rasul committed
128
        ("wuerstchen", WuerstchenCombinedPipeline),
129
        ("cascade", StableCascadeCombinedPipeline),
130
        ("lcm", LatentConsistencyModelPipeline),
131
132
        ("pixart-alpha", PixArtAlphaPipeline),
        ("pixart-sigma", PixArtSigmaPipeline),
133
134
        ("sana", SanaPipeline),
        ("sana-pag", SanaPAGPipeline),
135
        ("stable-diffusion-pag", StableDiffusionPAGPipeline),
136
        ("stable-diffusion-controlnet-pag", StableDiffusionControlNetPAGPipeline),
YiYi Xu's avatar
YiYi Xu committed
137
138
        ("stable-diffusion-xl-pag", StableDiffusionXLPAGPipeline),
        ("stable-diffusion-xl-controlnet-pag", StableDiffusionXLControlNetPAGPipeline),
139
        ("pixart-sigma-pag", PixArtSigmaPAGPipeline),
140
        ("auraflow", AuraFlowPipeline),
Sayak Paul's avatar
Sayak Paul committed
141
        ("flux", FluxPipeline),
142
        ("flux-control", FluxControlPipeline),
143
        ("flux-controlnet", FluxControlNetPipeline),
144
145
        ("lumina", LuminaPipeline),
        ("lumina2", Lumina2Pipeline),
Yuxuan.Zhang's avatar
Yuxuan.Zhang committed
146
        ("cogview3", CogView3PlusPipeline),
147
        ("cogview4", CogView4Pipeline),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
148
        ("cogview4-control", CogView4ControlPipeline),
YiYi Xu's avatar
YiYi Xu committed
149
150
151
152
153
154
155
    ]
)

AUTO_IMAGE2IMAGE_PIPELINES_MAPPING = OrderedDict(
    [
        ("stable-diffusion", StableDiffusionImg2ImgPipeline),
        ("stable-diffusion-xl", StableDiffusionXLImg2ImgPipeline),
156
        ("stable-diffusion-3", StableDiffusion3Img2ImgPipeline),
157
        ("stable-diffusion-3-pag", StableDiffusion3PAGImg2ImgPipeline),
YiYi Xu's avatar
YiYi Xu committed
158
        ("if", IFImg2ImgPipeline),
159
160
        ("kandinsky", KandinskyImg2ImgCombinedPipeline),
        ("kandinsky22", KandinskyV22Img2ImgCombinedPipeline),
161
        ("kandinsky3", Kandinsky3Img2ImgPipeline),
YiYi Xu's avatar
YiYi Xu committed
162
        ("stable-diffusion-controlnet", StableDiffusionControlNetImg2ImgPipeline),
163
        ("stable-diffusion-pag", StableDiffusionPAGImg2ImgPipeline),
164
        ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetImg2ImgPipeline),
165
        ("stable-diffusion-xl-controlnet-union", StableDiffusionXLControlNetUnionImg2ImgPipeline),
YiYi Xu's avatar
YiYi Xu committed
166
        ("stable-diffusion-xl-pag", StableDiffusionXLPAGImg2ImgPipeline),
167
        ("stable-diffusion-xl-controlnet-pag", StableDiffusionXLControlNetPAGImg2ImgPipeline),
168
        ("lcm", LatentConsistencyModelImg2ImgPipeline),
169
        ("flux", FluxImg2ImgPipeline),
170
        ("flux-controlnet", FluxControlNetImg2ImgPipeline),
171
        ("flux-control", FluxControlImg2ImgPipeline),
YiYi Xu's avatar
YiYi Xu committed
172
173
174
175
176
177
178
    ]
)

AUTO_INPAINT_PIPELINES_MAPPING = OrderedDict(
    [
        ("stable-diffusion", StableDiffusionInpaintPipeline),
        ("stable-diffusion-xl", StableDiffusionXLInpaintPipeline),
179
        ("stable-diffusion-3", StableDiffusion3InpaintPipeline),
YiYi Xu's avatar
YiYi Xu committed
180
        ("if", IFInpaintingPipeline),
181
182
183
        ("kandinsky", KandinskyInpaintCombinedPipeline),
        ("kandinsky22", KandinskyV22InpaintCombinedPipeline),
        ("stable-diffusion-controlnet", StableDiffusionControlNetInpaintPipeline),
184
        ("stable-diffusion-controlnet-pag", StableDiffusionControlNetPAGInpaintPipeline),
185
        ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetInpaintPipeline),
186
        ("stable-diffusion-xl-controlnet-union", StableDiffusionXLControlNetUnionInpaintPipeline),
187
        ("stable-diffusion-3-controlnet", StableDiffusion3ControlNetInpaintingPipeline),
YiYi Xu's avatar
YiYi Xu committed
188
        ("stable-diffusion-xl-pag", StableDiffusionXLPAGInpaintPipeline),
189
        ("flux", FluxInpaintPipeline),
190
        ("flux-controlnet", FluxControlNetInpaintPipeline),
191
        ("flux-control", FluxControlInpaintPipeline),
192
        ("stable-diffusion-pag", StableDiffusionPAGInpaintPipeline),
193
194
195
196
197
198
199
    ]
)

_AUTO_TEXT2IMAGE_DECODER_PIPELINES_MAPPING = OrderedDict(
    [
        ("kandinsky", KandinskyPipeline),
        ("kandinsky22", KandinskyV22Pipeline),
Kashif Rasul's avatar
Kashif Rasul committed
200
        ("wuerstchen", WuerstchenDecoderPipeline),
201
        ("cascade", StableCascadeDecoderPipeline),
202
203
204
205
206
207
208
209
210
211
    ]
)
_AUTO_IMAGE2IMAGE_DECODER_PIPELINES_MAPPING = OrderedDict(
    [
        ("kandinsky", KandinskyImg2ImgPipeline),
        ("kandinsky22", KandinskyV22Img2ImgPipeline),
    ]
)
_AUTO_INPAINT_DECODER_PIPELINES_MAPPING = OrderedDict(
    [
YiYi Xu's avatar
YiYi Xu committed
212
213
214
215
216
        ("kandinsky", KandinskyInpaintPipeline),
        ("kandinsky22", KandinskyV22InpaintPipeline),
    ]
)

217
if is_sentencepiece_available():
218
    from .kolors import KolorsImg2ImgPipeline, KolorsPipeline
Álvaro Somoza's avatar
Álvaro Somoza committed
219
    from .pag import KolorsPAGPipeline
220
221

    AUTO_TEXT2IMAGE_PIPELINES_MAPPING["kolors"] = KolorsPipeline
Álvaro Somoza's avatar
Álvaro Somoza committed
222
    AUTO_TEXT2IMAGE_PIPELINES_MAPPING["kolors-pag"] = KolorsPAGPipeline
223
    AUTO_IMAGE2IMAGE_PIPELINES_MAPPING["kolors"] = KolorsImg2ImgPipeline
224

YiYi Xu's avatar
YiYi Xu committed
225
226
227
228
SUPPORTED_TASKS_MAPPINGS = [
    AUTO_TEXT2IMAGE_PIPELINES_MAPPING,
    AUTO_IMAGE2IMAGE_PIPELINES_MAPPING,
    AUTO_INPAINT_PIPELINES_MAPPING,
229
230
231
    _AUTO_TEXT2IMAGE_DECODER_PIPELINES_MAPPING,
    _AUTO_IMAGE2IMAGE_DECODER_PIPELINES_MAPPING,
    _AUTO_INPAINT_DECODER_PIPELINES_MAPPING,
YiYi Xu's avatar
YiYi Xu committed
232
233
234
]


235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
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)


def _get_task_class(mapping, pipeline_class_name, throw_error_if_not_exist: bool = True):
YiYi Xu's avatar
YiYi Xu committed
250
251
252
253
254
255
256
257
258
259
260
261
    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

    model_name = get_model(pipeline_class_name)

    if model_name is not None:
        task_class = mapping.get(model_name, None)
        if task_class is not None:
            return task_class
262
263
264

    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
265
266
267
268
269


class AutoPipelineForText2Image(ConfigMixin):
    r"""

270
271
272
    [`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
273

274
    This class cannot be instantiated using `__init__()` (throws an error).
YiYi Xu's avatar
YiYi Xu committed
275
276
277
278
279
280
281

    Class attributes:

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

    """
282

YiYi Xu's avatar
YiYi Xu committed
283
284
285
286
287
288
289
290
291
292
    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
293
    @validate_hf_hub_args
YiYi Xu's avatar
YiYi Xu committed
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
    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:

        ```
311
        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
312
313
314
315
316
        - 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:
317
            pretrained_model_or_path (`str` or `os.PathLike`, *optional*):
YiYi Xu's avatar
YiYi Xu committed
318
319
320
321
322
323
324
                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`].
325
326
            torch_dtype (`torch.dtype`, *optional*):
                Override the default `torch.dtype` and load the model with another dtype.
YiYi Xu's avatar
YiYi Xu committed
327
328
329
330
331
332
            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.
333

YiYi Xu's avatar
YiYi Xu committed
334
335
336
337
338
339
340
341
            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.
342
            token (`str` or *bool*, *optional*):
YiYi Xu's avatar
YiYi Xu committed
343
344
345
346
347
348
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
398
399
                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>

        To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with
        `huggingface-cli login`.

        </Tip>

        Examples:

        ```py
400
        >>> from diffusers import AutoPipelineForText2Image
YiYi Xu's avatar
YiYi Xu committed
401

402
        >>> pipeline = AutoPipelineForText2Image.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5")
403
        >>> image = pipeline(prompt).images[0]
YiYi Xu's avatar
YiYi Xu committed
404
405
        ```
        """
406
        cache_dir = kwargs.pop("cache_dir", None)
407
408
        force_download = kwargs.pop("force_download", False)
        proxies = kwargs.pop("proxies", None)
409
        token = kwargs.pop("token", None)
410
411
412
413
414
415
416
        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,
417
            "token": token,
418
419
420
421
422
            "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
423
        orig_class_name = config["_class_name"]
424
425
426
427
        if "ControlPipeline" in orig_class_name:
            to_replace = "ControlPipeline"
        else:
            to_replace = "Pipeline"
YiYi Xu's avatar
YiYi Xu committed
428
429

        if "controlnet" in kwargs:
430
            if isinstance(kwargs["controlnet"], ControlNetUnionModel):
431
                orig_class_name = config["_class_name"].replace(to_replace, "ControlNetUnionPipeline")
432
            else:
433
                orig_class_name = config["_class_name"].replace(to_replace, "ControlNetPipeline")
YiYi Xu's avatar
YiYi Xu committed
434
435
436
        if "enable_pag" in kwargs:
            enable_pag = kwargs.pop("enable_pag")
            if enable_pag:
437
                orig_class_name = orig_class_name.replace(to_replace, "PAGPipeline")
YiYi Xu's avatar
YiYi Xu committed
438
439
440

        text_2_image_cls = _get_task_class(AUTO_TEXT2IMAGE_PIPELINES_MAPPING, orig_class_name)

441
        kwargs = {**load_config_kwargs, **kwargs}
YiYi Xu's avatar
YiYi Xu committed
442
443
444
445
446
447
448
449
450
451
452
        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
453
        additional memory.
YiYi Xu's avatar
YiYi Xu committed
454
455
456
457
458
459
460
461

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

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

        ```py
462
        >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForImage2Image
YiYi Xu's avatar
YiYi Xu committed
463
464

        >>> pipe_i2i = AutoPipelineForImage2Image.from_pretrained(
465
        ...     "stable-diffusion-v1-5/stable-diffusion-v1-5", requires_safety_checker=False
YiYi Xu's avatar
YiYi Xu committed
466
467
        ... )

468
469
        >>> pipe_t2i = AutoPipelineForText2Image.from_pipe(pipe_i2i)
        >>> image = pipe_t2i(prompt).images[0]
YiYi Xu's avatar
YiYi Xu committed
470
471
472
473
474
475
476
477
478
        ```
        """

        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)

479
480
        if "controlnet" in kwargs:
            if kwargs["controlnet"] is not None:
YiYi Xu's avatar
YiYi Xu committed
481
                to_replace = "PAGPipeline" if "PAG" in text_2_image_cls.__name__ else "Pipeline"
482
483
                text_2_image_cls = _get_task_class(
                    AUTO_TEXT2IMAGE_PIPELINES_MAPPING,
YiYi Xu's avatar
YiYi Xu committed
484
                    text_2_image_cls.__name__.replace("ControlNet", "").replace(to_replace, "ControlNet" + to_replace),
485
486
487
488
                )
            else:
                text_2_image_cls = _get_task_class(
                    AUTO_TEXT2IMAGE_PIPELINES_MAPPING,
YiYi Xu's avatar
YiYi Xu committed
489
490
491
492
493
494
495
496
497
498
499
500
501
502
                    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", ""),
503
504
                )

YiYi Xu's avatar
YiYi Xu committed
505
        # define expected module and optional kwargs given the pipeline signature
506
        expected_modules, optional_kwargs = text_2_image_cls._get_signature_keys(text_2_image_cls)
YiYi Xu's avatar
YiYi Xu committed
507
508
509
510
511
512
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

        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
        }

545
546
547
        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
548
549
550

        if len(missing_modules) > 0:
            raise ValueError(
551
                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
552
553
554
555
556
557
558
559
560
561
562
563
            )

        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"""

564
565
566
    [`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
567

568
    This class cannot be instantiated using `__init__()` (throws an error).
YiYi Xu's avatar
YiYi Xu committed
569
570
571
572
573
574
575

    Class attributes:

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

    """
576

YiYi Xu's avatar
YiYi Xu committed
577
578
579
580
581
582
583
584
585
586
    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
587
    @validate_hf_hub_args
YiYi Xu's avatar
YiYi Xu committed
588
589
590
591
592
593
594
595
596
597
    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.

598
599
        If a `controlnet` argument is passed, it will instantiate a [`StableDiffusionControlNetImg2ImgPipeline`]
        object.
YiYi Xu's avatar
YiYi Xu committed
600
601
602
603
604
605

        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:

        ```
606
        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
607
608
609
610
611
        - 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:
612
            pretrained_model_or_path (`str` or `os.PathLike`, *optional*):
YiYi Xu's avatar
YiYi Xu committed
613
614
615
616
617
618
619
620
                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*):
621
                Override the default `torch.dtype` and load the model with another dtype.
YiYi Xu's avatar
YiYi Xu committed
622
623
624
625
626
627
            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.
628

YiYi Xu's avatar
YiYi Xu committed
629
630
631
632
633
634
635
636
            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.
637
            token (`str` or *bool*, *optional*):
YiYi Xu's avatar
YiYi Xu committed
638
639
640
641
642
643
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
693
694
                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>

        To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with
        `huggingface-cli login`.

        </Tip>

        Examples:

        ```py
695
        >>> from diffusers import AutoPipelineForImage2Image
YiYi Xu's avatar
YiYi Xu committed
696

697
        >>> pipeline = AutoPipelineForImage2Image.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5")
698
        >>> image = pipeline(prompt, image).images[0]
YiYi Xu's avatar
YiYi Xu committed
699
700
        ```
        """
701
        cache_dir = kwargs.pop("cache_dir", None)
702
703
        force_download = kwargs.pop("force_download", False)
        proxies = kwargs.pop("proxies", None)
704
        token = kwargs.pop("token", None)
705
706
707
708
709
710
711
        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,
712
            "token": token,
713
714
715
716
717
            "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
718
719
        orig_class_name = config["_class_name"]

720
721
        # the `orig_class_name` can be:
        # `- *Pipeline` (for regular text-to-image checkpoint)
722
        #  - `*ControlPipeline` (for Flux tools specific checkpoint)
723
        # `- *Img2ImgPipeline` (for refiner checkpoint)
724
725
726
727
728
729
        if "Img2Img" in orig_class_name:
            to_replace = "Img2ImgPipeline"
        elif "ControlPipeline" in orig_class_name:
            to_replace = "ControlPipeline"
        else:
            to_replace = "Pipeline"
730

YiYi Xu's avatar
YiYi Xu committed
731
        if "controlnet" in kwargs:
732
733
734
735
            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
736
737
738
        if "enable_pag" in kwargs:
            enable_pag = kwargs.pop("enable_pag")
            if enable_pag:
739
                orig_class_name = orig_class_name.replace(to_replace, "PAG" + to_replace)
YiYi Xu's avatar
YiYi Xu committed
740

741
742
743
        if to_replace == "ControlPipeline":
            orig_class_name = orig_class_name.replace(to_replace, "ControlImg2ImgPipeline")

YiYi Xu's avatar
YiYi Xu committed
744
745
        image_2_image_cls = _get_task_class(AUTO_IMAGE2IMAGE_PIPELINES_MAPPING, orig_class_name)

746
        kwargs = {**load_config_kwargs, **kwargs}
YiYi Xu's avatar
YiYi Xu committed
747
748
749
750
751
752
753
754
755
756
757
        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
758
        additional memory.
YiYi Xu's avatar
YiYi Xu committed
759
760
761
762
763
764
765
766
767
768

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

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

        Examples:

        ```py
769
        >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForImage2Image
YiYi Xu's avatar
YiYi Xu committed
770
771

        >>> pipe_t2i = AutoPipelineForText2Image.from_pretrained(
772
        ...     "stable-diffusion-v1-5/stable-diffusion-v1-5", requires_safety_checker=False
YiYi Xu's avatar
YiYi Xu committed
773
774
        ... )

775
776
        >>> pipe_i2i = AutoPipelineForImage2Image.from_pipe(pipe_t2i)
        >>> image = pipe_i2i(prompt, image).images[0]
YiYi Xu's avatar
YiYi Xu committed
777
778
779
780
781
782
783
784
785
        ```
        """

        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)

786
787
        if "controlnet" in kwargs:
            if kwargs["controlnet"] is not None:
YiYi Xu's avatar
YiYi Xu committed
788
789
790
                to_replace = "Img2ImgPipeline"
                if "PAG" in image_2_image_cls.__name__:
                    to_replace = "PAG" + to_replace
791
792
                image_2_image_cls = _get_task_class(
                    AUTO_IMAGE2IMAGE_PIPELINES_MAPPING,
793
                    image_2_image_cls.__name__.replace("ControlNet", "").replace(
YiYi Xu's avatar
YiYi Xu committed
794
                        to_replace, "ControlNet" + to_replace
795
                    ),
796
797
798
799
                )
            else:
                image_2_image_cls = _get_task_class(
                    AUTO_IMAGE2IMAGE_PIPELINES_MAPPING,
YiYi Xu's avatar
YiYi Xu committed
800
801
802
803
804
805
806
807
808
809
810
811
812
813
                    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", ""),
814
815
                )

YiYi Xu's avatar
YiYi Xu committed
816
        # define expected module and optional kwargs given the pipeline signature
817
        expected_modules, optional_kwargs = image_2_image_cls._get_signature_keys(image_2_image_cls)
YiYi Xu's avatar
YiYi Xu committed
818
819
820
821
822
823
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

        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
        }

856
857
858
        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
859
860
861

        if len(missing_modules) > 0:
            raise ValueError(
862
                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
863
864
865
866
867
868
869
870
871
872
873
874
            )

        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"""

875
876
877
    [`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
878

879
    This class cannot be instantiated using `__init__()` (throws an error).
YiYi Xu's avatar
YiYi Xu committed
880
881
882
883
884
885
886

    Class attributes:

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

    """
887

YiYi Xu's avatar
YiYi Xu committed
888
889
890
891
892
893
894
895
896
897
    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
898
    @validate_hf_hub_args
YiYi Xu's avatar
YiYi Xu committed
899
900
901
902
903
904
905
906
907
    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.

908
909
        If a `controlnet` argument is passed, it will instantiate a [`StableDiffusionControlNetInpaintPipeline`]
        object.
YiYi Xu's avatar
YiYi Xu committed
910
911
912
913
914
915

        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:

        ```
916
        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
917
918
919
920
921
        - 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:
922
            pretrained_model_or_path (`str` or `os.PathLike`, *optional*):
YiYi Xu's avatar
YiYi Xu committed
923
924
925
926
927
928
929
930
                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*):
931
                Override the default `torch.dtype` and load the model with another dtype.
YiYi Xu's avatar
YiYi Xu committed
932
933
934
935
936
937
            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.
938

YiYi Xu's avatar
YiYi Xu committed
939
940
941
942
943
944
945
946
            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.
947
            token (`str` or *bool*, *optional*):
YiYi Xu's avatar
YiYi Xu committed
948
949
950
951
952
953
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
1003
1004
                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>

        To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with
        `huggingface-cli login`.

        </Tip>

        Examples:

        ```py
1005
        >>> from diffusers import AutoPipelineForInpainting
YiYi Xu's avatar
YiYi Xu committed
1006

1007
        >>> pipeline = AutoPipelineForInpainting.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5")
1008
        >>> image = pipeline(prompt, image=init_image, mask_image=mask_image).images[0]
YiYi Xu's avatar
YiYi Xu committed
1009
1010
        ```
        """
1011
        cache_dir = kwargs.pop("cache_dir", None)
1012
1013
        force_download = kwargs.pop("force_download", False)
        proxies = kwargs.pop("proxies", None)
1014
        token = kwargs.pop("token", None)
1015
1016
1017
1018
1019
1020
1021
        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,
1022
            "token": token,
1023
1024
1025
1026
1027
            "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
1028
1029
        orig_class_name = config["_class_name"]

1030
1031
        # The `orig_class_name`` can be:
        # `- *InpaintPipeline` (for inpaint-specific checkpoint)
1032
        #  - `*ControlPipeline` (for Flux tools specific checkpoint)
1033
        #  - or *Pipeline (for regular text-to-image checkpoint)
1034
1035
1036
1037
1038
1039
        if "Inpaint" in orig_class_name:
            to_replace = "InpaintPipeline"
        elif "ControlPipeline" in orig_class_name:
            to_replace = "ControlPipeline"
        else:
            to_replace = "Pipeline"
1040

YiYi Xu's avatar
YiYi Xu committed
1041
        if "controlnet" in kwargs:
1042
1043
1044
1045
            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
1046
1047
1048
        if "enable_pag" in kwargs:
            enable_pag = kwargs.pop("enable_pag")
            if enable_pag:
1049
                orig_class_name = orig_class_name.replace(to_replace, "PAG" + to_replace)
1050
1051
        if to_replace == "ControlPipeline":
            orig_class_name = orig_class_name.replace(to_replace, "ControlInpaintPipeline")
YiYi Xu's avatar
YiYi Xu committed
1052
1053
        inpainting_cls = _get_task_class(AUTO_INPAINT_PIPELINES_MAPPING, orig_class_name)

1054
        kwargs = {**load_config_kwargs, **kwargs}
YiYi Xu's avatar
YiYi Xu committed
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
        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
1066
        additional memory.
YiYi Xu's avatar
YiYi Xu committed
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076

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

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

        Examples:

        ```py
1077
        >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForInpainting
YiYi Xu's avatar
YiYi Xu committed
1078
1079
1080
1081
1082
1083

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

        >>> pipe_inpaint = AutoPipelineForInpainting.from_pipe(pipe_t2i)
1084
        >>> image = pipe_inpaint(prompt, image=init_image, mask_image=mask_image).images[0]
YiYi Xu's avatar
YiYi Xu committed
1085
1086
1087
1088
1089
1090
1091
1092
        ```
        """
        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)

1093
1094
1095
1096
        if "controlnet" in kwargs:
            if kwargs["controlnet"] is not None:
                inpainting_cls = _get_task_class(
                    AUTO_INPAINT_PIPELINES_MAPPING,
1097
1098
1099
                    inpainting_cls.__name__.replace("ControlNet", "").replace(
                        "InpaintPipeline", "ControlNetInpaintPipeline"
                    ),
1100
1101
1102
1103
1104
1105
1106
                )
            else:
                inpainting_cls = _get_task_class(
                    AUTO_INPAINT_PIPELINES_MAPPING,
                    inpainting_cls.__name__.replace("ControlNetInpaintPipeline", "InpaintPipeline"),
                )

YiYi Xu's avatar
YiYi Xu committed
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
        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
1120
        # define expected module and optional kwargs given the pipeline signature
1121
        expected_modules, optional_kwargs = inpainting_cls._get_signature_keys(inpainting_cls)
YiYi Xu's avatar
YiYi Xu committed
1122
1123
1124
1125
1126
1127
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

        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
        }

1160
1161
1162
        missing_modules = (
            set(expected_modules) - set(inpainting_cls._optional_components) - set(inpainting_kwargs.keys())
        )
YiYi Xu's avatar
YiYi Xu committed
1163
1164
1165

        if len(missing_modules) > 0:
            raise ValueError(
1166
                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
1167
1168
1169
1170
1171
1172
1173
            )

        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