test_stable_diffusion_inpaint.py 57.6 KB
Newer Older
1
# coding=utf-8
Patrick von Platen's avatar
Patrick von Platen committed
2
# Copyright 2023 HuggingFace Inc.
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.

import gc
import random
18
import traceback
19
20
21
22
import unittest

import numpy as np
import torch
23
from huggingface_hub import hf_hub_download
24
25
from PIL import Image
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
26
27

from diffusers import (
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
28
    AsymmetricAutoencoderKL,
29
    AutoencoderKL,
30
    DDIMScheduler,
31
    DPMSolverMultistepScheduler,
Patrick von Platen's avatar
Patrick von Platen committed
32
    LCMScheduler,
33
    LMSDiscreteScheduler,
34
35
36
37
    PNDMScheduler,
    StableDiffusionInpaintPipeline,
    UNet2DConditionModel,
)
38
from diffusers.models.attention_processor import AttnProcessor
39
from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint import prepare_mask_and_masked_image
40
41
from diffusers.utils.testing_utils import (
    enable_full_determinism,
Dhruv Nair's avatar
Dhruv Nair committed
42
43
44
45
    floats_tensor,
    load_image,
    load_numpy,
    nightly,
Dhruv Nair's avatar
Dhruv Nair committed
46
    require_python39_or_higher,
47
48
49
    require_torch_2,
    require_torch_gpu,
    run_test_in_subprocess,
Dhruv Nair's avatar
Dhruv Nair committed
50
51
    slow,
    torch_device,
52
)
53

54
55
56
57
58
from ..pipeline_params import (
    TEXT_GUIDED_IMAGE_INPAINTING_BATCH_PARAMS,
    TEXT_GUIDED_IMAGE_INPAINTING_PARAMS,
    TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS,
)
59
from ..test_pipelines_common import PipelineKarrasSchedulerTesterMixin, PipelineLatentTesterMixin, PipelineTesterMixin
60

61

62
enable_full_determinism()
63
64


65
66
67
68
69
70
71
72
73
74
75
76
# Will be run via run_test_in_subprocess
def _test_inpaint_compile(in_queue, out_queue, timeout):
    error = None
    try:
        inputs = in_queue.get(timeout=timeout)
        torch_device = inputs.pop("torch_device")
        seed = inputs.pop("seed")
        inputs["generator"] = torch.Generator(device=torch_device).manual_seed(seed)

        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
77
        pipe.unet.set_default_attn_processor()
78
79
80
81
82
83
84
85
86
87
88
        pipe.scheduler = PNDMScheduler.from_config(pipe.scheduler.config)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        pipe.unet.to(memory_format=torch.channels_last)
        pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)

        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
89
        expected_slice = np.array([0.0689, 0.0699, 0.0790, 0.0536, 0.0470, 0.0488, 0.041, 0.0508, 0.04179])
90
91
92
93
94
95
96
97
98
        assert np.abs(expected_slice - image_slice).max() < 3e-3
    except Exception:
        error = f"{traceback.format_exc()}"

    results = {"error": error}
    out_queue.put(results, timeout=timeout)
    out_queue.join()


99
100
101
class StableDiffusionInpaintPipelineFastTests(
    PipelineLatentTesterMixin, PipelineKarrasSchedulerTesterMixin, PipelineTesterMixin, unittest.TestCase
):
102
    pipeline_class = StableDiffusionInpaintPipeline
103
104
    params = TEXT_GUIDED_IMAGE_INPAINTING_PARAMS
    batch_params = TEXT_GUIDED_IMAGE_INPAINTING_BATCH_PARAMS
105
106
    image_params = frozenset([])
    # TO-DO: update image_params once pipeline is refactored with VaeImageProcessor.preprocess
107
    image_latents_params = frozenset([])
108
    callback_cfg_params = TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS.union({"mask", "masked_image_latents"})
109

Patrick von Platen's avatar
Patrick von Platen committed
110
    def get_dummy_components(self, time_cond_proj_dim=None):
111
        torch.manual_seed(0)
112
        unet = UNet2DConditionModel(
113
            block_out_channels=(32, 64),
Patrick von Platen's avatar
Patrick von Platen committed
114
            time_cond_proj_dim=time_cond_proj_dim,
115
116
117
118
119
120
121
122
            layers_per_block=2,
            sample_size=32,
            in_channels=9,
            out_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            cross_attention_dim=32,
        )
123
        scheduler = PNDMScheduler(skip_prk_steps=True)
124
        torch.manual_seed(0)
125
        vae = AutoencoderKL(
126
127
128
129
130
131
132
133
            block_out_channels=[32, 64],
            in_channels=3,
            out_channels=3,
            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
            latent_channels=4,
        )
        torch.manual_seed(0)
134
        text_encoder_config = CLIPTextConfig(
135
136
137
138
139
140
141
142
143
144
            bos_token_id=0,
            eos_token_id=2,
            hidden_size=32,
            intermediate_size=37,
            layer_norm_eps=1e-05,
            num_attention_heads=4,
            num_hidden_layers=5,
            pad_token_id=1,
            vocab_size=1000,
        )
145
        text_encoder = CLIPTextModel(text_encoder_config)
146
147
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

148
149
150
151
152
153
154
        components = {
            "unet": unet,
            "scheduler": scheduler,
            "vae": vae,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "safety_checker": None,
155
            "feature_extractor": None,
156
            "image_encoder": None,
157
158
159
        }
        return components

160
    def get_dummy_inputs(self, device, seed=0, img_res=64, output_pil=True):
161
        # TODO: use tensor inputs instead of PIL, this is here just to leave the old expected_slices untouched
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
        if output_pil:
            # Get random floats in [0, 1] as image
            image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
            image = image.cpu().permute(0, 2, 3, 1)[0]
            mask_image = torch.ones_like(image)
            # Convert image and mask_image to [0, 255]
            image = 255 * image
            mask_image = 255 * mask_image
            # Convert to PIL image
            init_image = Image.fromarray(np.uint8(image)).convert("RGB").resize((img_res, img_res))
            mask_image = Image.fromarray(np.uint8(mask_image)).convert("RGB").resize((img_res, img_res))
        else:
            # Get random floats in [0, 1] as image with spatial size (img_res, img_res)
            image = floats_tensor((1, 3, img_res, img_res), rng=random.Random(seed)).to(device)
            # Convert image to [-1, 1]
            init_image = 2.0 * image - 1.0
            mask_image = torch.ones((1, 1, img_res, img_res), device=device)

180
181
182
183
        if str(device).startswith("mps"):
            generator = torch.manual_seed(seed)
        else:
            generator = torch.Generator(device=device).manual_seed(seed)
184

185
186
187
188
189
190
191
192
193
194
        inputs = {
            "prompt": "A painting of a squirrel eating a burger",
            "image": init_image,
            "mask_image": mask_image,
            "generator": generator,
            "num_inference_steps": 2,
            "guidance_scale": 6.0,
            "output_type": "numpy",
        }
        return inputs
195

196
197
198
199
    def test_stable_diffusion_inpaint(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionInpaintPipeline(**components)
200
201
202
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

203
204
        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
205
206
        image_slice = image[0, -3:, -3:, -1]

207
        assert image.shape == (1, 64, 64, 3)
208
        expected_slice = np.array([0.4703, 0.5697, 0.3879, 0.5470, 0.6042, 0.4413, 0.5078, 0.4728, 0.4469])
209

210
211
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

Patrick von Platen's avatar
Patrick von Platen committed
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
    def test_stable_diffusion_inpaint_lcm(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components(time_cond_proj_dim=256)
        sd_pipe = StableDiffusionInpaintPipeline(**components)
        sd_pipe.scheduler = LCMScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
        expected_slice = np.array([0.4931, 0.5988, 0.4569, 0.5556, 0.6650, 0.5087, 0.5966, 0.5358, 0.5269])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
    def test_stable_diffusion_inpaint_lcm_custom_timesteps(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components(time_cond_proj_dim=256)
        sd_pipe = StableDiffusionInpaintPipeline(**components)
        sd_pipe.scheduler = LCMScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        del inputs["num_inference_steps"]
        inputs["timesteps"] = [999, 499]
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
        expected_slice = np.array([0.4931, 0.5988, 0.4569, 0.5556, 0.6650, 0.5087, 0.5966, 0.5358, 0.5269])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

248
249
    def test_stable_diffusion_inpaint_image_tensor(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
250
251
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionInpaintPipeline(**components)
252
253
254
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

255
256
257
        inputs = self.get_dummy_inputs(device)
        output = sd_pipe(**inputs)
        out_pil = output.images
258

259
260
261
262
263
        inputs = self.get_dummy_inputs(device)
        inputs["image"] = torch.tensor(np.array(inputs["image"]) / 127.5 - 1).permute(2, 0, 1).unsqueeze(0)
        inputs["mask_image"] = torch.tensor(np.array(inputs["mask_image"]) / 255).permute(2, 0, 1)[:1].unsqueeze(0)
        output = sd_pipe(**inputs)
        out_tensor = output.images
264

265
266
        assert out_pil.shape == (1, 64, 64, 3)
        assert np.abs(out_pil.flatten() - out_tensor.flatten()).max() < 5e-2
267

268
269
270
    def test_inference_batch_single_identical(self):
        super().test_inference_batch_single_identical(expected_max_diff=3e-3)

271
272
273
274
275
276
277
278
279
280
281
282
283
284
    def test_stable_diffusion_inpaint_strength_zero_test(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionInpaintPipeline(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)

        # check that the pipeline raises value error when num_inference_steps is < 1
        inputs["strength"] = 0.01
        with self.assertRaises(ValueError):
            sd_pipe(**inputs).images

285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
    def test_stable_diffusion_inpaint_mask_latents(self):
        device = "cpu"
        components = self.get_dummy_components()
        sd_pipe = self.pipeline_class(**components).to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        # normal mask + normal image
        ##  `image`: pil, `mask_image``: pil, `masked_image_latents``: None
        inputs = self.get_dummy_inputs(device)
        inputs["strength"] = 0.9
        out_0 = sd_pipe(**inputs).images

        # image latents + mask latents
        inputs = self.get_dummy_inputs(device)
        image = sd_pipe.image_processor.preprocess(inputs["image"]).to(sd_pipe.device)
        mask = sd_pipe.mask_processor.preprocess(inputs["mask_image"]).to(sd_pipe.device)
        masked_image = image * (mask < 0.5)

        generator = torch.Generator(device=device).manual_seed(0)
        image_latents = (
            sd_pipe.vae.encode(image).latent_dist.sample(generator=generator) * sd_pipe.vae.config.scaling_factor
        )
        torch.randn((1, 4, 32, 32), generator=generator)
        mask_latents = (
            sd_pipe.vae.encode(masked_image).latent_dist.sample(generator=generator)
            * sd_pipe.vae.config.scaling_factor
        )
        inputs["image"] = image_latents
        inputs["masked_image_latents"] = mask_latents
        inputs["mask_image"] = mask
        inputs["strength"] = 0.9
        generator = torch.Generator(device=device).manual_seed(0)
        torch.randn((1, 4, 32, 32), generator=generator)
        inputs["generator"] = generator
        out_1 = sd_pipe(**inputs).images
        assert np.abs(out_0 - out_1).max() < 1e-2

322

323
324
325
326
327
328
329
class StableDiffusionSimpleInpaintPipelineFastTests(StableDiffusionInpaintPipelineFastTests):
    pipeline_class = StableDiffusionInpaintPipeline
    params = TEXT_GUIDED_IMAGE_INPAINTING_PARAMS
    batch_params = TEXT_GUIDED_IMAGE_INPAINTING_BATCH_PARAMS
    image_params = frozenset([])
    # TO-DO: update image_params once pipeline is refactored with VaeImageProcessor.preprocess

Patrick von Platen's avatar
Patrick von Platen committed
330
    def get_dummy_components(self, time_cond_proj_dim=None):
331
332
333
334
        torch.manual_seed(0)
        unet = UNet2DConditionModel(
            block_out_channels=(32, 64),
            layers_per_block=2,
Patrick von Platen's avatar
Patrick von Platen committed
335
            time_cond_proj_dim=time_cond_proj_dim,
336
337
338
339
340
341
342
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
            sample_size=32,
            in_channels=4,
            out_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            cross_attention_dim=32,
        )
        scheduler = PNDMScheduler(skip_prk_steps=True)
        torch.manual_seed(0)
        vae = AutoencoderKL(
            block_out_channels=[32, 64],
            in_channels=3,
            out_channels=3,
            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
            latent_channels=4,
        )
        torch.manual_seed(0)
        text_encoder_config = CLIPTextConfig(
            bos_token_id=0,
            eos_token_id=2,
            hidden_size=32,
            intermediate_size=37,
            layer_norm_eps=1e-05,
            num_attention_heads=4,
            num_hidden_layers=5,
            pad_token_id=1,
            vocab_size=1000,
        )
        text_encoder = CLIPTextModel(text_encoder_config)
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        components = {
            "unet": unet,
            "scheduler": scheduler,
            "vae": vae,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "safety_checker": None,
            "feature_extractor": None,
376
            "image_encoder": None,
377
378
379
        }
        return components

380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
    def get_dummy_inputs_2images(self, device, seed=0, img_res=64):
        # Get random floats in [0, 1] as image with spatial size (img_res, img_res)
        image1 = floats_tensor((1, 3, img_res, img_res), rng=random.Random(seed)).to(device)
        image2 = floats_tensor((1, 3, img_res, img_res), rng=random.Random(seed + 22)).to(device)
        # Convert images to [-1, 1]
        init_image1 = 2.0 * image1 - 1.0
        init_image2 = 2.0 * image2 - 1.0

        # empty mask
        mask_image = torch.zeros((1, 1, img_res, img_res), device=device)

        if str(device).startswith("mps"):
            generator1 = torch.manual_seed(seed)
            generator2 = torch.manual_seed(seed)
        else:
            generator1 = torch.Generator(device=device).manual_seed(seed)
            generator2 = torch.Generator(device=device).manual_seed(seed)

        inputs = {
            "prompt": ["A painting of a squirrel eating a burger"] * 2,
            "image": [init_image1, init_image2],
            "mask_image": [mask_image] * 2,
            "generator": [generator1, generator2],
            "num_inference_steps": 2,
            "guidance_scale": 6.0,
            "output_type": "numpy",
        }
        return inputs

409
410
411
412
413
414
415
416
417
418
419
420
    def test_stable_diffusion_inpaint(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionInpaintPipeline(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
421
        expected_slice = np.array([0.6584, 0.5424, 0.5649, 0.5449, 0.5897, 0.6111, 0.5404, 0.5463, 0.5214])
422
423
424

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

Patrick von Platen's avatar
Patrick von Platen committed
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
    def test_stable_diffusion_inpaint_lcm(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components(time_cond_proj_dim=256)
        sd_pipe = StableDiffusionInpaintPipeline(**components)
        sd_pipe.scheduler = LCMScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
        expected_slice = np.array([0.6240, 0.5355, 0.5649, 0.5378, 0.5374, 0.6242, 0.5132, 0.5347, 0.5396])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
    def test_stable_diffusion_inpaint_lcm_custom_timesteps(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components(time_cond_proj_dim=256)
        sd_pipe = StableDiffusionInpaintPipeline(**components)
        sd_pipe.scheduler = LCMScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        del inputs["num_inference_steps"]
        inputs["timesteps"] = [999, 499]
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
        expected_slice = np.array([0.6240, 0.5355, 0.5649, 0.5378, 0.5374, 0.6242, 0.5132, 0.5347, 0.5396])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
    def test_stable_diffusion_inpaint_2_images(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        sd_pipe = self.pipeline_class(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        # test to confirm if we pass two same image, we will get same output
        inputs = self.get_dummy_inputs(device)
        gen1 = torch.Generator(device=device).manual_seed(0)
        gen2 = torch.Generator(device=device).manual_seed(0)
        for name in ["prompt", "image", "mask_image"]:
            inputs[name] = [inputs[name]] * 2
        inputs["generator"] = [gen1, gen2]
        images = sd_pipe(**inputs).images

        assert images.shape == (2, 64, 64, 3)

        image_slice1 = images[0, -3:, -3:, -1]
        image_slice2 = images[1, -3:, -3:, -1]
        assert np.abs(image_slice1.flatten() - image_slice2.flatten()).max() < 1e-4

        # test to confirm that if we pass two different images, we will get different output
        inputs = self.get_dummy_inputs_2images(device)
        images = sd_pipe(**inputs).images
        assert images.shape == (2, 64, 64, 3)

        image_slice1 = images[0, -3:, -3:, -1]
        image_slice2 = images[1, -3:, -3:, -1]
        assert np.abs(image_slice1.flatten() - image_slice2.flatten()).max() > 1e-2

492

493
@slow
494
@require_torch_gpu
495
496
497
498
class StableDiffusionInpaintPipelineSlowTests(unittest.TestCase):
    def setUp(self):
        super().setUp()

499
500
501
502
503
    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

504
505
    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
        generator = torch.Generator(device=generator_device).manual_seed(seed)
506
        init_image = load_image(
507
508
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/input_bench_image.png"
509
510
        )
        mask_image = load_image(
511
512
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/input_bench_mask.png"
513
        )
514
515
516
517
518
519
520
521
522
523
        inputs = {
            "prompt": "Face of a yellow cat, high resolution, sitting on a park bench",
            "image": init_image,
            "mask_image": mask_image,
            "generator": generator,
            "num_inference_steps": 3,
            "guidance_scale": 7.5,
            "output_type": "numpy",
        }
        return inputs
524

525
526
527
528
    def test_stable_diffusion_inpaint_ddim(self):
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
529
530
531
532
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

533
534
535
        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()
536

537
        assert image.shape == (1, 512, 512, 3)
538
539
        expected_slice = np.array([0.0427, 0.0460, 0.0483, 0.0460, 0.0584, 0.0521, 0.1549, 0.1695, 0.1794])

540
        assert np.abs(expected_slice - image_slice).max() < 6e-4
541
542
543

    def test_stable_diffusion_inpaint_fp16(self):
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
544
            "runwayml/stable-diffusion-inpainting", torch_dtype=torch.float16, safety_checker=None
545
        )
546
        pipe.unet.set_default_attn_processor()
547
548
549
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()
550

551
552
553
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()
554

555
        assert image.shape == (1, 512, 512, 3)
556
        expected_slice = np.array([0.1509, 0.1245, 0.1672, 0.1655, 0.1519, 0.1226, 0.1462, 0.1567, 0.2451])
557
        assert np.abs(expected_slice - image_slice).max() < 1e-1
558

559
    def test_stable_diffusion_inpaint_pndm(self):
560
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
561
            "runwayml/stable-diffusion-inpainting", safety_checker=None
562
        )
563
        pipe.scheduler = PNDMScheduler.from_config(pipe.scheduler.config)
564
565
566
567
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

568
569
570
        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()
571

572
        assert image.shape == (1, 512, 512, 3)
573
574
        expected_slice = np.array([0.0425, 0.0273, 0.0344, 0.1694, 0.1727, 0.1812, 0.3256, 0.3311, 0.3272])

575
        assert np.abs(expected_slice - image_slice).max() < 5e-3
576

577
578
579
    def test_stable_diffusion_inpaint_k_lms(self):
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
580
        )
581
        pipe.scheduler = LMSDiscreteScheduler.from_config(pipe.scheduler.config)
582
583
584
585
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

586
587
588
589
590
        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
591
592
        expected_slice = np.array([0.9314, 0.7575, 0.9432, 0.8885, 0.9028, 0.7298, 0.9811, 0.9667, 0.7633])

593
        assert np.abs(expected_slice - image_slice).max() < 6e-3
594

595
596
597
598
599
600
    def test_stable_diffusion_inpaint_with_sequential_cpu_offloading(self):
        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()

        pipe = StableDiffusionInpaintPipeline.from_pretrained(
601
            "runwayml/stable-diffusion-inpainting", safety_checker=None, torch_dtype=torch.float16
602
        )
603
604
605
606
607
608
609
610
611
612
613
614
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing(1)
        pipe.enable_sequential_cpu_offload()

        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        _ = pipe(**inputs)

        mem_bytes = torch.cuda.max_memory_allocated()
        # make sure that less than 2.2 GB is allocated
        assert mem_bytes < 2.2 * 10**9

Dhruv Nair's avatar
Dhruv Nair committed
615
    @require_python39_or_higher
616
    @require_torch_2
617
    def test_inpaint_compile(self):
618
619
620
621
622
623
624
        seed = 0
        inputs = self.get_inputs(torch_device, seed=seed)
        # Can't pickle a Generator object
        del inputs["generator"]
        inputs["torch_device"] = torch_device
        inputs["seed"] = seed
        run_test_in_subprocess(test_case=self, target_func=_test_inpaint_compile, inputs=inputs)
625

626
    def test_stable_diffusion_inpaint_pil_input_resolution_test(self):
Patrick von Platen's avatar
Patrick von Platen committed
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
        pipe.scheduler = LMSDiscreteScheduler.from_config(pipe.scheduler.config)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        inputs = self.get_inputs(torch_device)
        # change input image to a random size (one that would cause a tensor mismatch error)
        inputs["image"] = inputs["image"].resize((127, 127))
        inputs["mask_image"] = inputs["mask_image"].resize((127, 127))
        inputs["height"] = 128
        inputs["width"] = 128
        image = pipe(**inputs).images
        # verify that the returned image has the same height and width as the input height and width
        assert image.shape == (1, inputs["height"], inputs["width"], 3)
644

645
646
647
648
649
    def test_stable_diffusion_inpaint_strength_test(self):
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
        pipe.scheduler = LMSDiscreteScheduler.from_config(pipe.scheduler.config)
650
        pipe.unet.set_default_attn_processor()
651
652
653
654
655
656
657
658
659
660
661
662
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        inputs = self.get_inputs(torch_device)
        # change input strength
        inputs["strength"] = 0.75
        image = pipe(**inputs).images
        # verify that the returned image has the same height and width as the input height and width
        assert image.shape == (1, 512, 512, 3)

        image_slice = image[0, 253:256, 253:256, -1].flatten()
663
664
        expected_slice = np.array([0.2728, 0.2803, 0.2665, 0.2511, 0.2774, 0.2586, 0.2391, 0.2392, 0.2582])
        assert np.abs(expected_slice - image_slice).max() < 1e-3
665

666
667
    def test_stable_diffusion_simple_inpaint_ddim(self):
        pipe = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", safety_checker=None)
668
        pipe.unet.set_default_attn_processor()
669
670
671
672
673
674
675
676
677
678
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images

        image_slice = image[0, 253:256, 253:256, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
679
680
        expected_slice = np.array([0.3757, 0.3875, 0.4445, 0.4353, 0.3780, 0.4513, 0.3965, 0.3984, 0.4362])
        assert np.abs(expected_slice - image_slice).max() < 1e-3
681

682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
    def test_download_local(self):
        filename = hf_hub_download("runwayml/stable-diffusion-inpainting", filename="sd-v1-5-inpainting.ckpt")

        pipe = StableDiffusionInpaintPipeline.from_single_file(filename, torch_dtype=torch.float16)
        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
        pipe.to("cuda")

        inputs = self.get_inputs(torch_device)
        inputs["num_inference_steps"] = 1
        image_out = pipe(**inputs).images[0]

        assert image_out.shape == (512, 512, 3)

    def test_download_ckpt_diff_format_is_same(self):
        ckpt_path = "https://huggingface.co/runwayml/stable-diffusion-inpainting/blob/main/sd-v1-5-inpainting.ckpt"

        pipe = StableDiffusionInpaintPipeline.from_single_file(ckpt_path)
        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
        pipe.unet.set_attn_processor(AttnProcessor())
        pipe.to("cuda")

        inputs = self.get_inputs(torch_device)
        inputs["num_inference_steps"] = 5
        image_ckpt = pipe(**inputs).images[0]

        pipe = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-inpainting")
        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
        pipe.unet.set_attn_processor(AttnProcessor())
        pipe.to("cuda")

        inputs = self.get_inputs(torch_device)
        inputs["num_inference_steps"] = 5
        image = pipe(**inputs).images[0]

716
        assert np.max(np.abs(image - image_ckpt)) < 5e-4
717

718

Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
@slow
@require_torch_gpu
class StableDiffusionInpaintPipelineAsymmetricAutoencoderKLSlowTests(unittest.TestCase):
    def setUp(self):
        super().setUp()

    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
        generator = torch.Generator(device=generator_device).manual_seed(seed)
        init_image = load_image(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/input_bench_image.png"
        )
        mask_image = load_image(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/input_bench_mask.png"
        )
        inputs = {
            "prompt": "Face of a yellow cat, high resolution, sitting on a park bench",
            "image": init_image,
            "mask_image": mask_image,
            "generator": generator,
            "num_inference_steps": 3,
            "guidance_scale": 7.5,
            "output_type": "numpy",
        }
        return inputs

    def test_stable_diffusion_inpaint_ddim(self):
        vae = AsymmetricAutoencoderKL.from_pretrained("cross-attention/asymmetric-autoencoder-kl-x-1-5")
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
        pipe.vae = vae
757
        pipe.unet.set_default_attn_processor()
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
758
759
760
761
762
763
764
765
766
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
767
        expected_slice = np.array([0.0522, 0.0604, 0.0596, 0.0449, 0.0493, 0.0427, 0.1186, 0.1289, 0.1442])
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
768

769
        assert np.abs(expected_slice - image_slice).max() < 1e-3
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
770
771
772
773
774
775
776
777

    def test_stable_diffusion_inpaint_fp16(self):
        vae = AsymmetricAutoencoderKL.from_pretrained(
            "cross-attention/asymmetric-autoencoder-kl-x-1-5", torch_dtype=torch.float16
        )
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", torch_dtype=torch.float16, safety_checker=None
        )
778
        pipe.unet.set_default_attn_processor()
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
        pipe.vae = vae
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
        expected_slice = np.array([0.1343, 0.1406, 0.1440, 0.1504, 0.1729, 0.0989, 0.1807, 0.2822, 0.1179])

        assert np.abs(expected_slice - image_slice).max() < 5e-2

    def test_stable_diffusion_inpaint_pndm(self):
        vae = AsymmetricAutoencoderKL.from_pretrained("cross-attention/asymmetric-autoencoder-kl-x-1-5")
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
798
        pipe.unet.set_default_attn_processor()
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
799
800
801
802
803
804
805
806
807
808
809
        pipe.vae = vae
        pipe.scheduler = PNDMScheduler.from_config(pipe.scheduler.config)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
810
        expected_slice = np.array([0.0966, 0.1083, 0.1148, 0.1422, 0.1318, 0.1197, 0.3702, 0.3537, 0.3288])
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
811
812
813
814
815
816
817
818

        assert np.abs(expected_slice - image_slice).max() < 5e-3

    def test_stable_diffusion_inpaint_k_lms(self):
        vae = AsymmetricAutoencoderKL.from_pretrained("cross-attention/asymmetric-autoencoder-kl-x-1-5")
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
819
        pipe.unet.set_default_attn_processor()
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
820
821
822
823
824
825
826
827
828
829
        pipe.vae = vae
        pipe.scheduler = LMSDiscreteScheduler.from_config(pipe.scheduler.config)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, 253:256, 253:256, -1].flatten()
        assert image.shape == (1, 512, 512, 3)
830
        expected_slice = np.array([0.8931, 0.8683, 0.8965, 0.8501, 0.8592, 0.9118, 0.8734, 0.7463, 0.8990])
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
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
        assert np.abs(expected_slice - image_slice).max() < 6e-3

    def test_stable_diffusion_inpaint_with_sequential_cpu_offloading(self):
        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()

        vae = AsymmetricAutoencoderKL.from_pretrained(
            "cross-attention/asymmetric-autoencoder-kl-x-1-5", torch_dtype=torch.float16
        )
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None, torch_dtype=torch.float16
        )
        pipe.vae = vae
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing(1)
        pipe.enable_sequential_cpu_offload()

        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        _ = pipe(**inputs)

        mem_bytes = torch.cuda.max_memory_allocated()
        # make sure that less than 2.45 GB is allocated
        assert mem_bytes < 2.45 * 10**9

Dhruv Nair's avatar
Dhruv Nair committed
857
    @require_python39_or_higher
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
    @require_torch_2
    def test_inpaint_compile(self):
        pass

    def test_stable_diffusion_inpaint_pil_input_resolution_test(self):
        vae = AsymmetricAutoencoderKL.from_pretrained(
            "cross-attention/asymmetric-autoencoder-kl-x-1-5",
        )
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
        pipe.vae = vae
        pipe.scheduler = LMSDiscreteScheduler.from_config(pipe.scheduler.config)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        inputs = self.get_inputs(torch_device)
        # change input image to a random size (one that would cause a tensor mismatch error)
        inputs["image"] = inputs["image"].resize((127, 127))
        inputs["mask_image"] = inputs["mask_image"].resize((127, 127))
        inputs["height"] = 128
        inputs["width"] = 128
        image = pipe(**inputs).images
        # verify that the returned image has the same height and width as the input height and width
        assert image.shape == (1, inputs["height"], inputs["width"], 3)

    def test_stable_diffusion_inpaint_strength_test(self):
        vae = AsymmetricAutoencoderKL.from_pretrained("cross-attention/asymmetric-autoencoder-kl-x-1-5")
        pipe = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", safety_checker=None
        )
890
        pipe.unet.set_default_attn_processor()
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
        pipe.vae = vae
        pipe.scheduler = LMSDiscreteScheduler.from_config(pipe.scheduler.config)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        inputs = self.get_inputs(torch_device)
        # change input strength
        inputs["strength"] = 0.75
        image = pipe(**inputs).images
        # verify that the returned image has the same height and width as the input height and width
        assert image.shape == (1, 512, 512, 3)

        image_slice = image[0, 253:256, 253:256, -1].flatten()
        expected_slice = np.array([0.2458, 0.2576, 0.3124, 0.2679, 0.2669, 0.2796, 0.2872, 0.2975, 0.2661])
        assert np.abs(expected_slice - image_slice).max() < 3e-3

    def test_stable_diffusion_simple_inpaint_ddim(self):
        vae = AsymmetricAutoencoderKL.from_pretrained("cross-attention/asymmetric-autoencoder-kl-x-1-5")
        pipe = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", safety_checker=None)
        pipe.vae = vae
912
        pipe.unet.set_default_attn_processor()
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
913
914
915
916
917
918
919
920
921
922
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images

        image_slice = image[0, 253:256, 253:256, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
923
924
        expected_slice = np.array([0.3296, 0.4041, 0.4097, 0.4145, 0.4342, 0.4152, 0.4927, 0.4931, 0.4430])
        assert np.abs(expected_slice - image_slice).max() < 1e-3
Ruslan Vorovchenko's avatar
Ruslan Vorovchenko committed
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946

    def test_download_local(self):
        vae = AsymmetricAutoencoderKL.from_pretrained(
            "cross-attention/asymmetric-autoencoder-kl-x-1-5", torch_dtype=torch.float16
        )
        filename = hf_hub_download("runwayml/stable-diffusion-inpainting", filename="sd-v1-5-inpainting.ckpt")

        pipe = StableDiffusionInpaintPipeline.from_single_file(filename, torch_dtype=torch.float16)
        pipe.vae = vae
        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
        pipe.to("cuda")

        inputs = self.get_inputs(torch_device)
        inputs["num_inference_steps"] = 1
        image_out = pipe(**inputs).images[0]

        assert image_out.shape == (512, 512, 3)

    def test_download_ckpt_diff_format_is_same(self):
        pass


947
948
949
950
951
952
953
@nightly
@require_torch_gpu
class StableDiffusionInpaintPipelineNightlyTests(unittest.TestCase):
    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()
954

955
956
    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
        generator = torch.Generator(device=generator_device).manual_seed(seed)
957
        init_image = load_image(
958
959
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/input_bench_image.png"
960
961
        )
        mask_image = load_image(
962
963
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/input_bench_mask.png"
964
        )
965
966
967
968
969
970
971
972
973
974
        inputs = {
            "prompt": "Face of a yellow cat, high resolution, sitting on a park bench",
            "image": init_image,
            "mask_image": mask_image,
            "generator": generator,
            "num_inference_steps": 50,
            "guidance_scale": 7.5,
            "output_type": "numpy",
        }
        return inputs
975

976
977
978
979
    def test_inpaint_ddim(self):
        sd_pipe = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-inpainting")
        sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)
980

981
982
        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images[0]
983

984
985
986
987
988
989
        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/stable_diffusion_inpaint_ddim.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3
990

991
992
993
994
995
996
997
998
    def test_inpaint_pndm(self):
        sd_pipe = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-inpainting")
        sd_pipe.scheduler = PNDMScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images[0]
999

1000
1001
1002
        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/stable_diffusion_inpaint_pndm.npy"
1003
        )
1004
1005
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3
1006

1007
1008
1009
1010
1011
    def test_inpaint_lms(self):
        sd_pipe = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-inpainting")
        sd_pipe.scheduler = LMSDiscreteScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)
1012

1013
1014
        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images[0]
1015

1016
1017
1018
        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/stable_diffusion_inpaint_lms.npy"
1019
        )
1020
1021
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3
1022

1023
1024
1025
1026
1027
    def test_inpaint_dpm(self):
        sd_pipe = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-inpainting")
        sd_pipe.scheduler = DPMSolverMultistepScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)
1028

1029
1030
1031
        inputs = self.get_inputs(torch_device)
        inputs["num_inference_steps"] = 30
        image = sd_pipe(**inputs).images[0]
1032

1033
1034
1035
        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_inpaint/stable_diffusion_inpaint_dpm_multi.npy"
1036
        )
1037
1038
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3
1039

Patrick von Platen's avatar
Patrick von Platen committed
1040

1041
1042
class StableDiffusionInpaintingPrepareMaskAndMaskedImageTests(unittest.TestCase):
    def test_pil_inputs(self):
1043
1044
        height, width = 32, 32
        im = np.random.randint(0, 255, (height, width, 3), dtype=np.uint8)
1045
        im = Image.fromarray(im)
1046
        mask = np.random.randint(0, 255, (height, width), dtype=np.uint8) > 127.5
1047
1048
        mask = Image.fromarray((mask * 255).astype(np.uint8))

1049
        t_mask, t_masked, t_image = prepare_mask_and_masked_image(im, mask, height, width, return_image=True)
1050
1051
1052

        self.assertTrue(isinstance(t_mask, torch.Tensor))
        self.assertTrue(isinstance(t_masked, torch.Tensor))
1053
        self.assertTrue(isinstance(t_image, torch.Tensor))
1054
1055
1056

        self.assertEqual(t_mask.ndim, 4)
        self.assertEqual(t_masked.ndim, 4)
1057
        self.assertEqual(t_image.ndim, 4)
1058

1059
1060
        self.assertEqual(t_mask.shape, (1, 1, height, width))
        self.assertEqual(t_masked.shape, (1, 3, height, width))
1061
        self.assertEqual(t_image.shape, (1, 3, height, width))
1062
1063
1064

        self.assertTrue(t_mask.dtype == torch.float32)
        self.assertTrue(t_masked.dtype == torch.float32)
1065
        self.assertTrue(t_image.dtype == torch.float32)
1066
1067
1068
1069
1070

        self.assertTrue(t_mask.min() >= 0.0)
        self.assertTrue(t_mask.max() <= 1.0)
        self.assertTrue(t_masked.min() >= -1.0)
        self.assertTrue(t_masked.min() <= 1.0)
1071
1072
        self.assertTrue(t_image.min() >= -1.0)
        self.assertTrue(t_image.min() >= -1.0)
1073
1074
1075
1076

        self.assertTrue(t_mask.sum() > 0.0)

    def test_np_inputs(self):
1077
1078
1079
        height, width = 32, 32

        im_np = np.random.randint(0, 255, (height, width, 3), dtype=np.uint8)
1080
        im_pil = Image.fromarray(im_np)
Patrick von Platen's avatar
Patrick von Platen committed
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
        mask_np = (
            np.random.randint(
                0,
                255,
                (
                    height,
                    width,
                ),
                dtype=np.uint8,
            )
            > 127.5
        )
1093
1094
        mask_pil = Image.fromarray((mask_np * 255).astype(np.uint8))

1095
1096
1097
1098
1099
1100
        t_mask_np, t_masked_np, t_image_np = prepare_mask_and_masked_image(
            im_np, mask_np, height, width, return_image=True
        )
        t_mask_pil, t_masked_pil, t_image_pil = prepare_mask_and_masked_image(
            im_pil, mask_pil, height, width, return_image=True
        )
1101
1102
1103

        self.assertTrue((t_mask_np == t_mask_pil).all())
        self.assertTrue((t_masked_np == t_masked_pil).all())
1104
        self.assertTrue((t_image_np == t_image_pil).all())
1105
1106

    def test_torch_3D_2D_inputs(self):
1107
1108
        height, width = 32, 32

Patrick von Platen's avatar
Patrick von Platen committed
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
        im_tensor = torch.randint(
            0,
            255,
            (
                3,
                height,
                width,
            ),
            dtype=torch.uint8,
        )
        mask_tensor = (
            torch.randint(
                0,
                255,
                (
                    height,
                    width,
                ),
                dtype=torch.uint8,
            )
            > 127.5
        )
1131
1132
1133
        im_np = im_tensor.numpy().transpose(1, 2, 0)
        mask_np = mask_tensor.numpy()

1134
1135
1136
1137
1138
        t_mask_tensor, t_masked_tensor, t_image_tensor = prepare_mask_and_masked_image(
            im_tensor / 127.5 - 1, mask_tensor, height, width, return_image=True
        )
        t_mask_np, t_masked_np, t_image_np = prepare_mask_and_masked_image(
            im_np, mask_np, height, width, return_image=True
Patrick von Platen's avatar
Patrick von Platen committed
1139
        )
1140
1141
1142

        self.assertTrue((t_mask_tensor == t_mask_np).all())
        self.assertTrue((t_masked_tensor == t_masked_np).all())
1143
        self.assertTrue((t_image_tensor == t_image_np).all())
1144
1145

    def test_torch_3D_3D_inputs(self):
1146
1147
        height, width = 32, 32

Patrick von Platen's avatar
Patrick von Platen committed
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
        im_tensor = torch.randint(
            0,
            255,
            (
                3,
                height,
                width,
            ),
            dtype=torch.uint8,
        )
        mask_tensor = (
            torch.randint(
                0,
                255,
                (
                    1,
                    height,
                    width,
                ),
                dtype=torch.uint8,
            )
            > 127.5
        )
1171
1172
1173
        im_np = im_tensor.numpy().transpose(1, 2, 0)
        mask_np = mask_tensor.numpy()[0]

1174
1175
1176
1177
1178
        t_mask_tensor, t_masked_tensor, t_image_tensor = prepare_mask_and_masked_image(
            im_tensor / 127.5 - 1, mask_tensor, height, width, return_image=True
        )
        t_mask_np, t_masked_np, t_image_np = prepare_mask_and_masked_image(
            im_np, mask_np, height, width, return_image=True
Patrick von Platen's avatar
Patrick von Platen committed
1179
        )
1180
1181
1182

        self.assertTrue((t_mask_tensor == t_mask_np).all())
        self.assertTrue((t_masked_tensor == t_masked_np).all())
1183
        self.assertTrue((t_image_tensor == t_image_np).all())
1184
1185

    def test_torch_4D_2D_inputs(self):
1186
1187
        height, width = 32, 32

Patrick von Platen's avatar
Patrick von Platen committed
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
        im_tensor = torch.randint(
            0,
            255,
            (
                1,
                3,
                height,
                width,
            ),
            dtype=torch.uint8,
        )
        mask_tensor = (
            torch.randint(
                0,
                255,
                (
                    height,
                    width,
                ),
                dtype=torch.uint8,
            )
            > 127.5
        )
1211
1212
1213
        im_np = im_tensor.numpy()[0].transpose(1, 2, 0)
        mask_np = mask_tensor.numpy()

1214
1215
1216
1217
1218
        t_mask_tensor, t_masked_tensor, t_image_tensor = prepare_mask_and_masked_image(
            im_tensor / 127.5 - 1, mask_tensor, height, width, return_image=True
        )
        t_mask_np, t_masked_np, t_image_np = prepare_mask_and_masked_image(
            im_np, mask_np, height, width, return_image=True
Patrick von Platen's avatar
Patrick von Platen committed
1219
        )
1220
1221
1222

        self.assertTrue((t_mask_tensor == t_mask_np).all())
        self.assertTrue((t_masked_tensor == t_masked_np).all())
1223
        self.assertTrue((t_image_tensor == t_image_np).all())
1224
1225

    def test_torch_4D_3D_inputs(self):
1226
1227
        height, width = 32, 32

Patrick von Platen's avatar
Patrick von Platen committed
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
        im_tensor = torch.randint(
            0,
            255,
            (
                1,
                3,
                height,
                width,
            ),
            dtype=torch.uint8,
        )
        mask_tensor = (
            torch.randint(
                0,
                255,
                (
                    1,
                    height,
                    width,
                ),
                dtype=torch.uint8,
            )
            > 127.5
        )
1252
1253
1254
        im_np = im_tensor.numpy()[0].transpose(1, 2, 0)
        mask_np = mask_tensor.numpy()[0]

1255
1256
1257
1258
1259
        t_mask_tensor, t_masked_tensor, t_image_tensor = prepare_mask_and_masked_image(
            im_tensor / 127.5 - 1, mask_tensor, height, width, return_image=True
        )
        t_mask_np, t_masked_np, t_image_np = prepare_mask_and_masked_image(
            im_np, mask_np, height, width, return_image=True
Patrick von Platen's avatar
Patrick von Platen committed
1260
        )
1261
1262
1263

        self.assertTrue((t_mask_tensor == t_mask_np).all())
        self.assertTrue((t_masked_tensor == t_masked_np).all())
1264
        self.assertTrue((t_image_tensor == t_image_np).all())
1265
1266

    def test_torch_4D_4D_inputs(self):
1267
1268
        height, width = 32, 32

Patrick von Platen's avatar
Patrick von Platen committed
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
        im_tensor = torch.randint(
            0,
            255,
            (
                1,
                3,
                height,
                width,
            ),
            dtype=torch.uint8,
        )
        mask_tensor = (
            torch.randint(
                0,
                255,
                (
                    1,
                    1,
                    height,
                    width,
                ),
                dtype=torch.uint8,
            )
            > 127.5
        )
1294
1295
1296
        im_np = im_tensor.numpy()[0].transpose(1, 2, 0)
        mask_np = mask_tensor.numpy()[0][0]

1297
1298
1299
1300
1301
        t_mask_tensor, t_masked_tensor, t_image_tensor = prepare_mask_and_masked_image(
            im_tensor / 127.5 - 1, mask_tensor, height, width, return_image=True
        )
        t_mask_np, t_masked_np, t_image_np = prepare_mask_and_masked_image(
            im_np, mask_np, height, width, return_image=True
Patrick von Platen's avatar
Patrick von Platen committed
1302
        )
1303
1304
1305

        self.assertTrue((t_mask_tensor == t_mask_np).all())
        self.assertTrue((t_masked_tensor == t_masked_np).all())
1306
        self.assertTrue((t_image_tensor == t_image_np).all())
1307
1308

    def test_torch_batch_4D_3D(self):
1309
1310
        height, width = 32, 32

Patrick von Platen's avatar
Patrick von Platen committed
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
        im_tensor = torch.randint(
            0,
            255,
            (
                2,
                3,
                height,
                width,
            ),
            dtype=torch.uint8,
        )
        mask_tensor = (
            torch.randint(
                0,
                255,
                (
                    2,
                    height,
                    width,
                ),
                dtype=torch.uint8,
            )
            > 127.5
        )
1335
1336
1337
1338

        im_nps = [im.numpy().transpose(1, 2, 0) for im in im_tensor]
        mask_nps = [mask.numpy() for mask in mask_tensor]

1339
1340
        t_mask_tensor, t_masked_tensor, t_image_tensor = prepare_mask_and_masked_image(
            im_tensor / 127.5 - 1, mask_tensor, height, width, return_image=True
Patrick von Platen's avatar
Patrick von Platen committed
1341
        )
1342
        nps = [prepare_mask_and_masked_image(i, m, height, width, return_image=True) for i, m in zip(im_nps, mask_nps)]
1343
1344
        t_mask_np = torch.cat([n[0] for n in nps])
        t_masked_np = torch.cat([n[1] for n in nps])
1345
        t_image_np = torch.cat([n[2] for n in nps])
1346
1347
1348

        self.assertTrue((t_mask_tensor == t_mask_np).all())
        self.assertTrue((t_masked_tensor == t_masked_np).all())
1349
        self.assertTrue((t_image_tensor == t_image_np).all())
1350
1351

    def test_torch_batch_4D_4D(self):
1352
1353
        height, width = 32, 32

Patrick von Platen's avatar
Patrick von Platen committed
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
        im_tensor = torch.randint(
            0,
            255,
            (
                2,
                3,
                height,
                width,
            ),
            dtype=torch.uint8,
        )
        mask_tensor = (
            torch.randint(
                0,
                255,
                (
                    2,
                    1,
                    height,
                    width,
                ),
                dtype=torch.uint8,
            )
            > 127.5
        )
1379
1380
1381
1382

        im_nps = [im.numpy().transpose(1, 2, 0) for im in im_tensor]
        mask_nps = [mask.numpy()[0] for mask in mask_tensor]

1383
1384
        t_mask_tensor, t_masked_tensor, t_image_tensor = prepare_mask_and_masked_image(
            im_tensor / 127.5 - 1, mask_tensor, height, width, return_image=True
Patrick von Platen's avatar
Patrick von Platen committed
1385
        )
1386
        nps = [prepare_mask_and_masked_image(i, m, height, width, return_image=True) for i, m in zip(im_nps, mask_nps)]
1387
1388
        t_mask_np = torch.cat([n[0] for n in nps])
        t_masked_np = torch.cat([n[1] for n in nps])
1389
        t_image_np = torch.cat([n[2] for n in nps])
1390
1391
1392

        self.assertTrue((t_mask_tensor == t_mask_np).all())
        self.assertTrue((t_masked_tensor == t_masked_np).all())
1393
        self.assertTrue((t_image_tensor == t_image_np).all())
1394
1395

    def test_shape_mismatch(self):
1396
1397
        height, width = 32, 32

1398
1399
        # test height and width
        with self.assertRaises(AssertionError):
Patrick von Platen's avatar
Patrick von Platen committed
1400
1401
1402
1403
1404
1405
1406
1407
1408
            prepare_mask_and_masked_image(
                torch.randn(
                    3,
                    height,
                    width,
                ),
                torch.randn(64, 64),
                height,
                width,
1409
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1410
            )
1411
1412
        # test batch dim
        with self.assertRaises(AssertionError):
Patrick von Platen's avatar
Patrick von Platen committed
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
            prepare_mask_and_masked_image(
                torch.randn(
                    2,
                    3,
                    height,
                    width,
                ),
                torch.randn(4, 64, 64),
                height,
                width,
1423
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1424
            )
1425
1426
        # test batch dim
        with self.assertRaises(AssertionError):
Patrick von Platen's avatar
Patrick von Platen committed
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
            prepare_mask_and_masked_image(
                torch.randn(
                    2,
                    3,
                    height,
                    width,
                ),
                torch.randn(4, 1, 64, 64),
                height,
                width,
1437
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1438
            )
1439
1440

    def test_type_mismatch(self):
1441
1442
        height, width = 32, 32

1443
1444
        # test tensors-only
        with self.assertRaises(TypeError):
Patrick von Platen's avatar
Patrick von Platen committed
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
            prepare_mask_and_masked_image(
                torch.rand(
                    3,
                    height,
                    width,
                ),
                torch.rand(
                    3,
                    height,
                    width,
                ).numpy(),
                height,
                width,
1458
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1459
            )
1460
1461
        # test tensors-only
        with self.assertRaises(TypeError):
Patrick von Platen's avatar
Patrick von Platen committed
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
            prepare_mask_and_masked_image(
                torch.rand(
                    3,
                    height,
                    width,
                ).numpy(),
                torch.rand(
                    3,
                    height,
                    width,
                ),
                height,
                width,
1475
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1476
            )
1477
1478

    def test_channels_first(self):
1479
1480
        height, width = 32, 32

1481
1482
        # test channels first for 3D tensors
        with self.assertRaises(AssertionError):
Patrick von Platen's avatar
Patrick von Platen committed
1483
1484
1485
1486
1487
1488
1489
1490
1491
            prepare_mask_and_masked_image(
                torch.rand(height, width, 3),
                torch.rand(
                    3,
                    height,
                    width,
                ),
                height,
                width,
1492
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1493
            )
1494
1495

    def test_tensor_range(self):
1496
1497
        height, width = 32, 32

1498
1499
        # test im <= 1
        with self.assertRaises(ValueError):
Patrick von Platen's avatar
Patrick von Platen committed
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
            prepare_mask_and_masked_image(
                torch.ones(
                    3,
                    height,
                    width,
                )
                * 2,
                torch.rand(
                    height,
                    width,
                ),
                height,
                width,
1513
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1514
            )
1515
1516
        # test im >= -1
        with self.assertRaises(ValueError):
Patrick von Platen's avatar
Patrick von Platen committed
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
            prepare_mask_and_masked_image(
                torch.ones(
                    3,
                    height,
                    width,
                )
                * (-2),
                torch.rand(
                    height,
                    width,
                ),
                height,
                width,
1530
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1531
            )
1532
1533
        # test mask <= 1
        with self.assertRaises(ValueError):
Patrick von Platen's avatar
Patrick von Platen committed
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
            prepare_mask_and_masked_image(
                torch.rand(
                    3,
                    height,
                    width,
                ),
                torch.ones(
                    height,
                    width,
                )
                * 2,
                height,
                width,
1547
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1548
            )
1549
1550
        # test mask >= 0
        with self.assertRaises(ValueError):
Patrick von Platen's avatar
Patrick von Platen committed
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
            prepare_mask_and_masked_image(
                torch.rand(
                    3,
                    height,
                    width,
                ),
                torch.ones(
                    height,
                    width,
                )
                * -1,
                height,
                width,
1564
                return_image=True,
Patrick von Platen's avatar
Patrick von Platen committed
1565
            )