test_stable_diffusion.py 24.6 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
18
19
20
#
# 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 unittest

import numpy as np
import torch
21
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
22
23
24
25

from diffusers import (
    AutoencoderKL,
    DDIMScheduler,
26
    DPMSolverMultistepScheduler,
27
28
29
30
31
32
33
34
    EulerAncestralDiscreteScheduler,
    EulerDiscreteScheduler,
    LMSDiscreteScheduler,
    PNDMScheduler,
    StableDiffusionPipeline,
    UNet2DConditionModel,
    logging,
)
35
36
37
from diffusers.utils.testing_utils import (
    CaptureLogger,
    enable_full_determinism,
Dhruv Nair's avatar
Dhruv Nair committed
38
39
    load_numpy,
    nightly,
40
41
    numpy_cosine_similarity_distance,
    require_torch_gpu,
Dhruv Nair's avatar
Dhruv Nair committed
42
43
    slow,
    torch_device,
44
)
45

46
47
48
49
50
51
from ..pipeline_params import (
    TEXT_TO_IMAGE_BATCH_PARAMS,
    TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS,
    TEXT_TO_IMAGE_IMAGE_PARAMS,
    TEXT_TO_IMAGE_PARAMS,
)
52
from ..test_pipelines_common import PipelineKarrasSchedulerTesterMixin, PipelineLatentTesterMixin, PipelineTesterMixin
53
54


55
enable_full_determinism()
56
57


58
59
60
class StableDiffusion2PipelineFastTests(
    PipelineLatentTesterMixin, PipelineKarrasSchedulerTesterMixin, PipelineTesterMixin, unittest.TestCase
):
61
    pipeline_class = StableDiffusionPipeline
62
63
    params = TEXT_TO_IMAGE_PARAMS
    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
64
    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
65
    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
66
    callback_cfg_params = TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS
67

68
    def get_dummy_components(self):
69
        torch.manual_seed(0)
70
        unet = UNet2DConditionModel(
71
72
73
74
75
76
77
78
79
            block_out_channels=(32, 64),
            layers_per_block=2,
            sample_size=32,
            in_channels=4,
            out_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            cross_attention_dim=32,
            # SD2-specific config below
Will Berman's avatar
Will Berman committed
80
            attention_head_dim=(2, 4),
81
82
            use_linear_projection=True,
        )
83
84
85
86
87
88
89
        scheduler = DDIMScheduler(
            beta_start=0.00085,
            beta_end=0.012,
            beta_schedule="scaled_linear",
            clip_sample=False,
            set_alpha_to_one=False,
        )
90
        torch.manual_seed(0)
91
        vae = AutoencoderKL(
92
93
94
95
96
97
98
99
100
            block_out_channels=[32, 64],
            in_channels=3,
            out_channels=3,
            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
            latent_channels=4,
            sample_size=128,
        )
        torch.manual_seed(0)
101
        text_encoder_config = CLIPTextConfig(
102
103
104
105
106
107
108
109
110
111
112
113
114
            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,
            # SD2-specific config below
            hidden_act="gelu",
            projection_dim=512,
        )
115
        text_encoder = CLIPTextModel(text_encoder_config)
116
117
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
        components = {
            "unet": unet,
            "scheduler": scheduler,
            "vae": vae,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "safety_checker": None,
            "feature_extractor": None,
        }
        return components

    def get_dummy_inputs(self, device, seed=0):
        if str(device).startswith("mps"):
            generator = torch.manual_seed(seed)
        else:
            generator = torch.Generator(device=device).manual_seed(seed)
        inputs = {
            "prompt": "A painting of a squirrel eating a burger",
            "generator": generator,
            "num_inference_steps": 2,
            "guidance_scale": 6.0,
            "output_type": "numpy",
        }
        return inputs
142
143
144

    def test_stable_diffusion_ddim(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
145
146
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionPipeline(**components)
147
148
149
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

150
151
        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
152
153
154
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
155
        expected_slice = np.array([0.5753, 0.6113, 0.5005, 0.5036, 0.5464, 0.4725, 0.4982, 0.4865, 0.4861])
156
157
158
159
160

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

    def test_stable_diffusion_pndm(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
161
162
163
        components = self.get_dummy_components()
        components["scheduler"] = PNDMScheduler(skip_prk_steps=True)
        sd_pipe = StableDiffusionPipeline(**components)
164
165
166
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

167
168
        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
169
170
171
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
172
        expected_slice = np.array([0.5121, 0.5714, 0.4827, 0.5057, 0.5646, 0.4766, 0.5189, 0.4895, 0.4990])
173

174
175
176
177
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_k_lms(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
178
179
180
        components = self.get_dummy_components()
        components["scheduler"] = LMSDiscreteScheduler.from_config(components["scheduler"].config)
        sd_pipe = StableDiffusionPipeline(**components)
181
182
183
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

184
185
        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
186
187
188
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
189
        expected_slice = np.array([0.4865, 0.5439, 0.4840, 0.4995, 0.5543, 0.4846, 0.5199, 0.4942, 0.5061])
190

191
192
193
194
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_k_euler_ancestral(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
195
196
197
        components = self.get_dummy_components()
        components["scheduler"] = EulerAncestralDiscreteScheduler.from_config(components["scheduler"].config)
        sd_pipe = StableDiffusionPipeline(**components)
198
199
200
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

201
202
        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
203
204
205
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
206
        expected_slice = np.array([0.4864, 0.5440, 0.4842, 0.4994, 0.5543, 0.4846, 0.5196, 0.4942, 0.5063])
207

208
209
210
211
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_k_euler(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
212
213
214
        components = self.get_dummy_components()
        components["scheduler"] = EulerDiscreteScheduler.from_config(components["scheduler"].config)
        sd_pipe = StableDiffusionPipeline(**components)
215
216
217
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

218
219
        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
220
221
222
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
223
        expected_slice = np.array([0.4865, 0.5439, 0.4840, 0.4995, 0.5543, 0.4846, 0.5199, 0.4942, 0.5061])
224

225
226
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
    def test_stable_diffusion_unflawed(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        components["scheduler"] = DDIMScheduler.from_config(
            components["scheduler"].config, timestep_spacing="trailing"
        )
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        inputs["guidance_rescale"] = 0.7
        inputs["num_inference_steps"] = 10
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
        expected_slice = np.array([0.4736, 0.5405, 0.4705, 0.4955, 0.5675, 0.4812, 0.5310, 0.4967, 0.5064])

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

248
    def test_stable_diffusion_long_prompt(self):
249
250
251
        components = self.get_dummy_components()
        components["scheduler"] = LMSDiscreteScheduler.from_config(components["scheduler"].config)
        sd_pipe = StableDiffusionPipeline(**components)
252
253
254
255
256
257
258
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        do_classifier_free_guidance = True
        negative_prompt = None
        num_images_per_prompt = 1
        logger = logging.get_logger("diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion")
259
        logger.setLevel(logging.WARNING)
260
261
262

        prompt = 25 * "@"
        with CaptureLogger(logger) as cap_logger_3:
263
            text_embeddings_3, negeative_text_embeddings_3 = sd_pipe.encode_prompt(
264
265
                prompt, torch_device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
            )
266
267
            if negeative_text_embeddings_3 is not None:
                text_embeddings_3 = torch.cat([negeative_text_embeddings_3, text_embeddings_3])
268
269
270

        prompt = 100 * "@"
        with CaptureLogger(logger) as cap_logger:
271
            text_embeddings, negative_embeddings = sd_pipe.encode_prompt(
272
273
                prompt, torch_device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
            )
274
275
            if negative_embeddings is not None:
                text_embeddings = torch.cat([negative_embeddings, text_embeddings])
276
277
278

        negative_prompt = "Hello"
        with CaptureLogger(logger) as cap_logger_2:
279
            text_embeddings_2, negative_text_embeddings_2 = sd_pipe.encode_prompt(
280
281
                prompt, torch_device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
            )
282
283
            if negative_text_embeddings_2 is not None:
                text_embeddings_2 = torch.cat([negative_text_embeddings_2, text_embeddings_2])
284
285
286
287
288
289
290
291
292

        assert text_embeddings_3.shape == text_embeddings_2.shape == text_embeddings.shape
        assert text_embeddings.shape[1] == 77

        assert cap_logger.out == cap_logger_2.out
        # 100 - 77 + 1 (BOS token) + 1 (EOS token) = 25
        assert cap_logger.out.count("@") == 25
        assert cap_logger_3.out == ""

293
294
295
296
297
298
    def test_attention_slicing_forward_pass(self):
        super().test_attention_slicing_forward_pass(expected_max_diff=3e-3)

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

299
300
301

@slow
@require_torch_gpu
302
class StableDiffusion2PipelineSlowTests(unittest.TestCase):
303
304
305
306
307
    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

308
309
    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
        generator = torch.Generator(device=generator_device).manual_seed(seed)
310
311
312
313
314
315
316
317
318
319
320
        latents = np.random.RandomState(seed).standard_normal((1, 4, 64, 64))
        latents = torch.from_numpy(latents).to(device=device, dtype=dtype)
        inputs = {
            "prompt": "a photograph of an astronaut riding a horse",
            "latents": latents,
            "generator": generator,
            "num_inference_steps": 3,
            "guidance_scale": 7.5,
            "output_type": "numpy",
        }
        return inputs
321

322
323
324
325
    def test_stable_diffusion_default_ddim(self):
        pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-base")
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
326

327
328
329
        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()
330
331

        assert image.shape == (1, 512, 512, 3)
332
        expected_slice = np.array([0.49493, 0.47896, 0.40798, 0.54214, 0.53212, 0.48202, 0.47656, 0.46329, 0.48506])
333
        assert np.abs(image_slice - expected_slice).max() < 7e-3
334

335
336
337
338
339
    def test_stable_diffusion_pndm(self):
        pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-base")
        pipe.scheduler = PNDMScheduler.from_config(pipe.scheduler.config)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
340

341
342
343
        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()
344
345

        assert image.shape == (1, 512, 512, 3)
346
        expected_slice = np.array([0.49493, 0.47896, 0.40798, 0.54214, 0.53212, 0.48202, 0.47656, 0.46329, 0.48506])
347
        assert np.abs(image_slice - expected_slice).max() < 7e-3
348
349

    def test_stable_diffusion_k_lms(self):
350
351
352
353
        pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-base")
        pipe.scheduler = LMSDiscreteScheduler.from_config(pipe.scheduler.config)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
354

355
356
357
        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()
358
359

        assert image.shape == (1, 512, 512, 3)
360
        expected_slice = np.array([0.10440, 0.13115, 0.11100, 0.10141, 0.11440, 0.07215, 0.11332, 0.09693, 0.10006])
361
        assert np.abs(image_slice - expected_slice).max() < 3e-3
362

363
    def test_stable_diffusion_attention_slicing(self):
364
        torch.cuda.reset_peak_memory_stats()
365
        pipe = StableDiffusionPipeline.from_pretrained(
366
            "stabilityai/stable-diffusion-2-base", torch_dtype=torch.float16
367
        )
368
        pipe.unet.set_default_attn_processor()
369
        pipe = pipe.to(torch_device)
370
371
        pipe.set_progress_bar_config(disable=None)

372
        # enable attention slicing
373
        pipe.enable_attention_slicing()
374
375
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        image_sliced = pipe(**inputs).images
376
377
378

        mem_bytes = torch.cuda.max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()
379
380
        # make sure that less than 3.3 GB is allocated
        assert mem_bytes < 3.3 * 10**9
381

382
        # disable slicing
383
        pipe.disable_attention_slicing()
384
        pipe.unet.set_default_attn_processor()
385
386
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        image = pipe(**inputs).images
387

388
        # make sure that more than 3.3 GB is allocated
389
        mem_bytes = torch.cuda.max_memory_allocated()
390
        assert mem_bytes > 3.3 * 10**9
391
392
        max_diff = numpy_cosine_similarity_distance(image.flatten(), image_sliced.flatten())
        assert max_diff < 5e-3
393
394
395
396

    def test_stable_diffusion_text2img_intermediate_state(self):
        number_of_steps = 0

397
398
        def callback_fn(step: int, timestep: int, latents: torch.FloatTensor) -> None:
            callback_fn.has_been_called = True
399
400
            nonlocal number_of_steps
            number_of_steps += 1
401
            if step == 1:
402
403
404
                latents = latents.detach().cpu().numpy()
                assert latents.shape == (1, 4, 64, 64)
                latents_slice = latents[0, -3:, -3:, -1]
405
406
407
408
409
                expected_slice = np.array(
                    [-0.3862, -0.4507, -1.1729, 0.0686, -1.1045, 0.7124, -1.8301, 0.1903, 1.2773]
                )

                assert np.abs(latents_slice.flatten() - expected_slice).max() < 5e-2
410
            elif step == 2:
411
412
413
                latents = latents.detach().cpu().numpy()
                assert latents.shape == (1, 4, 64, 64)
                latents_slice = latents[0, -3:, -3:, -1]
414
415
416
417
418
                expected_slice = np.array(
                    [0.2720, -0.1863, -0.7383, -0.5029, -0.7534, 0.3970, -0.7646, 0.4468, 1.2686]
                )

                assert np.abs(latents_slice.flatten() - expected_slice).max() < 5e-2
419

420
        callback_fn.has_been_called = False
421
422

        pipe = StableDiffusionPipeline.from_pretrained(
423
            "stabilityai/stable-diffusion-2-base", torch_dtype=torch.float16
424
425
426
427
428
        )
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

429
430
431
432
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        pipe(**inputs, callback=callback_fn, callback_steps=1)
        assert callback_fn.has_been_called
        assert number_of_steps == inputs["num_inference_steps"]
433
434
435
436
437
438

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

439
        pipe = StableDiffusionPipeline.from_pretrained(
440
            "stabilityai/stable-diffusion-2-base", torch_dtype=torch.float16
441
442
443
444
445
        )
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing(1)
        pipe.enable_sequential_cpu_offload()
446

447
448
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        _ = pipe(**inputs)
449
450
451
452

        mem_bytes = torch.cuda.max_memory_allocated()
        # make sure that less than 2.8 GB is allocated
        assert mem_bytes < 2.8 * 10**9
453

454
455
456
457
458
459
460
461
462
463
464
465
466
    def test_stable_diffusion_pipeline_with_model_offloading(self):
        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()

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

        # Normal inference

        pipe = StableDiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-2-base",
            torch_dtype=torch.float16,
        )
467
        pipe.unet.set_default_attn_processor()
468
469
470
471
472
473
474
475
476
477
478
479
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        outputs = pipe(**inputs)
        mem_bytes = torch.cuda.max_memory_allocated()

        # With model offloading

        # Reload but don't move to cuda
        pipe = StableDiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-2-base",
            torch_dtype=torch.float16,
        )
480
        pipe.unet.set_default_attn_processor()
481
482
483
484
485
486
487

        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()

        pipe.enable_model_cpu_offload()
        pipe.set_progress_bar_config(disable=None)
488
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
489
490
491
        outputs_offloaded = pipe(**inputs)
        mem_bytes_offloaded = torch.cuda.max_memory_allocated()

492
493
494
495
        images = outputs.images
        images_offloaded = outputs_offloaded.images
        max_diff = numpy_cosine_similarity_distance(images.flatten(), images_offloaded.flatten())
        assert max_diff < 1e-3
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
        assert mem_bytes_offloaded < mem_bytes
        assert mem_bytes_offloaded < 3 * 10**9
        for module in pipe.text_encoder, pipe.unet, pipe.vae:
            assert module.device == torch.device("cpu")

        # With attention slicing
        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()

        pipe.enable_attention_slicing()
        _ = pipe(**inputs)
        mem_bytes_slicing = torch.cuda.max_memory_allocated()
        assert mem_bytes_slicing < mem_bytes_offloaded

511
512
513
514
515
516
517
518
519

@nightly
@require_torch_gpu
class StableDiffusion2PipelineNightlyTests(unittest.TestCase):
    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

520
521
    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
        generator = torch.Generator(device=generator_device).manual_seed(seed)
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
        latents = np.random.RandomState(seed).standard_normal((1, 4, 64, 64))
        latents = torch.from_numpy(latents).to(device=device, dtype=dtype)
        inputs = {
            "prompt": "a photograph of an astronaut riding a horse",
            "latents": latents,
            "generator": generator,
            "num_inference_steps": 50,
            "guidance_scale": 7.5,
            "output_type": "numpy",
        }
        return inputs

    def test_stable_diffusion_2_0_default_ddim(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-base").to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

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

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_2_text2img/stable_diffusion_2_0_base_ddim.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

    def test_stable_diffusion_2_1_default_pndm(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1-base").to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

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

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_2_text2img/stable_diffusion_2_1_base_pndm.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

    def test_stable_diffusion_ddim(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1-base").to(torch_device)
        sd_pipe.scheduler = DDIMScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.set_progress_bar_config(disable=None)

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

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_2_text2img/stable_diffusion_2_1_base_ddim.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

    def test_stable_diffusion_lms(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1-base").to(torch_device)
        sd_pipe.scheduler = LMSDiscreteScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.set_progress_bar_config(disable=None)

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

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_2_text2img/stable_diffusion_2_1_base_lms.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

    def test_stable_diffusion_euler(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1-base").to(torch_device)
        sd_pipe.scheduler = EulerDiscreteScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.set_progress_bar_config(disable=None)

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

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_2_text2img/stable_diffusion_2_1_base_euler.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

    def test_stable_diffusion_dpm(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1-base").to(torch_device)
        sd_pipe.scheduler = DPMSolverMultistepScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        inputs["num_inference_steps"] = 25
        image = sd_pipe(**inputs).images[0]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_2_text2img/stable_diffusion_2_1_base_dpm_multi.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3