test_stable_diffusion_img2img.py 22.5 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
21
#
# 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
import unittest

import numpy as np
import torch
22
from packaging import version
23
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
24
25
26

from diffusers import (
    AutoencoderKL,
27
    DDIMScheduler,
28
    DPMSolverMultistepScheduler,
29
    HeunDiscreteScheduler,
30
31
32
33
34
    LMSDiscreteScheduler,
    PNDMScheduler,
    StableDiffusionImg2ImgPipeline,
    UNet2DConditionModel,
)
YiYi Xu's avatar
YiYi Xu committed
35
from diffusers.image_processor import VaeImageProcessor
36
from diffusers.utils import floats_tensor, load_image, load_numpy, nightly, slow, torch_device
37
from diffusers.utils.testing_utils import enable_full_determinism, require_torch_gpu, skip_mps
38

39
40
41
42
43
44
from ..pipeline_params import (
    IMAGE_TO_IMAGE_IMAGE_PARAMS,
    TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS,
    TEXT_GUIDED_IMAGE_VARIATION_PARAMS,
)
from ..test_pipelines_common import PipelineLatentTesterMixin, PipelineTesterMixin
45

46

47
enable_full_determinism()
48
49


50
class StableDiffusionImg2ImgPipelineFastTests(PipelineLatentTesterMixin, PipelineTesterMixin, unittest.TestCase):
51
    pipeline_class = StableDiffusionImg2ImgPipeline
52
53
54
    params = TEXT_GUIDED_IMAGE_VARIATION_PARAMS - {"height", "width"}
    required_optional_params = PipelineTesterMixin.required_optional_params - {"latents"}
    batch_params = TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS
55
    image_params = IMAGE_TO_IMAGE_IMAGE_PARAMS
56

57
    def get_dummy_components(self):
58
        torch.manual_seed(0)
59
        unet = UNet2DConditionModel(
60
61
62
63
64
65
66
67
68
            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,
        )
69
        scheduler = PNDMScheduler(skip_prk_steps=True)
70
        torch.manual_seed(0)
71
        vae = AutoencoderKL(
72
73
74
75
76
77
78
79
            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)
80
        text_encoder_config = CLIPTextConfig(
81
82
83
84
85
86
87
88
89
90
            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,
        )
91
92
        text_encoder = CLIPTextModel(text_encoder_config)
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
93

94
95
96
97
98
99
100
        components = {
            "unet": unet,
            "scheduler": scheduler,
            "vae": vae,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "safety_checker": None,
101
            "feature_extractor": None,
102
103
104
        }
        return components

105
    def get_dummy_inputs(self, device, seed=0):
106
107
108
109
110
111
112
        image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
        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",
113
            "image": image,
114
115
116
            "generator": generator,
            "num_inference_steps": 2,
            "guidance_scale": 6.0,
117
            "output_type": "numpy",
118
119
        }
        return inputs
120

121
    def test_stable_diffusion_img2img_default_case(self):
122
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
123
124
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionImg2ImgPipeline(**components)
125
        sd_pipe.image_processor = VaeImageProcessor(vae_scale_factor=sd_pipe.vae_scale_factor, do_normalize=True)
126
127
128
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

129
        inputs = self.get_dummy_inputs(device)
130
        inputs["image"] = inputs["image"] / 2 + 0.5
131
        image = sd_pipe(**inputs).images
132
133
134
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 32, 32, 3)
135
        expected_slice = np.array([0.4555, 0.3216, 0.4049, 0.4620, 0.4618, 0.4126, 0.4122, 0.4629, 0.4579])
136

137
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
138
139
140

    def test_stable_diffusion_img2img_negative_prompt(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
141
142
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionImg2ImgPipeline(**components)
143
        sd_pipe.image_processor = VaeImageProcessor(vae_scale_factor=sd_pipe.vae_scale_factor, do_normalize=True)
144
145
146
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

147
        inputs = self.get_dummy_inputs(device)
148
        inputs["image"] = inputs["image"] / 2 + 0.5
149
        negative_prompt = "french fries"
150
        output = sd_pipe(**inputs, negative_prompt=negative_prompt)
151
152
153
154
        image = output.images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 32, 32, 3)
155
        expected_slice = np.array([0.4593, 0.3408, 0.4232, 0.4749, 0.4476, 0.4115, 0.4357, 0.4733, 0.4663])
156

157
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
158
159
160

    def test_stable_diffusion_img2img_multiple_init_images(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
161
162
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionImg2ImgPipeline(**components)
163
        sd_pipe.image_processor = VaeImageProcessor(vae_scale_factor=sd_pipe.vae_scale_factor, do_normalize=True)
164
165
166
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

167
168
169
        inputs = self.get_dummy_inputs(device)
        inputs["prompt"] = [inputs["prompt"]] * 2
        inputs["image"] = inputs["image"].repeat(2, 1, 1, 1)
170
        inputs["image"] = inputs["image"] / 2 + 0.5
171
        image = sd_pipe(**inputs).images
172
173
174
        image_slice = image[-1, -3:, -3:, -1]

        assert image.shape == (2, 32, 32, 3)
175
        expected_slice = np.array([0.4241, 0.5576, 0.5711, 0.4792, 0.4311, 0.5952, 0.5827, 0.5138, 0.5109])
176

177
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
178
179
180

    def test_stable_diffusion_img2img_k_lms(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
181
182
183
        components = self.get_dummy_components()
        components["scheduler"] = LMSDiscreteScheduler(
            beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear"
184
        )
185
        sd_pipe = StableDiffusionImg2ImgPipeline(**components)
186
        sd_pipe.image_processor = VaeImageProcessor(vae_scale_factor=sd_pipe.vae_scale_factor, do_normalize=True)
187
188
189
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

190
        inputs = self.get_dummy_inputs(device)
191
        inputs["image"] = inputs["image"] / 2 + 0.5
192
        image = sd_pipe(**inputs).images
193
194
195
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 32, 32, 3)
196
        expected_slice = np.array([0.4398, 0.4949, 0.4337, 0.6580, 0.5555, 0.4338, 0.5769, 0.5955, 0.5175])
197

198
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
199

200
201
202
203
204
205
206
207
208
209
210
211
212
213
    @skip_mps
    def test_save_load_local(self):
        return super().test_save_load_local()

    @skip_mps
    def test_dict_tuple_outputs_equivalent(self):
        return super().test_dict_tuple_outputs_equivalent()

    @skip_mps
    def test_save_load_optional_components(self):
        return super().test_save_load_optional_components()

    @skip_mps
    def test_attention_slicing_forward_pass(self):
214
215
216
217
        return super().test_attention_slicing_forward_pass(expected_max_diff=5e-3)

    def test_inference_batch_single_identical(self):
        super().test_inference_batch_single_identical(expected_max_diff=3e-3)
218

219
220

@slow
221
@require_torch_gpu
222
class StableDiffusionImg2ImgPipelineSlowTests(unittest.TestCase):
223
224
225
226
227
    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

228
229
    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
        generator = torch.Generator(device=generator_device).manual_seed(seed)
230
        init_image = load_image(
231
232
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_img2img/sketch-mountains-input.png"
233
        )
234
235
236
237
238
239
240
        inputs = {
            "prompt": "a fantasy landscape, concept art, high resolution",
            "image": init_image,
            "generator": generator,
            "num_inference_steps": 3,
            "strength": 0.75,
            "guidance_scale": 7.5,
YiYi Xu's avatar
YiYi Xu committed
241
            "output_type": "np",
242
243
        }
        return inputs
244

245
246
    def test_stable_diffusion_img2img_default(self):
        pipe = StableDiffusionImg2ImgPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", safety_checker=None)
247
248
249
250
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

251
252
253
        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()
254

255
        assert image.shape == (1, 512, 768, 3)
256
257
        expected_slice = np.array([0.4300, 0.4662, 0.4930, 0.3990, 0.4307, 0.4525, 0.3719, 0.4064, 0.3923])

258
        assert np.abs(expected_slice - image_slice).max() < 1e-3
259

260
261
262
    def test_stable_diffusion_img2img_k_lms(self):
        pipe = StableDiffusionImg2ImgPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", safety_checker=None)
        pipe.scheduler = LMSDiscreteScheduler.from_config(pipe.scheduler.config)
263
264
265
266
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

267
268
269
        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()
270

271
        assert image.shape == (1, 512, 768, 3)
272
273
        expected_slice = np.array([0.0389, 0.0346, 0.0415, 0.0290, 0.0218, 0.0210, 0.0408, 0.0567, 0.0271])

274
        assert np.abs(expected_slice - image_slice).max() < 1e-3
275

276
277
278
    def test_stable_diffusion_img2img_ddim(self):
        pipe = StableDiffusionImg2ImgPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", safety_checker=None)
        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
279
280
281
282
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

283
284
285
        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()
286

287
        assert image.shape == (1, 512, 768, 3)
288
289
        expected_slice = np.array([0.0593, 0.0607, 0.0851, 0.0582, 0.0636, 0.0721, 0.0751, 0.0981, 0.0781])

290
        assert np.abs(expected_slice - image_slice).max() < 1e-3
291
292
293
294

    def test_stable_diffusion_img2img_intermediate_state(self):
        number_of_steps = 0

295
296
        def callback_fn(step: int, timestep: int, latents: torch.FloatTensor) -> None:
            callback_fn.has_been_called = True
297
298
            nonlocal number_of_steps
            number_of_steps += 1
299
            if step == 1:
300
301
302
                latents = latents.detach().cpu().numpy()
                assert latents.shape == (1, 4, 64, 96)
                latents_slice = latents[0, -3:, -3:, -1]
303
304
305
                expected_slice = np.array([-0.4958, 0.5107, 1.1045, 2.7539, 4.6680, 3.8320, 1.5049, 1.8633, 2.6523])

                assert np.abs(latents_slice.flatten() - expected_slice).max() < 5e-2
306
            elif step == 2:
307
308
309
                latents = latents.detach().cpu().numpy()
                assert latents.shape == (1, 4, 64, 96)
                latents_slice = latents[0, -3:, -3:, -1]
310
311
312
                expected_slice = np.array([-0.4956, 0.5078, 1.0918, 2.7520, 4.6484, 3.8125, 1.5146, 1.8633, 2.6367])

                assert np.abs(latents_slice.flatten() - expected_slice).max() < 5e-2
313

314
        callback_fn.has_been_called = False
315
316

        pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
317
            "CompVis/stable-diffusion-v1-4", safety_checker=None, torch_dtype=torch.float16
318
        )
319
        pipe = pipe.to(torch_device)
320
321
322
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

323
324
325
326
        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 == 2
327
328
329
330

    def test_stable_diffusion_pipeline_with_sequential_cpu_offloading(self):
        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()
Anton Lozhkov's avatar
Anton Lozhkov committed
331
        torch.cuda.reset_peak_memory_stats()
332
333

        pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
334
            "CompVis/stable-diffusion-v1-4", safety_checker=None, torch_dtype=torch.float16
335
        )
336
        pipe = pipe.to(torch_device)
337
338
339
340
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing(1)
        pipe.enable_sequential_cpu_offload()

341
342
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        _ = pipe(**inputs)
343
344

        mem_bytes = torch.cuda.max_memory_allocated()
Anton Lozhkov's avatar
Anton Lozhkov committed
345
346
        # make sure that less than 2.2 GB is allocated
        assert mem_bytes < 2.2 * 10**9
347

348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
    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 = StableDiffusionImg2ImgPipeline.from_pretrained(
            "CompVis/stable-diffusion-v1-4",
            safety_checker=None,
            torch_dtype=torch.float16,
        )
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe(**inputs)
        mem_bytes = torch.cuda.max_memory_allocated()

        # With model offloading

        # Reload but don't move to cuda
        pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
            "CompVis/stable-diffusion-v1-4",
            safety_checker=None,
            torch_dtype=torch.float16,
        )

        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)
        _ = pipe(**inputs)
        mem_bytes_offloaded = torch.cuda.max_memory_allocated()

        assert mem_bytes_offloaded < mem_bytes
        for module in pipe.text_encoder, pipe.unet, pipe.vae:
            assert module.device == torch.device("cpu")

389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
    def test_img2img_2nd_order(self):
        sd_pipe = StableDiffusionImg2ImgPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
        sd_pipe.scheduler = HeunDiscreteScheduler.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)
        inputs["num_inference_steps"] = 10
        inputs["strength"] = 0.75
        image = sd_pipe(**inputs).images[0]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/img2img/img2img_heun.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 5e-2

        inputs = self.get_inputs(torch_device)
        inputs["num_inference_steps"] = 11
        inputs["strength"] = 0.75
        image_other = sd_pipe(**inputs).images[0]

        mean_diff = np.abs(image - image_other).mean()

        # images should be very similar
        assert mean_diff < 5e-2

416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
    def test_stable_diffusion_img2img_pipeline_multiple_of_8(self):
        init_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
            "/img2img/sketch-mountains-input.jpg"
        )
        # resize to resolution that is divisible by 8 but not 16 or 32
        init_image = init_image.resize((760, 504))

        model_id = "CompVis/stable-diffusion-v1-4"
        pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
            model_id,
            safety_checker=None,
        )
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        prompt = "A fantasy landscape, trending on artstation"

435
        generator = torch.manual_seed(0)
436
437
438
439
440
441
442
443
444
445
446
447
448
        output = pipe(
            prompt=prompt,
            image=init_image,
            strength=0.75,
            guidance_scale=7.5,
            generator=generator,
            output_type="np",
        )
        image = output.images[0]

        image_slice = image[255:258, 383:386, -1]

        assert image.shape == (504, 760, 3)
449
450
451
        expected_slice = np.array([0.9393, 0.9500, 0.9399, 0.9438, 0.9458, 0.9400, 0.9455, 0.9414, 0.9423])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 5e-3
452

453
454
455
456
457
458
459
460
461
462
463
464
465
466
    def test_img2img_safety_checker_works(self):
        sd_pipe = StableDiffusionImg2ImgPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
        sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        inputs["num_inference_steps"] = 20
        # make sure the safety checker is activated
        inputs["prompt"] = "naked, sex, porn"
        out = sd_pipe(**inputs)

        assert out.nsfw_content_detected[0], f"Safety checker should work for prompt: {inputs['prompt']}"
        assert np.abs(out.images[0]).sum() < 1e-5  # should be all zeros

467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
    def test_img2img_compile(self):
        if version.parse(torch.__version__) < version.parse("2.0"):
            print(f"Test `test_stable_diffusion_ddim` is skipped because {torch.__version__} is < 2.0")
            return

        pipe = StableDiffusionImg2ImgPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", safety_checker=None)
        pipe.scheduler = DDIMScheduler.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)

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

        assert image.shape == (1, 512, 768, 3)
        expected_slice = np.array([0.0593, 0.0607, 0.0851, 0.0582, 0.0636, 0.0721, 0.0751, 0.0981, 0.0781])

        assert np.abs(expected_slice - image_slice).max() < 1e-3

489
490
491
492
493
494
495
496
497

@nightly
@require_torch_gpu
class StableDiffusionImg2ImgPipelineNightlyTests(unittest.TestCase):
    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

498
499
    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
        generator = torch.Generator(device=generator_device).manual_seed(seed)
500
501
502
503
504
505
506
507
508
509
510
        init_image = load_image(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_img2img/sketch-mountains-input.png"
        )
        inputs = {
            "prompt": "a fantasy landscape, concept art, high resolution",
            "image": init_image,
            "generator": generator,
            "num_inference_steps": 50,
            "strength": 0.75,
            "guidance_scale": 7.5,
YiYi Xu's avatar
YiYi Xu committed
511
            "output_type": "np",
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
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
        }
        return inputs

    def test_img2img_pndm(self):
        sd_pipe = StableDiffusionImg2ImgPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
        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]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_img2img/stable_diffusion_1_5_pndm.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

    def test_img2img_ddim(self):
        sd_pipe = StableDiffusionImg2ImgPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
        sd_pipe.scheduler = DDIMScheduler.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]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_img2img/stable_diffusion_1_5_ddim.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

    def test_img2img_lms(self):
        sd_pipe = StableDiffusionImg2ImgPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
        sd_pipe.scheduler = LMSDiscreteScheduler.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]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_img2img/stable_diffusion_1_5_lms.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

    def test_img2img_dpm(self):
        sd_pipe = StableDiffusionImg2ImgPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
        sd_pipe.scheduler = DPMSolverMultistepScheduler.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)
        inputs["num_inference_steps"] = 30
        image = sd_pipe(**inputs).images[0]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_img2img/stable_diffusion_1_5_dpm.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3