test_stable_diffusion_img2img.py 27.9 KB
Newer Older
1
# coding=utf-8
2
# Copyright 2024 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 transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
24
25
26

from diffusers import (
    AutoencoderKL,
27
    AutoencoderTiny,
28
    DDIMScheduler,
29
    DPMSolverMultistepScheduler,
30
    HeunDiscreteScheduler,
Patrick von Platen's avatar
Patrick von Platen committed
31
    LCMScheduler,
32
33
34
35
36
    LMSDiscreteScheduler,
    PNDMScheduler,
    StableDiffusionImg2ImgPipeline,
    UNet2DConditionModel,
)
37
38
from diffusers.utils.testing_utils import (
    enable_full_determinism,
Dhruv Nair's avatar
Dhruv Nair committed
39
40
41
42
    floats_tensor,
    load_image,
    load_numpy,
    nightly,
Dhruv Nair's avatar
Dhruv Nair committed
43
    require_python39_or_higher,
44
45
46
47
    require_torch_2,
    require_torch_gpu,
    run_test_in_subprocess,
    skip_mps,
Dhruv Nair's avatar
Dhruv Nair committed
48
49
    slow,
    torch_device,
50
)
51

52
53
54
55
from ..pipeline_params import (
    IMAGE_TO_IMAGE_IMAGE_PARAMS,
    TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS,
    TEXT_GUIDED_IMAGE_VARIATION_PARAMS,
56
    TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS,
57
)
Aryan's avatar
Aryan committed
58
59
60
61
62
63
from ..test_pipelines_common import (
    IPAdapterTesterMixin,
    PipelineKarrasSchedulerTesterMixin,
    PipelineLatentTesterMixin,
    PipelineTesterMixin,
)
64

65

66
enable_full_determinism()
67
68


69
70
71
72
73
74
75
76
77
78
79
# Will be run via run_test_in_subprocess
def _test_img2img_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 = StableDiffusionImg2ImgPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", safety_checker=None)
        pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
80
        pipe.unet.set_default_attn_processor()
81
82
83
84
85
86
87
88
89
        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, -3:, -3:, -1].flatten()

        assert image.shape == (1, 512, 768, 3)
90
        expected_slice = np.array([0.0606, 0.0570, 0.0805, 0.0579, 0.0628, 0.0623, 0.0843, 0.1115, 0.0806])
91
92
93
94
95
96
97
98
99
100

        assert np.abs(expected_slice - image_slice).max() < 1e-3
    except Exception:
        error = f"{traceback.format_exc()}"

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


101
class StableDiffusionImg2ImgPipelineFastTests(
Aryan's avatar
Aryan committed
102
103
104
105
106
    IPAdapterTesterMixin,
    PipelineLatentTesterMixin,
    PipelineKarrasSchedulerTesterMixin,
    PipelineTesterMixin,
    unittest.TestCase,
107
):
108
    pipeline_class = StableDiffusionImg2ImgPipeline
109
110
111
    params = TEXT_GUIDED_IMAGE_VARIATION_PARAMS - {"height", "width"}
    required_optional_params = PipelineTesterMixin.required_optional_params - {"latents"}
    batch_params = TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS
112
    image_params = IMAGE_TO_IMAGE_IMAGE_PARAMS
113
    image_latents_params = IMAGE_TO_IMAGE_IMAGE_PARAMS
114
    callback_cfg_params = TEXT_TO_IMAGE_CALLBACK_CFG_PARAMS
115

Patrick von Platen's avatar
Patrick von Platen committed
116
    def get_dummy_components(self, time_cond_proj_dim=None):
117
        torch.manual_seed(0)
118
        unet = UNet2DConditionModel(
119
120
            block_out_channels=(32, 64),
            layers_per_block=2,
Patrick von Platen's avatar
Patrick von Platen committed
121
            time_cond_proj_dim=time_cond_proj_dim,
122
123
124
125
126
127
128
            sample_size=32,
            in_channels=4,
            out_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            cross_attention_dim=32,
        )
129
        scheduler = PNDMScheduler(skip_prk_steps=True)
130
        torch.manual_seed(0)
131
        vae = AutoencoderKL(
132
133
134
135
136
137
138
139
            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)
140
        text_encoder_config = CLIPTextConfig(
141
142
143
144
145
146
147
148
149
150
            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,
        )
151
152
        text_encoder = CLIPTextModel(text_encoder_config)
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
153

154
155
156
157
158
159
160
        components = {
            "unet": unet,
            "scheduler": scheduler,
            "vae": vae,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "safety_checker": None,
161
            "feature_extractor": None,
162
            "image_encoder": None,
163
164
165
        }
        return components

166
167
168
    def get_dummy_tiny_autoencoder(self):
        return AutoencoderTiny(in_channels=3, out_channels=3, latent_channels=4)

169
    def get_dummy_inputs(self, device, seed=0):
170
        image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
171
        image = image / 2 + 0.5
172
173
174
175
176
177
        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",
178
            "image": image,
179
180
181
            "generator": generator,
            "num_inference_steps": 2,
            "guidance_scale": 6.0,
182
            "output_type": "np",
183
184
        }
        return inputs
185

186
    def test_stable_diffusion_img2img_default_case(self):
187
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
188
189
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionImg2ImgPipeline(**components)
190
191
192
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

193
194
        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
195
196
197
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 32, 32, 3)
198
        expected_slice = np.array([0.4555, 0.3216, 0.4049, 0.4620, 0.4618, 0.4126, 0.4122, 0.4629, 0.4579])
199

200
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
201

Patrick von Platen's avatar
Patrick von Platen committed
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
    def test_stable_diffusion_img2img_default_case_lcm(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components(time_cond_proj_dim=256)
        sd_pipe = StableDiffusionImg2ImgPipeline(**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, 32, 32, 3)
        expected_slice = np.array([0.5709, 0.4614, 0.4587, 0.5978, 0.5298, 0.6910, 0.6240, 0.5212, 0.5454])

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

219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
    def test_stable_diffusion_img2img_default_case_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 = StableDiffusionImg2ImgPipeline(**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, 32, 32, 3)
        expected_slice = np.array([0.5709, 0.4614, 0.4587, 0.5978, 0.5298, 0.6910, 0.6240, 0.5212, 0.5454])

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

238
239
    def test_stable_diffusion_img2img_negative_prompt(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
240
241
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionImg2ImgPipeline(**components)
242
243
244
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

245
        inputs = self.get_dummy_inputs(device)
246
        negative_prompt = "french fries"
247
        output = sd_pipe(**inputs, negative_prompt=negative_prompt)
248
249
250
251
        image = output.images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 32, 32, 3)
252
        expected_slice = np.array([0.4593, 0.3408, 0.4232, 0.4749, 0.4476, 0.4115, 0.4357, 0.4733, 0.4663])
253

254
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
255

256
257
258
259
260
261
    def test_ip_adapter_single(self):
        expected_pipe_slice = None
        if torch_device == "cpu":
            expected_pipe_slice = np.array([0.4932, 0.5092, 0.5135, 0.5517, 0.5626, 0.6621, 0.6490, 0.5021, 0.5441])
        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)

262
263
    def test_stable_diffusion_img2img_multiple_init_images(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
264
265
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionImg2ImgPipeline(**components)
266
267
268
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

269
270
271
272
        inputs = self.get_dummy_inputs(device)
        inputs["prompt"] = [inputs["prompt"]] * 2
        inputs["image"] = inputs["image"].repeat(2, 1, 1, 1)
        image = sd_pipe(**inputs).images
273
274
275
        image_slice = image[-1, -3:, -3:, -1]

        assert image.shape == (2, 32, 32, 3)
276
        expected_slice = np.array([0.4241, 0.5576, 0.5711, 0.4792, 0.4311, 0.5952, 0.5827, 0.5138, 0.5109])
277

278
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
279
280
281

    def test_stable_diffusion_img2img_k_lms(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
282
283
284
        components = self.get_dummy_components()
        components["scheduler"] = LMSDiscreteScheduler(
            beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear"
285
        )
286
        sd_pipe = StableDiffusionImg2ImgPipeline(**components)
287
288
289
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

290
291
        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
292
293
294
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 32, 32, 3)
295
        expected_slice = np.array([0.4398, 0.4949, 0.4337, 0.6580, 0.5555, 0.4338, 0.5769, 0.5955, 0.5175])
296

297
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
298

299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
    def test_stable_diffusion_img2img_tiny_autoencoder(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionImg2ImgPipeline(**components)
        sd_pipe.vae = self.get_dummy_tiny_autoencoder()
        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, 32, 32, 3)
        expected_slice = np.array([0.00669, 0.00669, 0.0, 0.00693, 0.00858, 0.0, 0.00567, 0.00515, 0.00125])

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

316
317
318
319
320
321
322
323
324
325
326
327
328
329
    @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):
330
331
332
333
        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)
334

335
336
337
    def test_float16_inference(self):
        super().test_float16_inference(expected_max_diff=5e-1)

Dhruv Nair's avatar
Dhruv Nair committed
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
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
    def test_pipeline_interrupt(self):
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionImg2ImgPipeline(**components)
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(torch_device)

        prompt = "hey"
        num_inference_steps = 3

        # store intermediate latents from the generation process
        class PipelineState:
            def __init__(self):
                self.state = []

            def apply(self, pipe, i, t, callback_kwargs):
                self.state.append(callback_kwargs["latents"])
                return callback_kwargs

        pipe_state = PipelineState()
        sd_pipe(
            prompt,
            image=inputs["image"],
            num_inference_steps=num_inference_steps,
            output_type="np",
            generator=torch.Generator("cpu").manual_seed(0),
            callback_on_step_end=pipe_state.apply,
        ).images

        # interrupt generation at step index
        interrupt_step_idx = 1

        def callback_on_step_end(pipe, i, t, callback_kwargs):
            if i == interrupt_step_idx:
                pipe._interrupt = True

            return callback_kwargs

        output_interrupted = sd_pipe(
            prompt,
            image=inputs["image"],
            num_inference_steps=num_inference_steps,
            output_type="latent",
            generator=torch.Generator("cpu").manual_seed(0),
            callback_on_step_end=callback_on_step_end,
        ).images

        # fetch intermediate latents at the interrupted step
        # from the completed generation process
        intermediate_latent = pipe_state.state[interrupt_step_idx]

        # compare the intermediate latent to the output of the interrupted process
        # they should be the same
        assert torch.allclose(intermediate_latent, output_interrupted, atol=1e-4)

394
395

@slow
396
@require_torch_gpu
397
class StableDiffusionImg2ImgPipelineSlowTests(unittest.TestCase):
398
399
400
401
402
    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

403
404
    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
        generator = torch.Generator(device=generator_device).manual_seed(seed)
405
        init_image = load_image(
406
407
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_img2img/sketch-mountains-input.png"
408
        )
409
410
411
412
413
414
415
        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
416
            "output_type": "np",
417
418
        }
        return inputs
419

420
421
    def test_stable_diffusion_img2img_default(self):
        pipe = StableDiffusionImg2ImgPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", safety_checker=None)
422
423
424
425
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

426
427
428
        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()
429

430
        assert image.shape == (1, 512, 768, 3)
431
432
        expected_slice = np.array([0.4300, 0.4662, 0.4930, 0.3990, 0.4307, 0.4525, 0.3719, 0.4064, 0.3923])

433
        assert np.abs(expected_slice - image_slice).max() < 1e-3
434

435
436
437
    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)
438
439
440
441
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

442
443
444
        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()
445

446
        assert image.shape == (1, 512, 768, 3)
447
448
        expected_slice = np.array([0.0389, 0.0346, 0.0415, 0.0290, 0.0218, 0.0210, 0.0408, 0.0567, 0.0271])

449
        assert np.abs(expected_slice - image_slice).max() < 1e-3
450

451
452
453
    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)
454
455
456
457
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

458
459
460
        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()
461

462
        assert image.shape == (1, 512, 768, 3)
463
464
        expected_slice = np.array([0.0593, 0.0607, 0.0851, 0.0582, 0.0636, 0.0721, 0.0751, 0.0981, 0.0781])

465
        assert np.abs(expected_slice - image_slice).max() < 1e-3
466
467
468
469

    def test_stable_diffusion_img2img_intermediate_state(self):
        number_of_steps = 0

470
471
        def callback_fn(step: int, timestep: int, latents: torch.FloatTensor) -> None:
            callback_fn.has_been_called = True
472
473
            nonlocal number_of_steps
            number_of_steps += 1
474
            if step == 1:
475
476
477
                latents = latents.detach().cpu().numpy()
                assert latents.shape == (1, 4, 64, 96)
                latents_slice = latents[0, -3:, -3:, -1]
478
479
480
                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
481
            elif step == 2:
482
483
484
                latents = latents.detach().cpu().numpy()
                assert latents.shape == (1, 4, 64, 96)
                latents_slice = latents[0, -3:, -3:, -1]
485
486
487
                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
488

489
        callback_fn.has_been_called = False
490
491

        pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
492
            "CompVis/stable-diffusion-v1-4", safety_checker=None, torch_dtype=torch.float16
493
        )
494
        pipe = pipe.to(torch_device)
495
496
497
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

498
499
500
501
        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
502
503
504
505

    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
506
        torch.cuda.reset_peak_memory_stats()
507
508

        pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
509
            "CompVis/stable-diffusion-v1-4", safety_checker=None, torch_dtype=torch.float16
510
        )
511
        pipe = pipe.to(torch_device)
512
513
514
515
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing(1)
        pipe.enable_sequential_cpu_offload()

516
517
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        _ = pipe(**inputs)
518
519

        mem_bytes = torch.cuda.max_memory_allocated()
Anton Lozhkov's avatar
Anton Lozhkov committed
520
521
        # make sure that less than 2.2 GB is allocated
        assert mem_bytes < 2.2 * 10**9
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
    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")

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
    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

591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
    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"

610
        generator = torch.manual_seed(0)
611
612
613
614
615
616
617
618
619
620
621
622
623
        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)
624
625
626
        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
627

628
629
630
631
632
633
634
635
636
637
638
639
640
641
    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

Dhruv Nair's avatar
Dhruv Nair committed
642
    @require_python39_or_higher
643
    @require_torch_2
644
    def test_img2img_compile(self):
645
646
647
648
649
650
651
        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_img2img_compile, inputs=inputs)
652

653
654
655
656
657
658
659
660
661

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

662
663
    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
        generator = torch.Generator(device=generator_device).manual_seed(seed)
664
665
666
667
668
669
670
671
672
673
674
        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
675
            "output_type": "np",
676
677
678
679
680
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
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
        }
        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