File size: 11,920 Bytes
87d40d2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
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
322
323
324
325
326
327
328
329
330
<!--Copyright 2024 The HuggingFace Team. All rights reserved.

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.
-->

# μŠ€μΌ€μ€„λŸ¬

diffusion νŒŒμ΄ν”„λΌμΈμ€ diffusion λͺ¨λΈ, μŠ€μΌ€μ€„λŸ¬ λ“±μ˜ μ»΄ν¬λ„ŒνŠΈλ“€λ‘œ κ΅¬μ„±λ©λ‹ˆλ‹€. 그리고 νŒŒμ΄ν”„λΌμΈ μ•ˆμ˜ 일뢀 μ»΄ν¬λ„ŒνŠΈλ₯Ό λ‹€λ₯Έ μ»΄ν¬λ„ŒνŠΈλ‘œ κ΅μ²΄ν•˜λŠ” μ‹μ˜ μ»€μŠ€ν„°λ§ˆμ΄μ§• μ—­μ‹œ κ°€λŠ₯ν•©λ‹ˆλ‹€.  이와 같은 μ»΄ν¬λ„ŒνŠΈ μ»€μŠ€ν„°λ§ˆμ΄μ§•μ˜ κ°€μž₯ λŒ€ν‘œμ μΈ μ˜ˆμ‹œκ°€ λ°”λ‘œ [μŠ€μΌ€μ€„λŸ¬](../api/schedulers/overview.md)λ₯Ό κ΅μ²΄ν•˜λŠ” κ²ƒμž…λ‹ˆλ‹€.



μŠ€μΌ€μ₯΄λŸ¬λŠ” λ‹€μŒκ³Ό 같이 diffusion μ‹œμŠ€ν…œμ˜ μ „λ°˜μ μΈ 디노이징 ν”„λ‘œμ„ΈμŠ€λ₯Ό μ •μ˜ν•©λ‹ˆλ‹€.

- 디노이징 μŠ€ν…μ„ μ–Όλ§ˆλ‚˜ κ°€μ Έκ°€μ•Ό ν• κΉŒ?
- ν™•λ₯ μ μœΌλ‘œ(stochastic) ν˜Ήμ€ ν™•μ •μ μœΌλ‘œ(deterministic)?
- 디노이징 된 μƒ˜ν”Œμ„ μ°Ύμ•„λ‚΄κΈ° μœ„ν•΄ μ–΄λ–€ μ•Œκ³ λ¦¬μ¦˜μ„ μ‚¬μš©ν•΄μ•Ό ν• κΉŒ?

μ΄λŸ¬ν•œ ν”„λ‘œμ„ΈμŠ€λŠ” λ‹€μ†Œ λ‚œν•΄ν•˜κ³ , 디노이징 속도와 디노이징 퀄리티 μ‚¬μ΄μ˜ νŠΈλ ˆμ΄λ“œ μ˜€ν”„λ₯Ό μ •μ˜ν•΄μ•Ό ν•˜λŠ” λ¬Έμ œκ°€ 될 수 μžˆμŠ΅λ‹ˆλ‹€. 주어진 νŒŒμ΄ν”„λΌμΈμ— μ–΄λ–€ μŠ€μΌ€μ€„λŸ¬κ°€ κ°€μž₯ μ ν•©ν•œμ§€λ₯Ό μ •λŸ‰μ μœΌλ‘œ νŒλ‹¨ν•˜λŠ” 것은 맀우 μ–΄λ €μš΄ μΌμž…λ‹ˆλ‹€. 이둜 인해 일단 ν•΄λ‹Ή μŠ€μΌ€μ€„λŸ¬λ₯Ό 직접 μ‚¬μš©ν•˜μ—¬, μƒμ„±λ˜λŠ” 이미지λ₯Ό 직접 눈으둜 보며, μ •μ„±μ μœΌλ‘œ μ„±λŠ₯을 νŒλ‹¨ν•΄λ³΄λŠ” 것이 μΆ”μ²œλ˜κ³€ ν•©λ‹ˆλ‹€.





## νŒŒμ΄ν”„λΌμΈ 뢈러였기

λ¨Όμ € μŠ€ν…Œμ΄λΈ” diffusion νŒŒμ΄ν”„λΌμΈμ„ λΆˆλŸ¬μ˜€λ„λ‘ ν•΄λ³΄κ² μŠ΅λ‹ˆλ‹€. λ¬Όλ‘  μŠ€ν…Œμ΄λΈ” diffusion을 μ‚¬μš©ν•˜κΈ° μœ„ν•΄μ„œλŠ”, ν—ˆκΉ…νŽ˜μ΄μŠ€ ν—ˆλΈŒμ— λ“±λ‘λœ μ‚¬μš©μžμ—¬μ•Ό ν•˜λ©°, κ΄€λ ¨ [λΌμ΄μ„ΌμŠ€](https://huggingface.co/runwayml/stable-diffusion-v1-5)에 λ™μ˜ν•΄μ•Ό ν•œλ‹€λŠ” 점을 μžŠμ§€ λ§μ•„μ£Όμ„Έμš”. 

*μ—­μž μ£Ό: λ‹€λ§Œ, ν˜„μž¬ μ‹ κ·œλ‘œ μƒμ„±ν•œ ν—ˆκΉ…νŽ˜μ΄μŠ€ 계정에 λŒ€ν•΄μ„œλŠ” λΌμ΄μ„ΌμŠ€ λ™μ˜λ₯Ό μš”κ΅¬ν•˜μ§€ μ•ŠλŠ” κ²ƒμœΌλ‘œ λ³΄μž…λ‹ˆλ‹€!*

```python
from huggingface_hub import login
from diffusers import DiffusionPipeline
import torch

# first we need to login with our access token
login()

# Now we can download the pipeline
pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)
```

λ‹€μŒμœΌλ‘œ, GPU둜 μ΄λ™ν•©λ‹ˆλ‹€.

```python
pipeline.to("cuda")
```





## μŠ€μΌ€μ€„λŸ¬ μ•‘μ„ΈμŠ€

μŠ€μΌ€μ€„λŸ¬λŠ” μ–Έμ œλ‚˜ νŒŒμ΄ν”„λΌμΈμ˜ μ»΄ν¬λ„ŒνŠΈλ‘œμ„œ μ‘΄μž¬ν•˜λ©°, 일반적으둜 νŒŒμ΄ν”„λΌμΈ μΈμŠ€ν„΄μŠ€ 내에 `scheduler`λΌλŠ” μ΄λ¦„μ˜ 속성(property)으둜 μ •μ˜λ˜μ–΄ μžˆμŠ΅λ‹ˆλ‹€. 

```python
pipeline.scheduler
```

**Output**:

```
PNDMScheduler {
  "_class_name": "PNDMScheduler",
  "_diffusers_version": "0.8.0.dev0",
  "beta_end": 0.012,
  "beta_schedule": "scaled_linear",
  "beta_start": 0.00085,
  "clip_sample": false,
  "num_train_timesteps": 1000,
  "set_alpha_to_one": false,
  "skip_prk_steps": true,
  "steps_offset": 1,
  "trained_betas": null
}
```

좜λ ₯ κ²°κ³Όλ₯Ό 톡해, μš°λ¦¬λŠ” ν•΄λ‹Ή μŠ€μΌ€μ€„λŸ¬κ°€ [`PNDMScheduler`]의 μΈμŠ€ν„΄μŠ€λΌλŠ” 것을 μ•Œ 수 μžˆμŠ΅λ‹ˆλ‹€. 이제 [`PNDMScheduler`]와 λ‹€λ₯Έ μŠ€μΌ€μ€„λŸ¬λ“€μ˜ μ„±λŠ₯을 비ꡐ해보도둝 ν•˜κ² μŠ΅λ‹ˆλ‹€. λ¨Όμ € ν…ŒμŠ€νŠΈμ— μ‚¬μš©ν•  ν”„λ‘¬ν”„νŠΈλ₯Ό λ‹€μŒκ³Ό 같이 μ •μ˜ν•΄λ³΄λ„λ‘ ν•˜κ² μŠ΅λ‹ˆλ‹€. 

```python
prompt = "A photograph of an astronaut riding a horse on Mars, high resolution, high definition."
```

λ‹€μŒμœΌλ‘œ μœ μ‚¬ν•œ 이미지 생성을 보μž₯ν•˜κΈ° μœ„ν•΄μ„œ, λ‹€μŒκ³Ό 같이 λžœλ€μ‹œλ“œλ₯Ό 고정해주도둝 ν•˜κ² μŠ΅λ‹ˆλ‹€. 

```python
generator = torch.Generator(device="cuda").manual_seed(8)
image = pipeline(prompt, generator=generator).images[0]
image
```

<p align="center">
    <br>
    <img src="https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/diffusers_docs/astronaut_pndm.png" width="400"/>
    <br>
</p>




## μŠ€μΌ€μ€„λŸ¬ κ΅μ²΄ν•˜κΈ°

λ‹€μŒμœΌλ‘œ νŒŒμ΄ν”„λΌμΈμ˜ μŠ€μΌ€μ€„λŸ¬λ₯Ό λ‹€λ₯Έ μŠ€μΌ€μ€„λŸ¬λ‘œ κ΅μ²΄ν•˜λŠ” 방법에 λŒ€ν•΄ μ•Œμ•„λ³΄κ² μŠ΅λ‹ˆλ‹€. λͺ¨λ“  μŠ€μΌ€μ€„λŸ¬λŠ” [`SchedulerMixin.compatibles`]λΌλŠ” 속성(property)을 κ°–κ³  μžˆμŠ΅λ‹ˆλ‹€. ν•΄λ‹Ή 속성은 **ν˜Έν™˜ κ°€λŠ₯ν•œ** μŠ€μΌ€μ€„λŸ¬λ“€μ— λŒ€ν•œ 정보λ₯Ό λ‹΄κ³  μžˆμŠ΅λ‹ˆλ‹€. 

```python
pipeline.scheduler.compatibles
```

**Output**:

```
[diffusers.schedulers.scheduling_lms_discrete.LMSDiscreteScheduler,
 diffusers.schedulers.scheduling_ddim.DDIMScheduler,
 diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler,
 diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler,
 diffusers.schedulers.scheduling_pndm.PNDMScheduler,
 diffusers.schedulers.scheduling_ddpm.DDPMScheduler,
 diffusers.schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteScheduler]
```

ν˜Έν™˜λ˜λŠ” μŠ€μΌ€μ€„λŸ¬λ“€μ„ μ‚΄νŽ΄λ³΄λ©΄ μ•„λž˜μ™€ κ°™μŠ΅λ‹ˆλ‹€.

- [`LMSDiscreteScheduler`], 
- [`DDIMScheduler`], 
- [`DPMSolverMultistepScheduler`], 
- [`EulerDiscreteScheduler`], 
- [`PNDMScheduler`], 
- [`DDPMScheduler`], 
- [`EulerAncestralDiscreteScheduler`].

μ•žμ„œ μ •μ˜ν–ˆλ˜ ν”„λ‘¬ν”„νŠΈλ₯Ό μ‚¬μš©ν•΄μ„œ 각각의 μŠ€μΌ€μ€„λŸ¬λ“€μ„ 비ꡐ해보도둝 ν•˜κ² μŠ΅λ‹ˆλ‹€.

λ¨Όμ € νŒŒμ΄ν”„λΌμΈ μ•ˆμ˜ μŠ€μΌ€μ€„λŸ¬λ₯Ό λ°”κΎΈκΈ° μœ„ν•΄ [`ConfigMixin.config`] 속성과 [`ConfigMixin.from_config`] λ©”μ„œλ“œλ₯Ό ν™œμš©ν•΄λ³΄λ €κ³  ν•©λ‹ˆλ‹€.



```python
pipeline.scheduler.config
```

**Output**:

```
FrozenDict([('num_train_timesteps', 1000),
            ('beta_start', 0.00085),
            ('beta_end', 0.012),
            ('beta_schedule', 'scaled_linear'),
            ('trained_betas', None),
            ('skip_prk_steps', True),
            ('set_alpha_to_one', False),
            ('steps_offset', 1),
            ('_class_name', 'PNDMScheduler'),
            ('_diffusers_version', '0.8.0.dev0'),
            ('clip_sample', False)])
```

κΈ°μ‘΄ μŠ€μΌ€μ€„λŸ¬μ˜ configλ₯Ό ν˜Έν™˜ κ°€λŠ₯ν•œ λ‹€λ₯Έ μŠ€μΌ€μ€„λŸ¬μ— μ΄μ‹ν•˜λŠ” 것 μ—­μ‹œ κ°€λŠ₯ν•©λ‹ˆλ‹€. 

λ‹€μŒ μ˜ˆμ‹œλŠ” κΈ°μ‘΄ μŠ€μΌ€μ€„λŸ¬(`pipeline.scheduler`)λ₯Ό λ‹€λ₯Έ μ’…λ₯˜μ˜ μŠ€μΌ€μ€„λŸ¬(`DDIMScheduler`)둜 λ°”κΎΈλŠ” μ½”λ“œμž…λ‹ˆλ‹€. κΈ°μ‘΄ μŠ€μΌ€μ€„λŸ¬κ°€ κ°–κ³  있던 configλ₯Ό `.from_config` λ©”μ„œλ“œμ˜ 인자둜 μ „λ‹¬ν•˜λŠ” 것을 확인할 수 μžˆμŠ΅λ‹ˆλ‹€.

```python
from diffusers import DDIMScheduler

pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
```



이제 νŒŒμ΄ν”„λΌμΈμ„ μ‹€ν–‰ν•΄μ„œ 두 μŠ€μΌ€μ€„λŸ¬ μ‚¬μ΄μ˜ μƒμ„±λœ μ΄λ―Έμ§€μ˜ 퀄리티λ₯Ό λΉ„κ΅ν•΄λ΄…μ‹œλ‹€.

```python
generator = torch.Generator(device="cuda").manual_seed(8)
image = pipeline(prompt, generator=generator).images[0]
image
```

<p align="center">
    <br>
    <img src="https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/diffusers_docs/astronaut_ddim.png" width="400"/>
    <br>
</p>




## μŠ€μΌ€μ€„λŸ¬λ“€ 비ꡐ해보기

μ§€κΈˆκΉŒμ§€λŠ” [`PNDMScheduler`]와 [`DDIMScheduler`] μŠ€μΌ€μ€„λŸ¬λ₯Ό μ‹€ν–‰ν•΄λ³΄μ•˜μŠ΅λ‹ˆλ‹€. 아직 비ꡐ해볼 μŠ€μΌ€μ€„λŸ¬λ“€μ΄ 더 많이 λ‚¨μ•„μžˆμœΌλ‹ˆ 계속 비ꡐ해보도둝 ν•˜κ² μŠ΅λ‹ˆλ‹€.



[`LMSDiscreteScheduler`]을 일반적으둜 더 쒋은 κ²°κ³Όλ₯Ό λ³΄μ—¬μ€λ‹ˆλ‹€.

```python
from diffusers import LMSDiscreteScheduler

pipeline.scheduler = LMSDiscreteScheduler.from_config(pipeline.scheduler.config)

generator = torch.Generator(device="cuda").manual_seed(8)
image = pipeline(prompt, generator=generator).images[0]
image
```

<p align="center">
    <br>
    <img src="https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/diffusers_docs/astronaut_lms.png" width="400"/>
    <br>
</p>


[`EulerDiscreteScheduler`]와 [`EulerAncestralDiscreteScheduler`] κ³ μž‘ 30번의 inference stepλ§ŒμœΌλ‘œλ„ 높은 ν€„λ¦¬ν‹°μ˜ 이미지λ₯Ό μƒμ„±ν•˜λŠ” 것을 μ•Œ 수 μžˆμŠ΅λ‹ˆλ‹€.

```python
from diffusers import EulerDiscreteScheduler

pipeline.scheduler = EulerDiscreteScheduler.from_config(pipeline.scheduler.config)

generator = torch.Generator(device="cuda").manual_seed(8)
image = pipeline(prompt, generator=generator, num_inference_steps=30).images[0]
image
```

<p align="center">
    <br>
    <img src="https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/diffusers_docs/astronaut_euler_discrete.png" width="400"/>
    <br>
</p>


```python
from diffusers import EulerAncestralDiscreteScheduler

pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config(pipeline.scheduler.config)

generator = torch.Generator(device="cuda").manual_seed(8)
image = pipeline(prompt, generator=generator, num_inference_steps=30).images[0]
image
```

<p align="center">
    <br>
    <img src="https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/diffusers_docs/astronaut_euler_ancestral.png" width="400"/>
    <br>
</p>


μ§€κΈˆ 이 λ¬Έμ„œλ₯Ό μž‘μ„±ν•˜λŠ” ν˜„μ‹œμ  기쀀에선, [`DPMSolverMultistepScheduler`]κ°€ μ‹œκ°„ λŒ€λΉ„ κ°€μž₯ 쒋은 ν’ˆμ§ˆμ˜ 이미지λ₯Ό μƒμ„±ν•˜λŠ” 것 κ°™μŠ΅λ‹ˆλ‹€. 20번 μ •λ„μ˜ μŠ€ν…λ§ŒμœΌλ‘œλ„ 싀행될 수 μžˆμŠ΅λ‹ˆλ‹€.



```python
from diffusers import DPMSolverMultistepScheduler

pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config)

generator = torch.Generator(device="cuda").manual_seed(8)
image = pipeline(prompt, generator=generator, num_inference_steps=20).images[0]
image
```

<p align="center">
    <br>
    <img src="https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/diffusers_docs/astronaut_dpm.png" width="400"/>
    <br>
</p>


λ³΄μ‹œλ‹€μ‹œν”Ό μƒμ„±λœ 이미지듀은 맀우 λΉ„μŠ·ν•˜κ³ , λΉ„μŠ·ν•œ 퀄리티λ₯Ό λ³΄μ΄λŠ” 것 κ°™μŠ΅λ‹ˆλ‹€. μ‹€μ œλ‘œ μ–΄λ–€ μŠ€μΌ€μ€„λŸ¬λ₯Ό 선택할 κ²ƒμΈκ°€λŠ” μ’…μ’… νŠΉμ • 이용 사둀에 κΈ°λ°˜ν•΄μ„œ κ²°μ •λ˜κ³€ ν•©λ‹ˆλ‹€. κ²°κ΅­ μ—¬λŸ¬ μ’…λ₯˜μ˜ μŠ€μΌ€μ€„λŸ¬λ₯Ό 직접 μ‹€ν–‰μ‹œμΌœλ³΄κ³  눈으둜 직접 λΉ„κ΅ν•΄μ„œ νŒλ‹¨ν•˜λŠ” 게 쒋은 선택일 것 κ°™μŠ΅λ‹ˆλ‹€.



## Flaxμ—μ„œ μŠ€μΌ€μ€„λŸ¬ κ΅μ²΄ν•˜κΈ°

JAX/Flax μ‚¬μš©μžμΈ 경우 κΈ°λ³Έ νŒŒμ΄ν”„λΌμΈ μŠ€μΌ€μ€„λŸ¬λ₯Ό λ³€κ²½ν•  μˆ˜λ„ μžˆμŠ΅λ‹ˆλ‹€. λ‹€μŒμ€ Flax Stable Diffusion νŒŒμ΄ν”„λΌμΈκ³Ό μ΄ˆκ³ μ† [DDPM-Solver++ μŠ€μΌ€μ€„λŸ¬λ₯Ό](../api/schedulers/multistep_dpm_solver) μ‚¬μš©ν•˜μ—¬ 좔둠을 μ‹€ν–‰ν•˜λŠ” 방법에 λŒ€ν•œ μ˜ˆμ‹œμž…λ‹ˆλ‹€ .

```Python
import jax
import numpy as np
from flax.jax_utils import replicate
from flax.training.common_utils import shard

from diffusers import FlaxStableDiffusionPipeline, FlaxDPMSolverMultistepScheduler

model_id = "runwayml/stable-diffusion-v1-5"
scheduler, scheduler_state = FlaxDPMSolverMultistepScheduler.from_pretrained(
    model_id,
    subfolder="scheduler"
)
pipeline, params = FlaxStableDiffusionPipeline.from_pretrained(
    model_id,
    scheduler=scheduler,
    revision="bf16",
    dtype=jax.numpy.bfloat16,
)
params["scheduler"] = scheduler_state

# Generate 1 image per parallel device (8 on TPUv2-8 or TPUv3-8)
prompt = "a photo of an astronaut riding a horse on mars"
num_samples = jax.device_count()
prompt_ids = pipeline.prepare_inputs([prompt] * num_samples)

prng_seed = jax.random.PRNGKey(0)
num_inference_steps = 25

# shard inputs and rng
params = replicate(params)
prng_seed = jax.random.split(prng_seed, jax.device_count())
prompt_ids = shard(prompt_ids)

images = pipeline(prompt_ids, params, prng_seed, num_inference_steps, jit=True).images
images = pipeline.numpy_to_pil(np.asarray(images.reshape((num_samples,) + images.shape[-3:])))
```

<Tip warning={true}>

λ‹€μŒ Flax μŠ€μΌ€μ€„λŸ¬λŠ” *아직* Flax Stable Diffusion νŒŒμ΄ν”„λΌμΈκ³Ό ν˜Έν™˜λ˜μ§€ μ•ŠμŠ΅λ‹ˆλ‹€.

- `FlaxLMSDiscreteScheduler`
- `FlaxDDPMScheduler`

</Tip>