patrickvonplaten commited on
Commit
9c24957
1 Parent(s): 4d5006d

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +433 -0
README.md ADDED
@@ -0,0 +1,433 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ tags:
4
+ - stable-diffusion
5
+ - text-to-image
6
+ inference: false
7
+ ---
8
+
9
+ # Stable Diffusion v1 Model Card
10
+ This model card focuses on the model associated with the Stable Diffusion model, available [here](https://github.com/CompVis/stable-diffusion).
11
+
12
+ ## Model Details
13
+ - **Developed by:** Robin Rombach, Patrick Esser
14
+ - **Model type:** Diffusion-based text-to-image generation model
15
+ - **Language(s):** English
16
+ - **License:** [Proprietary](LICENSE)
17
+ - **Model Description:** This is a model that can be used to generate and modify images based on text prompts. It is a [Latent Diffusion Model](https://arxiv.org/abs/2112.10752) that uses a fixed, pretrained text encoder ([CLIP ViT-L/14](https://arxiv.org/abs/2103.00020)) as suggested in the [Imagen paper](https://arxiv.org/abs/2205.11487).
18
+ - **Resources for more information:** [GitHub Repository](https://github.com/CompVis/stable-diffusion), [Paper](https://arxiv.org/abs/2112.10752).
19
+ - **Cite as:**
20
+
21
+ @InProceedings{Rombach_2022_CVPR,
22
+ author = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn},
23
+ title = {High-Resolution Image Synthesis With Latent Diffusion Models},
24
+ booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
25
+ month = {June},
26
+ year = {2022},
27
+ pages = {10684-10695}
28
+ }
29
+
30
+ ## Usage examples
31
+
32
+ ```bash
33
+ pip install --upgrade diffusers transformers scipy
34
+ ```
35
+
36
+ Run this command to log in with your HF Hub token if you haven't before:
37
+ ```bash
38
+ huggingface-cli login
39
+ ```
40
+
41
+ Running the pipeline with the default PLMS scheduler:
42
+ ```python
43
+ from torch import autocast
44
+ from diffusers import StableDiffusionPipeline
45
+
46
+ model_id = "CompVis/stable-diffusion-v1-3-diffusers"
47
+ pipe = StableDiffusionPipeline.from_pretrained(model_id, use_auth_token=True).to("cuda")
48
+
49
+ prompt = "a photograph of an astronaut riding a horse"
50
+ with autocast("cuda"):
51
+ image = pipe(prompt, guidance_scale=7)["sample"][0] # image here is in PIL format
52
+
53
+ image.save(f"astronaut_rides_horse.png")
54
+ ```
55
+
56
+ To swap out the noise scheduler, pass it to `from_pretrained`:
57
+
58
+ ```python
59
+ from diffusers import StableDiffusionPipeline, LMSDiscreteScheduler
60
+
61
+ model_id = "CompVis/stable-diffusion-v1-3-diffusers"
62
+ # Use the K-LMS scheduler here instead
63
+ scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000)
64
+ pipe = StableDiffusionPipeline.from_pretrained(model_id, scheduler=scheduler, use_auth_token=True).to("cuda")
65
+ ```
66
+
67
+ # Uses
68
+
69
+ ## Direct Use
70
+ The model is intended for research purposes only. Possible research areas and
71
+ tasks include
72
+
73
+ - Safe deployment of models which have the potential to generate harmful content.
74
+ - Probing and understanding the limitations and biases of generative models.
75
+ - Generation of artworks and use in design and other artistic processes.
76
+ - Applications in educational or creative tools.
77
+ - Research on generative models.
78
+
79
+ Excluded uses are described below.
80
+
81
+ ### Misuse, Malicious Use, and Out-of-Scope Use
82
+ _Note: This section is taken from the [DALLE-MINI model card](https://huggingface.co/dalle-mini/dalle-mini), but applies in the same way to Stable Diffusion v1_.
83
+
84
+
85
+ The model should not be used to intentionally create or disseminate images that create hostile or alienating environments for people. This includes generating images that people would foreseeably find disturbing, distressing, or offensive; or content that propagates historical or current stereotypes.
86
+ #### Out-of-Scope Use
87
+ The model was not trained to be factual or true representations of people or events, and therefore using the model to generate such content is out-of-scope for the abilities of this model.
88
+ #### Misuse and Malicious Use
89
+ Using the model to generate content that is cruel to individuals is a misuse of this model. This includes, but is not limited to:
90
+
91
+ - Generating demeaning, dehumanizing, or otherwise harmful representations of people or their environments, cultures, religions, etc.
92
+ - Intentionally promoting or propagating discriminatory content or harmful stereotypes.
93
+ - Impersonating individuals without their consent.
94
+ - Sexual content without consent of the people who might see it.
95
+ - Mis- and disinformation
96
+ - Representations of egregious violence and gore
97
+ - Sharing of copyrighted or licensed material in violation of its terms of use.
98
+ - Sharing content that is an alteration of copyrighted or licensed material in violation of its terms of use.
99
+
100
+ ## Limitations and Bias
101
+
102
+ ### Limitations
103
+
104
+ - The model does not achieve perfect photorealism
105
+ - The model cannot render legible text
106
+ - The model does not perform well on more difficult tasks which involve compositionality, such as rendering an image corresponding to “A red cube on top of a blue sphere”
107
+ - Faces and people in general may not be generated properly.
108
+ - The model was trained mainly with English captions and will not work as well in other languages.
109
+ - The autoencoding part of the model is lossy
110
+ - The model was trained on a large-scale dataset
111
+ [LAION-5B](https://laion.ai/blog/laion-5b/) which contains adult material
112
+ and is not fit for product use without additional safety mechanisms and
113
+ considerations.
114
+
115
+ ### Bias
116
+ While the capabilities of image generation models are impressive, they can also reinforce or exacerbate social biases.
117
+ Stable Diffusion v1 was trained on subsets of [LAION-2B(en)](https://laion.ai/blog/laion-5b/),
118
+ which consists of images that are primarily limited to English descriptions.
119
+ Texts and images from communities and cultures that use other languages are likely to be insufficiently accounted for.
120
+ This affects the overall output of the model, as white and western cultures are often set as the default. Further, the
121
+ ability of the model to generate content with non-English prompts is significantly worse than with English-language prompts.
122
+
123
+
124
+ ## Training
125
+
126
+ **Training Data**
127
+ The model developers used the following dataset for training the model:
128
+
129
+ - LAION-2B (en) and subsets thereof (see next section)
130
+
131
+ **Training Procedure**
132
+ Stable Diffusion v1 is a latent diffusion model which combines an autoencoder with a diffusion model that is trained in the latent space of the autoencoder. During training,
133
+
134
+ - Images are encoded through an encoder, which turns images into latent representations. The autoencoder uses a relative downsampling factor of 8 and maps images of shape H x W x 3 to latents of shape H/f x W/f x 4
135
+ - Text prompts are encoded through a ViT-L/14 text-encoder.
136
+ - The non-pooled output of the text encoder is fed into the UNet backbone of the latent diffusion model via cross-attention.
137
+ - The loss is a reconstruction objective between the noise that was added to the latent and the prediction made by the UNet.
138
+
139
+ We currently provide three checkpoints, `sd-v1-1.ckpt`, `sd-v1-2.ckpt` and `sd-v1-3.ckpt`,
140
+ which were trained as follows,
141
+
142
+ - `sd-v1-1.ckpt`: 237k steps at resolution `256x256` on [laion2B-en](https://huggingface.co/datasets/laion/laion2B-en).
143
+ 194k steps at resolution `512x512` on [laion-high-resolution](https://huggingface.co/datasets/laion/laion-high-resolution) (170M examples from LAION-5B with resolution `>= 1024x1024`).
144
+ - `sd-v1-2.ckpt`: Resumed from `sd-v1-1.ckpt`.
145
+ 515k steps at res---
146
+ license: other
147
+ tags:
148
+ - stable-diffusion
149
+ - text-to-image
150
+ inference: false
151
+ ---
152
+
153
+ # Stable Diffusion v1 Model Card
154
+ This model card focuses on the model associated with the Stable Diffusion model, available [here](https://github.com/CompVis/stable-diffusion).
155
+
156
+ ## Model Details
157
+ - **Developed by:** Robin Rombach, Patrick Esser
158
+ - **Model type:** Diffusion-based text-to-image generation model
159
+ - **Language(s):** English
160
+ - **License:** [Proprietary](LICENSE)
161
+ - **Model Description:** This is a model that can be used to generate and modify images based on text prompts. It is a [Latent Diffusion Model](https://arxiv.org/abs/2112.10752) that uses a fixed, pretrained text encoder ([CLIP ViT-L/14](https://arxiv.org/abs/2103.00020)) as suggested in the [Imagen paper](https://arxiv.org/abs/2205.11487).
162
+ - **Resources for more information:** [GitHub Repository](https://github.com/CompVis/stable-diffusion), [Paper](https://arxiv.org/abs/2112.10752).
163
+ - **Cite as:**
164
+
165
+ @InProceedings{Rombach_2022_CVPR,
166
+ author = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn},
167
+ title = {High-Resolution Image Synthesis With Latent Diffusion Models},
168
+ booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
169
+ month = {June},
170
+ year = {2022},
171
+ pages = {10684-10695}
172
+ }
173
+
174
+ ## Usage examples
175
+
176
+ ```bash
177
+ pip install --upgrade diffusers transformers scipy
178
+ ```
179
+
180
+ Run this command to log in with your HF Hub token if you haven't before:
181
+ ```bash
182
+ huggingface-cli login
183
+ ```
184
+
185
+ Running the pipeline with the default PLMS scheduler:
186
+ ```python
187
+ from torch import autocast
188
+ from diffusers import StableDiffusionPipeline
189
+
190
+ model_id = "CompVis/stable-diffusion-v1-3-diffusers"
191
+ pipe = StableDiffusionPipeline.from_pretrained(model_id, use_auth_token=True).to("cuda")
192
+
193
+ prompt = "a photograph of an astronaut riding a horse"
194
+ with autocast("cuda"):
195
+ image = pipe(prompt, guidance_scale=7)["sample"][0] # image here is in PIL format
196
+
197
+ image.save(f"astronaut_rides_horse.png")
198
+ ```
199
+
200
+ To swap out the noise scheduler, pass it to `from_pretrained`:
201
+
202
+ ```python
203
+ from diffusers import StableDiffusionPipeline, LMSDiscreteScheduler
204
+
205
+ model_id = "CompVis/stable-diffusion-v1-3-diffusers"
206
+ # Use the K-LMS scheduler here instead
207
+ scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000)
208
+ pipe = StableDiffusionPipeline.from_pretrained(model_id, scheduler=scheduler, use_auth_token=True).to("cuda")
209
+ ```
210
+
211
+ # Uses
212
+
213
+ ## Direct Use
214
+ The model is intended for research purposes only. Possible research areas and
215
+ tasks include
216
+
217
+ - Safe deployment of models which have the potential to generate harmful content.
218
+ - Probing and understanding the limitations and biases of generative models.
219
+ - Generation of artworks and use in design and other artistic processes.
220
+ - Applications in educational or creative tools.
221
+ - Research on generative models.
222
+
223
+ Excluded uses are described below.
224
+
225
+ ### Misuse, Malicious Use, and Out-of-Scope Use
226
+ _Note: This section is taken from the [DALLE-MINI model card](https://huggingface.co/dalle-mini/dalle-mini), but applies in the same way to Stable Diffusion v1_.
227
+
228
+
229
+ The model should not be used to intentionally create or disseminate images that create hostile or alienating environments for people. This includes generating images that people would foreseeably find disturbing, distressing, or offensive; or content that propagates historical or current stereotypes.
230
+ #### Out-of-Scope Use
231
+ The model was not trained to be factual or true representations of people or events, and therefore using the model to generate such content is out-of-scope for the abilities of this model.
232
+ #### Misuse and Malicious Use
233
+ Using the model to generate content that is cruel to individuals is a misuse of this model. This includes, but is not limited to:
234
+
235
+ - Generating demeaning, dehumanizing, or otherwise harmful representations of people or their environments, cultures, religions, etc.
236
+ - Intentionally promoting or propagating discriminatory content or harmful stereotypes.
237
+ - Impersonating individuals without their consent.
238
+ - Sexual content without consent of the people who might see it.
239
+ - Mis- and disinformation
240
+ - Representations of egregious violence and gore
241
+ - Sharing of copyrighted or licensed material in violation of its terms of use.
242
+ - Sharing content that is an alteration of copyrighted or licensed material in violation of its terms of use.
243
+
244
+ ## Limitations and Bias
245
+
246
+ ### Limitations
247
+
248
+ - The model does not achieve perfect photorealism
249
+ - The model cannot render legible text
250
+ - The model does not perform well on more difficult tasks which involve compositionality, such as rendering an image corresponding to “A red cube on top of a blue sphere”
251
+ - Faces and people in general may not be generated properly.
252
+ - The model was trained mainly with English captions and will not work as well in other languages.
253
+ - The autoencoding part of the model is lossy
254
+ - The model was trained on a large-scale dataset
255
+ [LAION-5B](https://laion.ai/blog/laion-5b/) which contains adult material
256
+ and is not fit for product use without additional safety mechanisms and
257
+ considerations.
258
+
259
+ ### Bias
260
+ While the capabilities of image generation models are impressive, they can also reinforce or exacerbate social biases.
261
+ Stable Diffusion v1 was trained on subsets of [LAION-2B(en)](https://laion.ai/blog/laion-5b/),
262
+ which consists of images that are primarily limited to English descriptions.
263
+ Texts and images from communities and cultures that use other languages are likely to be insufficiently accounted for.
264
+ This affects the overall output of the model, as white and western cultures are often set as the default. Further, the
265
+ ability of the model to generate content with non-English prompts is significantly worse than with English-language prompts.
266
+
267
+
268
+ ## Training
269
+
270
+ **Training Data**
271
+ The model developers used the following dataset for training the model:
272
+
273
+ - LAION-2B (en) and subsets thereof (see next section)
274
+
275
+ **Training Procedure**
276
+ Stable Diffusion v1 is a latent diffusion model which combines an autoencoder with a diffusion model that is trained in the latent space of the autoencoder. During training,
277
+
278
+ - Images are encoded through an encoder, which turns images into latent representations. The autoencoder uses a relative downsampling factor of 8 and maps images of shape H x W x 3 to latents of shape H/f x W/f x 4
279
+ - Text prompts are encoded through a ViT-L/14 text-encoder.
280
+ - The non-pooled output of the text encoder is fed into the UNet backbone of the latent diffusion model via cross-attention.
281
+ - The loss is a reconstruction objective between the noise that was added to the latent and the prediction made by the UNet.
282
+
283
+ We currently provide three checkpoints, `sd-v1-1.ckpt`, `sd-v1-2.ckpt` and `sd-v1-3.ckpt`,
284
+ which were trained as follows,
285
+
286
+ - `sd-v1-1.ckpt`: 237k steps at resolution `256x256` on [laion2B-en](https://huggingface.co/datasets/laion/laion2B-en).
287
+ 194k steps at resolution `512x512` on [laion-high-resolution](https://huggingface.co/datasets/laion/laion-high-resolution) (170M examples from LAION-5B with resolution `>= 1024x1024`).
288
+ - `sd-v1-2.ckpt`: Resumed from `sd-v1-1.ckpt`.
289
+ 515k steps at resolution `512x512` on "laion-improved-aesthetics" (a subset of laion2B-en,
290
+ filtered to images with an original size `>= 512x512`, estimated aesthetics score `> 5.0`, and an estimated watermark probability `< 0.5`. The watermark estimate is from the LAION-5B metadata, the aesthetics score is estimated using an [improved aesthetics estimator](https://github.com/christophschuhmann/improved-aesthetic-predictor)).
291
+ - `sd-v1-3.ckpt`: Resumed from `sd-v1-2.ckpt`. 195k steps at resolution `512x512` on "laion-improved-aesthetics" and 10\% dropping of the text-conditioning to improve [classifier-free guidance sampling](https://arxiv.org/abs/2207.12598).
292
+ - `sd-v1-4.ckpt`: Resumed from `sd-v1-2.ckpt`. 195k steps at resolution `512x512` on "laion-improved-aesthetics" and 10\% dropping of the text-conditioning to improve [classifier-free guidance sampling](https://arxiv.org/abs/2207.12598).
293
+
294
+
295
+ - **Hardware:** 32 x 8 x A100 GPUs
296
+ - **Optimizer:** AdamW
297
+ - **Gradient Accumulations**: 2
298
+ - **Batch:** 32 x 8 x 2 x 4 = 2048
299
+ - **Learning rate:** warmup to 0.0001 for 10,000 steps and then kept constant
300
+
301
+ ## Evaluation Results
302
+ Evaluations with different classifier-free guidance scales (1.5, 2.0, 3.0, 4.0,
303
+ 5.0, 6.0, 7.0, 8.0) and 50 PLMS sampling
304
+ steps show the relative improvements of the checkpoints:
305
+
306
+ ![pareto](v1-variants-scores.jpg)
307
+
308
+ Evaluated using 50 PLMS steps and 10000 random prompts from the COCO2017 validation set, evaluated at 512x512 resolution. Not optimized for FID scores.
309
+ ## Environmental Impact
310
+
311
+ **Stable Diffusion v1** **Estimated Emissions**
312
+ Based on that information, we estimate the following CO2 emissions using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700). The hardware, runtime, cloud provider, and compute region were utilized to estimate the carbon impact.
313
+
314
+ - **Hardware Type:** A100 PCIe 40GB
315
+ - **Hours used:** 150000
316
+ - **Cloud Provider:** AWS
317
+ - **Compute Region:** US-east
318
+ - **Carbon Emitted (Power consumption x Time x Carbon produced based on location of power grid):** 11250 kg CO2 eq.
319
+
320
+ ## Usage
321
+
322
+ ### Setup
323
+
324
+ - Install `diffusers` with
325
+
326
+ `pip install -U git+https://github.com/huggingface/diffusers.git`
327
+ - Install `transformers` with
328
+
329
+ `pip install transformers`
330
+
331
+ ```python
332
+ import torch
333
+ from diffusers import StableDiffusionPipeline
334
+
335
+ pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-3-diffusers")
336
+
337
+ prompt = "19th Century wooden engraving of Elon musk"
338
+
339
+ seed = torch.manual_seed(1024)
340
+ images = pipe([prompt], num_inference_steps=50, guidance_scale=7.5, generator=seed)["sample"]
341
+
342
+ # save images
343
+ for idx, image in enumerate(images):
344
+ image.save(f"image-{idx}.png")
345
+ ```
346
+
347
+
348
+ ## Citation
349
+
350
+ ```bibtex
351
+ @InProceedings{Rombach_2022_CVPR,
352
+ author = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn},
353
+ title = {High-Resolution Image Synthesis With Latent Diffusion Models},
354
+ booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
355
+ month = {June},
356
+ year = {2022},
357
+ pages = {10684-10695}
358
+ }
359
+ ```
360
+
361
+ *This model card was written by: Robin Rombach and Patrick Esser and is based on the [DALL-E Mini model card](https://huggingface.co/dalle-mini/dalle-mini).*
362
+ olution `512x512` on "laion-improved-aesthetics" (a subset of laion2B-en,
363
+ filtered to images with an original size `>= 512x512`, estimated aesthetics score `> 5.0`, and an estimated watermark probability `< 0.5`. The watermark estimate is from the LAION-5B metadata, the aesthetics score is estimated using an [improved aesthetics estimator](https://github.com/christophschuhmann/improved-aesthetic-predictor)).
364
+ - `sd-v1-3.ckpt`: Resumed from `sd-v1-2.ckpt`. 195k steps at resolution `512x512` on "laion-improved-aesthetics" and 10\% dropping of the text-conditioning to improve [classifier-free guidance sampling](https://arxiv.org/abs/2207.12598).
365
+
366
+
367
+ - **Hardware:** 32 x 8 x A100 GPUs
368
+ - **Optimizer:** AdamW
369
+ - **Gradient Accumulations**: 2
370
+ - **Batch:** 32 x 8 x 2 x 4 = 2048
371
+ - **Learning rate:** warmup to 0.0001 for 10,000 steps and then kept constant
372
+
373
+ ## Evaluation Results
374
+ Evaluations with different classifier-free guidance scales (1.5, 2.0, 3.0, 4.0,
375
+ 5.0, 6.0, 7.0, 8.0) and 50 PLMS sampling
376
+ steps show the relative improvements of the checkpoints:
377
+
378
+ ![pareto](v1-variants-scores.jpg)
379
+
380
+ Evaluated using 50 PLMS steps and 10000 random prompts from the COCO2017 validation set, evaluated at 512x512 resolution. Not optimized for FID scores.
381
+ ## Environmental Impact
382
+
383
+ **Stable Diffusion v1** **Estimated Emissions**
384
+ Based on that information, we estimate the following CO2 emissions using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700). The hardware, runtime, cloud provider, and compute region were utilized to estimate the carbon impact.
385
+
386
+ - **Hardware Type:** A100 PCIe 40GB
387
+ - **Hours used:** 150000
388
+ - **Cloud Provider:** AWS
389
+ - **Compute Region:** US-east
390
+ - **Carbon Emitted (Power consumption x Time x Carbon produced based on location of power grid):** 11250 kg CO2 eq.
391
+
392
+ ## Usage
393
+
394
+ ### Setup
395
+
396
+ - Install `diffusers` with
397
+
398
+ `pip install -U git+https://github.com/huggingface/diffusers.git`
399
+ - Install `transformers` with
400
+
401
+ `pip install transformers`
402
+
403
+ ```python
404
+ import torch
405
+ from diffusers import StableDiffusionPipeline
406
+
407
+ pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-3-diffusers")
408
+
409
+ prompt = "19th Century wooden engraving of Elon musk"
410
+
411
+ seed = torch.manual_seed(1024)
412
+ images = pipe([prompt], num_inference_steps=50, guidance_scale=7.5, generator=seed)["sample"]
413
+
414
+ # save images
415
+ for idx, image in enumerate(images):
416
+ image.save(f"image-{idx}.png")
417
+ ```
418
+
419
+
420
+ ## Citation
421
+
422
+ ```bibtex
423
+ @InProceedings{Rombach_2022_CVPR,
424
+ author = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn},
425
+ title = {High-Resolution Image Synthesis With Latent Diffusion Models},
426
+ booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
427
+ month = {June},
428
+ year = {2022},
429
+ pages = {10684-10695}
430
+ }
431
+ ```
432
+
433
+ *This model card was written by: Robin Rombach and Patrick Esser and is based on the [DALL-E Mini model card](https://huggingface.co/dalle-mini/dalle-mini).*