README.md
16.7 KB · 324 lines · markdown Raw
1 ---
2 license: creativeml-openrail-m
3 tags:
4 - stable-diffusion
5 - stable-diffusion-diffusers
6 - text-to-image
7 widget:
8 - text: "A high tech solarpunk utopia in the Amazon rainforest"
9 example_title: Amazon rainforest
10 - text: "A pikachu fine dining with a view to the Eiffel Tower"
11 example_title: Pikachu in Paris
12 - text: "A mecha robot in a favela in expressionist style"
13 example_title: Expressionist robot
14 - text: "an insect robot preparing a delicious meal"
15 example_title: Insect robot
16 - text: "A small cabin on top of a snowy mountain in the style of Disney, artstation"
17 example_title: Snowy disney cabin
18 extra_gated_prompt: |-
19 This model is open access and available to all, with a CreativeML OpenRAIL-M license further specifying rights and usage.
20 The CreativeML OpenRAIL License specifies:
21
22 1. You can't use the model to deliberately produce nor share illegal or harmful outputs or content
23 2. The authors claim no rights on the outputs you generate, you are free to use them and are accountable for their use which must not go against the provisions set in the license
24 3. You may re-distribute the weights and use the model commercially and/or as a service. If you do, please be aware you have to include the same use restrictions as the ones in the license and share a copy of the CreativeML OpenRAIL-M to all your users (please read the license entirely and carefully)
25 Please read the full license carefully here: https://huggingface.co/spaces/CompVis/stable-diffusion-license
26
27 extra_gated_heading: Please read the LICENSE to access this model
28 ---
29
30 # Stable Diffusion v1-4 Model Card
31
32 Stable Diffusion is a latent text-to-image diffusion model capable of generating photo-realistic images given any text input.
33 For more information about how Stable Diffusion functions, please have a look at [🤗's Stable Diffusion with 🧨Diffusers blog](https://huggingface.co/blog/stable_diffusion).
34
35 The **Stable-Diffusion-v1-4** checkpoint was initialized with the weights of the [Stable-Diffusion-v1-2](https:/steps/huggingface.co/CompVis/stable-diffusion-v1-2)
36 checkpoint and subsequently fine-tuned on 225k steps at resolution 512x512 on "laion-aesthetics v2 5+" and 10% dropping of the text-conditioning to improve [classifier-free guidance sampling](https://arxiv.org/abs/2207.12598).
37
38 This weights here are intended to be used with the 🧨 Diffusers library. If you are looking for the weights to be loaded into the CompVis Stable Diffusion codebase, [come here](https://huggingface.co/CompVis/stable-diffusion-v-1-4-original)
39
40 ## Model Details
41 - **Developed by:** Robin Rombach, Patrick Esser
42 - **Model type:** Diffusion-based text-to-image generation model
43 - **Language(s):** English
44 - **License:** [The CreativeML OpenRAIL M license](https://huggingface.co/spaces/CompVis/stable-diffusion-license) is an [Open RAIL M license](https://www.licenses.ai/blog/2022/8/18/naming-convention-of-responsible-ai-licenses), adapted from the work that [BigScience](https://bigscience.huggingface.co/) and [the RAIL Initiative](https://www.licenses.ai/) are jointly carrying in the area of responsible AI licensing. See also [the article about the BLOOM Open RAIL license](https://bigscience.huggingface.co/blog/the-bigscience-rail-license) on which our license is based.
45 - **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).
46 - **Resources for more information:** [GitHub Repository](https://github.com/CompVis/stable-diffusion), [Paper](https://arxiv.org/abs/2112.10752).
47 - **Cite as:**
48
49 @InProceedings{Rombach_2022_CVPR,
50 author = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn},
51 title = {High-Resolution Image Synthesis With Latent Diffusion Models},
52 booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
53 month = {June},
54 year = {2022},
55 pages = {10684-10695}
56 }
57
58 ## Examples
59
60 We recommend using [🤗's Diffusers library](https://github.com/huggingface/diffusers) to run Stable Diffusion.
61
62 ### PyTorch
63
64 ```bash
65 pip install --upgrade diffusers transformers scipy
66 ```
67
68 Running the pipeline with the default PNDM scheduler:
69
70 ```python
71 import torch
72 from diffusers import StableDiffusionPipeline
73
74 model_id = "CompVis/stable-diffusion-v1-4"
75 device = "cuda"
76
77
78 pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
79 pipe = pipe.to(device)
80
81 prompt = "a photo of an astronaut riding a horse on mars"
82 image = pipe(prompt).images[0]
83
84 image.save("astronaut_rides_horse.png")
85 ```
86
87 **Note**:
88 If you are limited by GPU memory and have less than 4GB of GPU RAM available, please make sure to load the StableDiffusionPipeline in float16 precision instead of the default float32 precision as done above. You can do so by telling diffusers to expect the weights to be in float16 precision:
89
90
91 ```py
92 import torch
93
94 pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
95 pipe = pipe.to(device)
96 pipe.enable_attention_slicing()
97
98 prompt = "a photo of an astronaut riding a horse on mars"
99 image = pipe(prompt).images[0]
100
101 image.save("astronaut_rides_horse.png")
102 ```
103
104 To swap out the noise scheduler, pass it to `from_pretrained`:
105
106 ```python
107 from diffusers import StableDiffusionPipeline, EulerDiscreteScheduler
108
109 model_id = "CompVis/stable-diffusion-v1-4"
110
111 # Use the Euler scheduler here instead
112 scheduler = EulerDiscreteScheduler.from_pretrained(model_id, subfolder="scheduler")
113 pipe = StableDiffusionPipeline.from_pretrained(model_id, scheduler=scheduler, torch_dtype=torch.float16)
114 pipe = pipe.to("cuda")
115
116 prompt = "a photo of an astronaut riding a horse on mars"
117 image = pipe(prompt).images[0]
118
119 image.save("astronaut_rides_horse.png")
120 ```
121
122 ### JAX/Flax
123
124 To use StableDiffusion on TPUs and GPUs for faster inference you can leverage JAX/Flax.
125
126 Running the pipeline with default PNDMScheduler
127
128 ```python
129 import jax
130 import numpy as np
131 from flax.jax_utils import replicate
132 from flax.training.common_utils import shard
133
134 from diffusers import FlaxStableDiffusionPipeline
135
136 pipeline, params = FlaxStableDiffusionPipeline.from_pretrained(
137 "CompVis/stable-diffusion-v1-4", revision="flax", dtype=jax.numpy.bfloat16
138 )
139
140 prompt = "a photo of an astronaut riding a horse on mars"
141
142 prng_seed = jax.random.PRNGKey(0)
143 num_inference_steps = 50
144
145 num_samples = jax.device_count()
146 prompt = num_samples * [prompt]
147 prompt_ids = pipeline.prepare_inputs(prompt)
148
149 # shard inputs and rng
150 params = replicate(params)
151 prng_seed = jax.random.split(prng_seed, num_samples)
152 prompt_ids = shard(prompt_ids)
153
154 images = pipeline(prompt_ids, params, prng_seed, num_inference_steps, jit=True).images
155 images = pipeline.numpy_to_pil(np.asarray(images.reshape((num_samples,) + images.shape[-3:])))
156 ```
157
158 **Note**:
159 If you are limited by TPU memory, please make sure to load the `FlaxStableDiffusionPipeline` in `bfloat16` precision instead of the default `float32` precision as done above. You can do so by telling diffusers to load the weights from "bf16" branch.
160
161 ```python
162 import jax
163 import numpy as np
164 from flax.jax_utils import replicate
165 from flax.training.common_utils import shard
166
167 from diffusers import FlaxStableDiffusionPipeline
168
169 pipeline, params = FlaxStableDiffusionPipeline.from_pretrained(
170 "CompVis/stable-diffusion-v1-4", revision="bf16", dtype=jax.numpy.bfloat16
171 )
172
173 prompt = "a photo of an astronaut riding a horse on mars"
174
175 prng_seed = jax.random.PRNGKey(0)
176 num_inference_steps = 50
177
178 num_samples = jax.device_count()
179 prompt = num_samples * [prompt]
180 prompt_ids = pipeline.prepare_inputs(prompt)
181
182 # shard inputs and rng
183 params = replicate(params)
184 prng_seed = jax.random.split(prng_seed, num_samples)
185 prompt_ids = shard(prompt_ids)
186
187 images = pipeline(prompt_ids, params, prng_seed, num_inference_steps, jit=True).images
188 images = pipeline.numpy_to_pil(np.asarray(images.reshape((num_samples,) + images.shape[-3:])))
189 ```
190
191 # Uses
192
193 ## Direct Use
194 The model is intended for research purposes only. Possible research areas and
195 tasks include
196
197 - Safe deployment of models which have the potential to generate harmful content.
198 - Probing and understanding the limitations and biases of generative models.
199 - Generation of artworks and use in design and other artistic processes.
200 - Applications in educational or creative tools.
201 - Research on generative models.
202
203 Excluded uses are described below.
204
205 ### Misuse, Malicious Use, and Out-of-Scope Use
206 _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_.
207
208
209 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.
210
211 #### Out-of-Scope Use
212 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.
213
214 #### Misuse and Malicious Use
215 Using the model to generate content that is cruel to individuals is a misuse of this model. This includes, but is not limited to:
216
217 - Generating demeaning, dehumanizing, or otherwise harmful representations of people or their environments, cultures, religions, etc.
218 - Intentionally promoting or propagating discriminatory content or harmful stereotypes.
219 - Impersonating individuals without their consent.
220 - Sexual content without consent of the people who might see it.
221 - Mis- and disinformation
222 - Representations of egregious violence and gore
223 - Sharing of copyrighted or licensed material in violation of its terms of use.
224 - Sharing content that is an alteration of copyrighted or licensed material in violation of its terms of use.
225
226 ## Limitations and Bias
227
228 ### Limitations
229
230 - The model does not achieve perfect photorealism
231 - The model cannot render legible text
232 - 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”
233 - Faces and people in general may not be generated properly.
234 - The model was trained mainly with English captions and will not work as well in other languages.
235 - The autoencoding part of the model is lossy
236 - The model was trained on a large-scale dataset
237 [LAION-5B](https://laion.ai/blog/laion-5b/) which contains adult material
238 and is not fit for product use without additional safety mechanisms and
239 considerations.
240 - No additional measures were used to deduplicate the dataset. As a result, we observe some degree of memorization for images that are duplicated in the training data.
241 The training data can be searched at [https://rom1504.github.io/clip-retrieval/](https://rom1504.github.io/clip-retrieval/) to possibly assist in the detection of memorized images.
242
243 ### Bias
244
245 While the capabilities of image generation models are impressive, they can also reinforce or exacerbate social biases.
246 Stable Diffusion v1 was trained on subsets of [LAION-2B(en)](https://laion.ai/blog/laion-5b/),
247 which consists of images that are primarily limited to English descriptions.
248 Texts and images from communities and cultures that use other languages are likely to be insufficiently accounted for.
249 This affects the overall output of the model, as white and western cultures are often set as the default. Further, the
250 ability of the model to generate content with non-English prompts is significantly worse than with English-language prompts.
251
252 ### Safety Module
253
254 The intended use of this model is with the [Safety Checker](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/safety_checker.py) in Diffusers.
255 This checker works by checking model outputs against known hard-coded NSFW concepts.
256 The concepts are intentionally hidden to reduce the likelihood of reverse-engineering this filter.
257 Specifically, the checker compares the class probability of harmful concepts in the embedding space of the `CLIPTextModel` *after generation* of the images.
258 The concepts are passed into the model with the generated image and compared to a hand-engineered weight for each NSFW concept.
259
260
261 ## Training
262
263 **Training Data**
264 The model developers used the following dataset for training the model:
265
266 - LAION-2B (en) and subsets thereof (see next section)
267
268 **Training Procedure**
269 Stable Diffusion v1-4 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,
270
271 - 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
272 - Text prompts are encoded through a ViT-L/14 text-encoder.
273 - The non-pooled output of the text encoder is fed into the UNet backbone of the latent diffusion model via cross-attention.
274 - The loss is a reconstruction objective between the noise that was added to the latent and the prediction made by the UNet.
275
276 We currently provide four checkpoints, which were trained as follows.
277 - [`stable-diffusion-v1-1`](https://huggingface.co/CompVis/stable-diffusion-v1-1): 237,000 steps at resolution `256x256` on [laion2B-en](https://huggingface.co/datasets/laion/laion2B-en).
278 194,000 steps at resolution `512x512` on [laion-high-resolution](https://huggingface.co/datasets/laion/laion-high-resolution) (170M examples from LAION-5B with resolution `>= 1024x1024`).
279 - [`stable-diffusion-v1-2`](https://huggingface.co/CompVis/stable-diffusion-v1-2): Resumed from `stable-diffusion-v1-1`.
280 515,000 steps at resolution `512x512` on "laion-improved-aesthetics" (a subset of laion2B-en,
281 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)).
282 - [`stable-diffusion-v1-3`](https://huggingface.co/CompVis/stable-diffusion-v1-3): Resumed from `stable-diffusion-v1-2`. 195,000 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).
283 - [`stable-diffusion-v1-4`](https://huggingface.co/CompVis/stable-diffusion-v1-4) Resumed from `stable-diffusion-v1-2`.225,000 steps at resolution `512x512` on "laion-aesthetics v2 5+" and 10 % dropping of the text-conditioning to improve [classifier-free guidance sampling](https://arxiv.org/abs/2207.12598).
284
285 - **Hardware:** 32 x 8 x A100 GPUs
286 - **Optimizer:** AdamW
287 - **Gradient Accumulations**: 2
288 - **Batch:** 32 x 8 x 2 x 4 = 2048
289 - **Learning rate:** warmup to 0.0001 for 10,000 steps and then kept constant
290
291 ## Evaluation Results
292 Evaluations with different classifier-free guidance scales (1.5, 2.0, 3.0, 4.0,
293 5.0, 6.0, 7.0, 8.0) and 50 PLMS sampling
294 steps show the relative improvements of the checkpoints:
295
296 ![pareto](https://huggingface.co/CompVis/stable-diffusion/resolve/main/v1-variants-scores.jpg)
297
298 Evaluated using 50 PLMS steps and 10000 random prompts from the COCO2017 validation set, evaluated at 512x512 resolution. Not optimized for FID scores.
299 ## Environmental Impact
300
301 **Stable Diffusion v1** **Estimated Emissions**
302 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.
303
304 - **Hardware Type:** A100 PCIe 40GB
305 - **Hours used:** 150000
306 - **Cloud Provider:** AWS
307 - **Compute Region:** US-east
308 - **Carbon Emitted (Power consumption x Time x Carbon produced based on location of power grid):** 11250 kg CO2 eq.
309
310
311 ## Citation
312
313 ```bibtex
314 @InProceedings{Rombach_2022_CVPR,
315 author = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn},
316 title = {High-Resolution Image Synthesis With Latent Diffusion Models},
317 booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
318 month = {June},
319 year = {2022},
320 pages = {10684-10695}
321 }
322 ```
323
324 *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).*