<!--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.
-->
# Textual inversion
[[open-in-colab]]
The [`StableDiffusionPipeline`] supports textual inversion, a technique that enables a model like Stable Diffusion to learn a new concept from just a few sample images. This gives you more control over the generated images and allows you to tailor the model towards specific concepts. You can get started quickly with a collection of community created concepts in the [Stable Diffusion Conceptualizer](https://huggingface.co/spaces/sd-concepts-library/stable-diffusion-conceptualizer).
This guide will show you how to run inference with textual inversion using a pre-learned concept from the Stable Diffusion Conceptualizer. If you're interested in teaching a model new concepts with textual inversion, take a look at the [Textual Inversion](../training/text_inversion) training guide.
Import the necessary libraries:
```py
importtorch
fromdiffusersimportStableDiffusionPipeline
fromdiffusers.utilsimportmake_image_grid
```
## Stable Diffusion 1 and 2
Pick a Stable Diffusion checkpoint and a pre-learned concept from the [Stable Diffusion Conceptualizer](https://huggingface.co/spaces/sd-concepts-library/stable-diffusion-conceptualizer):
Create a prompt with the pre-learned concept by using the special placeholder token `<cat-toy>`, and choose the number of samples and rows of images you'd like to generate:
```py
prompt="a grafitti in a favela wall with a <cat-toy> on it"
num_samples_per_row=2
num_rows=2
```
Then run the pipeline (feel free to adjust the parameters like `num_inference_steps` and `guidance_scale` to see how they affect image quality), save the generated images and visualize them with the helper function you created at the beginning:
Stable Diffusion XL (SDXL) can also use textual inversion vectors for inference. In contrast to Stable Diffusion 1 and 2, SDXL has two text encoders so you'll need two textual inversion embeddings - one for each text encoder model.
Let's download the SDXL textual inversion embeddings and have a closer look at it's structure:
<!--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.
-->
# Unconditional image generation
[[open-in-colab]]
Unconditional image generation generates images that look like a random sample from the training data the model was trained on because the denoising process is not guided by any additional context like text or image.
To get started, use the [`DiffusionPipeline`] to load the [anton-l/ddpm-butterflies-128](https://huggingface.co/anton-l/ddpm-butterflies-128) checkpoint to generate images of butterflies. The [`DiffusionPipeline`] downloads and caches all the model components required to generate an image.
Want to generate images of something else? Take a look at the training [guide](../training/unconditional_training) to learn how to train a model to generate your own images.
</Tip>
The output image is a [`PIL.Image`](https://pillow.readthedocs.io/en/stable/reference/Image.html?highlight=image#the-image-class) object that can be saved:
```py
image.save("generated_image.png")
```
You can also try experimenting with the `num_inference_steps` parameter, which controls the number of denoising steps. More denoising steps typically produce higher quality images, but it'll take longer to generate. Feel free to play around with this parameter to see how it affects the image quality.
<!--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.
-->
# Load safetensors
[[open-in-colab]]
[safetensors](https://github.com/huggingface/safetensors) is a safe and fast file format for storing and loading tensors. Typically, PyTorch model weights are saved or *pickled* into a `.bin` file with Python's [`pickle`](https://docs.python.org/3/library/pickle.html) utility. However, `pickle` is not secure and pickled files may contain malicious code that can be executed. safetensors is a secure alternative to `pickle`, making it ideal for sharing model weights.
This guide will show you how you load `.safetensor` files, and how to convert Stable Diffusion model weights stored in other formats to `.safetensor`. Before you start, make sure you have safetensors installed:
```py
# uncomment to install the necessary libraries in Colab
#!pip install safetensors
```
If you look at the [`runwayml/stable-diffusion-v1-5`](https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/main) repository, you'll see weights inside the `text_encoder`, `unet` and `vae` subfolders are stored in the `.safetensors` format. By default, 🤗 Diffusers automatically loads these `.safetensors` files from their subfolders if they're available in the model repository.
For more explicit control, you can optionally set `use_safetensors=True` (if `safetensors` is not installed, you'll get an error message asking you to install it):
However, model weights are not necessarily stored in separate subfolders like in the example above. Sometimes, all the weights are stored in a single `.safetensors` file. In this case, if the weights are Stable Diffusion weights, you can load the file directly with the [`~diffusers.loaders.FromSingleFileMixin.from_single_file`] method:
Not all weights on the Hub are available in the `.safetensors` format, and you may encounter weights stored as `.bin`. In this case, use the [Convert Space](https://huggingface.co/spaces/diffusers/convert) to convert the weights to `.safetensors`. The Convert Space downloads the pickled weights, converts them, and opens a Pull Request to upload the newly converted `.safetensors` file on the Hub. This way, if there is any malicious code contained in the pickled files, they're uploaded to the Hub - which has a [security scanner](https://huggingface.co/docs/hub/security-pickle#hubs-security-scanner) to detect unsafe files and suspicious pickle imports - instead of your computer.
You can use the model with the new `.safetensors` weights by specifying the reference to the Pull Request in the `revision` parameter (you can also test it in this [Check PR](https://huggingface.co/spaces/diffusers/check_pr) Space on the Hub), for example `refs/pr/22`:
- Safety is the number one reason for using safetensors. As open-source and model distribution grows, it is important to be able to trust the model weights you downloaded don't contain any malicious code. The current size of the header in safetensors prevents parsing extremely large JSON files.
- Loading speed between switching models is another reason to use safetensors, which performs zero-copy of the tensors. It is especially fast compared to `pickle` if you're loading the weights to CPU (the default case), and just as fast if not faster when directly loading the weights to GPU. You'll only notice the performance difference if the model is already loaded, and not if you're downloading the weights or loading the model for the first time.
But the actual time it takes to load 500MB of the model weights is only:
```bash
safetensors: 3.4873ms
PyTorch: 172.7537ms
```
- Lazy loading is also supported in safetensors, which is useful in distributed settings to only load some of the tensors. This format allowed the [BLOOM](https://huggingface.co/bigscience/bloom) model to be loaded in 45 seconds on 8 GPUs instead of 10 minutes with regular PyTorch weights.
<!--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.
-->
# Prompt weighting
[[open-in-colab]]
Prompt weighting provides a way to emphasize or de-emphasize certain parts of a prompt, allowing for more control over the generated image. A prompt can include several concepts, which gets turned into contextualized text embeddings. The embeddings are used by the model to condition its cross-attention layers to generate an image (read the Stable Diffusion [blog post](https://huggingface.co/blog/stable_diffusion) to learn more about how it works).
Prompt weighting works by increasing or decreasing the scale of the text embedding vector that corresponds to its concept in the prompt because you may not necessarily want the model to focus on all concepts equally. The easiest way to prepare the prompt-weighted embeddings is to use [Compel](https://github.com/damian0815/compel), a text prompt-weighting and blending library. Once you have the prompt-weighted embeddings, you can pass them to any pipeline that has a [`prompt_embeds`](https://huggingface.co/docs/diffusers/en/api/pipelines/stable_diffusion/text2img#diffusers.StableDiffusionPipeline.__call__.prompt_embeds)(and optionally [`negative_prompt_embeds`](https://huggingface.co/docs/diffusers/en/api/pipelines/stable_diffusion/text2img#diffusers.StableDiffusionPipeline.__call__.negative_prompt_embeds)) parameter, such as [`StableDiffusionPipeline`], [`StableDiffusionControlNetPipeline`], and [`StableDiffusionXLPipeline`].
<Tip>
If your favorite pipeline doesn't have a `prompt_embeds` parameter, please open an [issue](https://github.com/huggingface/diffusers/issues/new/choose) so we can add it!
</Tip>
This guide will show you how to weight and blend your prompts with Compel in 🤗 Diffusers.
Before you begin, make sure you have the latest version of Compel installed:
```py
# uncomment to install in Colab
#!pip install compel --upgrade
```
For this guide, let's generate an image with the prompt `"a red cat playing with a ball"` using the [`StableDiffusionPipeline`]:
You'll notice there is no "ball" in the image! Let's use compel to upweight the concept of "ball" in the prompt. Create a [`Compel`](https://github.com/damian0815/compel/blob/main/doc/compel.md#compel-objects) object, and pass it a tokenizer and text encoder:
compel uses `+` or `-` to increase or decrease the weight of a word in the prompt. To increase the weight of "ball":
<Tip>
`+` corresponds to the value `1.1`, `++` corresponds to `1.1^2`, and so on. Similarly, `-` corresponds to `0.9` and `--` corresponds to `0.9^2`. Feel free to experiment with adding more `+` or `-` in your prompt!
</Tip>
```py
prompt="a red cat playing with a ball++"
```
Pass the prompt to `compel_proc` to create the new prompt embeddings which are passed to the pipeline:
You can also create a weighted *blend* of prompts by adding `.blend()` to a list of prompts and passing it some weights. Your blend may not always produce the result you expect because it breaks some assumptions about how the text encoder functions, so just have fun and experiment with it!
```py
prompt_embeds=compel_proc('("a red cat playing with a ball", "jungle").blend(0.7, 0.8)')
A conjunction diffuses each prompt independently and concatenates their results by their weighted sum. Add `.and()` to the end of a list of prompts to create a conjunction:
```py
prompt_embeds=compel_proc('["a red cat", "playing with a", "ball"].and()')
[Textual inversion](../training/text_inversion) is a technique for learning a specific concept from some images which you can use to generate new images conditioned on that concept.
Create a pipeline and use the [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] function to load the textual inversion embeddings (feel free to browse the [Stable Diffusion Conceptualizer](https://huggingface.co/spaces/sd-concepts-library/stable-diffusion-conceptualizer) for 100+ trained concepts):
Compel provides a `DiffusersTextualInversionManager` class to simplify prompt weighting with textual inversion. Instantiate `DiffusersTextualInversionManager` and pass it to the `Compel` class:
[DreamBooth](../training/dreambooth) is a technique for generating contextualized images of a subject given just a few images of the subject to train on. It is similar to textual inversion, but DreamBooth trains the full model whereas textual inversion only fine-tunes the text embeddings. This means you should use [`~DiffusionPipeline.from_pretrained`] to load the DreamBooth model (feel free to browse the [Stable Diffusion Dreambooth Concepts Library](https://huggingface.co/sd-dreambooth-library) for 100+ trained models):
Create a `Compel` class with a tokenizer and text encoder, and pass your prompt to it. Depending on the model you use, you'll need to incorporate the model's unique identifier into your prompt. For example, the `dndcoverart-v1` model uses the identifier `dndcoverart`:
Stable Diffusion XL (SDXL) has two tokenizers and text encoders so it's usage is a bit different. To address this, you should pass both tokenizers and encoders to the `Compel` class:
This time, let's upweight "ball" by a factor of 1.5 for the first prompt, and downweight "ball" by 0.6 for the second prompt. The [`StableDiffusionXLPipeline`] also requires [`pooled_prompt_embeds`](https://huggingface.co/docs/diffusers/en/api/pipelines/stable_diffusion/stable_diffusion_xl#diffusers.StableDiffusionXLInpaintPipeline.__call__.pooled_prompt_embeds)(and optionally [`negative_pooled_prompt_embeds`](https://huggingface.co/docs/diffusers/en/api/pipelines/stable_diffusion/stable_diffusion_xl#diffusers.StableDiffusionXLInpaintPipeline.__call__.negative_pooled_prompt_embeds)) so you should pass those to the pipeline along with the conditioning tensors:
```py
# apply weights
prompt=["a red cat playing with a (ball)1.5","a red cat playing with a (ball)0.6"]
<!--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.
-->
# Understanding pipelines, models and schedulers
[[open-in-colab]]
🧨 Diffusers is designed to be a user-friendly and flexible toolbox for building diffusion systems tailored to your use-case. At the core of the toolbox are models and schedulers. While the [`DiffusionPipeline`] bundles these components together for convenience, you can also unbundle the pipeline and use the models and schedulers separately to create new diffusion systems.
In this tutorial, you'll learn how to use models and schedulers to assemble a diffusion system for inference, starting with a basic pipeline and then progressing to the Stable Diffusion pipeline.
## Deconstruct a basic pipeline
A pipeline is a quick and easy way to run a model for inference, requiring no more than four lines of code to generate an image:
<imgsrc="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/ddpm-cat.png"alt="Image of cat created from DDPMPipeline"/>
</div>
That was super easy, but how did the pipeline do that? Let's breakdown the pipeline and take a look at what's happening under the hood.
In the example above, the pipeline contains a [`UNet2DModel`] model and a [`DDPMScheduler`]. The pipeline denoises an image by taking random noise the size of the desired output and passing it through the model several times. At each timestep, the model predicts the *noise residual* and the scheduler uses it to predict a less noisy image. The pipeline repeats this process until it reaches the end of the specified number of inference steps.
To recreate the pipeline with the model and scheduler separately, let's write our own denoising process.
2. Set the number of timesteps to run the denoising process for:
```py
>>>scheduler.set_timesteps(50)
```
3. Setting the scheduler timesteps creates a tensor with evenly spaced elements in it, 50 in this example. Each element corresponds to a timestep at which the model denoises an image. When you create the denoising loop later, you'll iterate over this tensor to denoise an image:
5. Now write a loop to iterate over the timesteps. At each timestep, the model does a [`UNet2DModel.forward`] pass and returns the noisy residual. The scheduler's [`~DDPMScheduler.step`] method takes the noisy residual, timestep, and input and it predicts the image at the previous timestep. This output becomes the next input to the model in the denoising loop, and it'll repeat until it reaches the end of the `timesteps` array.
In the next section, you'll put your skills to the test and breakdown the more complex Stable Diffusion pipeline. The steps are more or less the same. You'll initialize the necessary components, and set the number of timesteps to create a `timestep` array. The `timestep` array is used in the denoising loop, and for each element in this array, the model predicts a less noisy image. The denoising loop iterates over the `timestep`'s, and at each timestep, it outputs a noisy residual and the scheduler uses it to predict a less noisy image at the previous timestep. This process is repeated until you reach the end of the `timestep` array.
Let's try it out!
## Deconstruct the Stable Diffusion pipeline
Stable Diffusion is a text-to-image *latent diffusion* model. It is called a latent diffusion model because it works with a lower-dimensional representation of the image instead of the actual pixel space, which makes it more memory efficient. The encoder compresses the image into a smaller representation, and a decoder to convert the compressed representation back into an image. For text-to-image models, you'll need a tokenizer and an encoder to generate text embeddings. From the previous example, you already know you need a UNet model and a scheduler.
As you can see, this is already more complex than the DDPM pipeline which only contains a UNet model. The Stable Diffusion model has three separate pretrained models.
<Tip>
💡 Read the [How does Stable Diffusion work?](https://huggingface.co/blog/stable_diffusion#how-does-stable-diffusion-work) blog for more details about how the VAE, UNet, and text encoder models work.
</Tip>
Now that you know what you need for the Stable Diffusion pipeline, load all these components with the [`~ModelMixin.from_pretrained`] method. You can find them in the pretrained [`runwayml/stable-diffusion-v1-5`](https://huggingface.co/runwayml/stable-diffusion-v1-5) checkpoint, and each component is stored in a separate subfolder:
To speed up inference, move the models to a GPU since, unlike the scheduler, they have trainable weights:
```py
>>>torch_device="cuda"
>>>vae.to(torch_device)
>>>text_encoder.to(torch_device)
>>>unet.to(torch_device)
```
### Create text embeddings
The next step is to tokenize the text to generate embeddings. The text is used to condition the UNet model and steer the diffusion process towards something that resembles the input prompt.
<Tip>
💡 The `guidance_scale` parameter determines how much weight should be given to the prompt when generating an image.
</Tip>
Feel free to choose any prompt you like if you want to generate something else!
```py
>>>prompt=["a photograph of an astronaut riding a horse"]
>>>height=512# default height of Stable Diffusion
>>>width=512# default width of Stable Diffusion
>>>num_inference_steps=25# Number of denoising steps
>>>guidance_scale=7.5# Scale for classifier-free guidance
>>>generator=torch.manual_seed(0)# Seed generator to create the initial latent noise
>>>batch_size=len(prompt)
```
Tokenize the text and generate the embeddings from the prompt:
You'll also need to generate the *unconditional text embeddings* which are the embeddings for the padding token. These need to have the same shape (`batch_size` and `seq_length`) as the conditional `text_embeddings`:
Next, generate some initial random noise as a starting point for the diffusion process. This is the latent representation of the image, and it'll be gradually denoised. At this point, the `latent` image is smaller than the final image size but that's okay though because the model will transform it into the final 512x512 image dimensions later.
<Tip>
💡 The height and width are divided by 8 because the `vae` model has 3 down-sampling layers. You can check by running the following:
Start by scaling the input with the initial noise distribution, *sigma*, the noise scale value, which is required for improved schedulers like [`UniPCMultistepScheduler`]:
```py
>>>latents=latents*scheduler.init_noise_sigma
```
The last step is to create the denoising loop that'll progressively transform the pure noise in `latents` to an image described by your prompt. Remember, the denoising loop needs to do three things:
1. Set the scheduler's timesteps to use during denoising.
2. Iterate over the timesteps.
3. At each timestep, call the UNet model to predict the noise residual and pass it to the scheduler to compute the previous noisy sample.
```py
>>>fromtqdm.autoimporttqdm
>>>scheduler.set_timesteps(num_inference_steps)
>>>fortintqdm(scheduler.timesteps):
...# expand the latents if we are doing classifier-free guidance to avoid doing two forward passes.
From basic to complex pipelines, you've seen that all you really need to write your own diffusion system is a denoising loop. The loop should set the scheduler's timesteps, iterate over them, and alternate between calling the UNet model to predict the noise residual and passing it to the scheduler to compute the previous noisy sample.
This is really what 🧨 Diffusers is designed for: to make it intuitive and easy to write your own diffusion system using models and schedulers.
For your next steps, feel free to:
* Learn how to [build and contribute a pipeline](../using-diffusers/contribute_pipeline) to 🧨 Diffusers. We can't wait and see what you'll come up with!
* Explore [existing pipelines](../api/pipelines/overview) in the library, and see if you can deconstruct and build a pipeline from scratch using the models and schedulers separately.
* 様々な使用例については、公式およびコミュニティの[training or finetuning scripts](https://github.com/huggingface/diffusers/tree/main/examples#-diffusers-examples)の例を参照してください。
* スケジューラのロード、アクセス、変更、比較については[Using different Schedulers](./using-diffusers/schedulers)ガイドを参照してください。
* 🧨 Diffusers の高速化については、最適化された [PyTorch on a GPU](./optimization/fp16)のガイド、[Stable Diffusion on Apple Silicon (M1/M2)](./optimization/mps)と[ONNX Runtime](./optimization/onnx)を参照してください。
"portrait photo of the oldest warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3 --beta --upbeta",
"portrait photo of a old warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3 --beta --upbeta",
"portrait photo of a warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3 --beta --upbeta",
"portrait photo of a young warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes 50mm portrait photography, hard rim lighting photography--beta --ar 2:3 --beta --upbeta",
]
generator = [torch.Generator("cuda").manual_seed(1) for _ in range(len(prompts))]
<imgsrc="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/autopipeline-text2img.png"alt="generated image of peasant fighting dragon in wood cutting style"/>
<imgsrc="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/autopipeline-img2img.png"alt="generated image of a vermeer portrait of a dog wearing a pearl earring"/>
<imgsrc="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/autopipeline-inpaint.png"alt="generated image of a tiger sitting on a bench"/>
<!--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.
-->
# Stable diffusion XL
Stable Diffusion XL은 Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, Robin Rombach에 의해 [SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis](https://arxiv.org/abs/2307.01952)에서 제안되었습니다.
논문 초록은 다음을 따릅니다:
*text-to-image의 latent diffusion 모델인 SDXL을 소개합니다. 이전 버전의 Stable Diffusion과 비교하면, SDXL은 세 배 더큰 규모의 UNet 백본을 포함합니다: 모델 파라미터의 증가는 많은 attention 블럭을 사용하고 더 큰 cross-attention context를 SDXL의 두 번째 텍스트 인코더에 사용하기 때문입니다. 다중 종횡비에 다수의 새로운 conditioning 방법을 구성했습니다. 또한 후에 수정하는 image-to-image 기술을 사용함으로써 SDXL에 의해 생성된 시각적 품질을 향상하기 위해 정제된 모델을 소개합니다. SDXL은 이전 버전의 Stable Diffusion보다 성능이 향상되었고, 이러한 black-box 최신 이미지 생성자와 경쟁력있는 결과를 달성했습니다.*
## 팁
- Stable Diffusion XL은 특히 786과 1024사이의 이미지에 잘 작동합니다.
- Stable Diffusion XL은 아래와 같이 학습된 각 텍스트 인코더에 대해 서로 다른 프롬프트를 전달할 수 있습니다. 동일한 프롬프트의 다른 부분을 텍스트 인코더에 전달할 수도 있습니다.
- Stable Diffusion XL 결과 이미지는 아래에 보여지듯이 정제기(refiner)를 사용함으로써 향상될 수 있습니다.
### 이용가능한 체크포인트:
-*Text-to-Image (1024x1024 해상도)*: [`StableDiffusionXLPipeline`]을 사용한 [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
-*Image-to-Image / 정제기(refiner) (1024x1024 해상도)*: [`StableDiffusionXLImg2ImgPipeline`]를 사용한 [stabilityai/stable-diffusion-xl-refiner-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0)
## 사용 예시
SDXL을 사용하기 전에 `transformers`, `accelerate`, `safetensors` 와 `invisible_watermark`를 설치하세요.
다음과 같이 라이브러리를 설치할 수 있습니다:
```
pip install transformers
pip install accelerate
pip install safetensors
pip install invisible-watermark>=0.2.0
```
### 워터마커
Stable Diffusion XL로 이미지를 생성할 때 워터마크가 보이지 않도록 추가하는 것을 권장하는데, 이는 다운스트림(downstream) 어플리케이션에서 기계에 합성되었는지를 식별하는데 도움을 줄 수 있습니다. 그렇게 하려면 [invisible_watermark 라이브러리](https://pypi.org/project/invisible-watermark/)를 통해 설치해주세요:
```
pip install invisible-watermark>=0.2.0
```
`invisible-watermark` 라이브러리가 설치되면 워터마커가 **기본적으로** 사용될 것입니다.
생성 또는 안전하게 이미지를 배포하기 위해 다른 규정이 있다면, 다음과 같이 워터마커를 비활성화할 수 있습니다:
[base 모델 체크포인트](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)에서, StableDiffusion-XL 또한 고주파 품질을 향상시키는 이미지를 생성하기 위해 낮은 노이즈 단계 이미지를 제거하는데 특화된 [refiner 체크포인트](huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0)를 포함하고 있습니다. 이 refiner 체크포인트는 이미지 품질을 향상시키기 위해 base 체크포인트를 실행한 후 "두 번째 단계" 파이프라인에 사용될 수 있습니다.
refiner를 사용할 때, 쉽게 사용할 수 있습니다
- 1.) base 모델과 refiner을 사용하는데, 이는 *Denoisers의 앙상블*을 위한 첫 번째 제안된 [eDiff-I](https://research.nvidia.com/labs/dir/eDiff-I/)를 사용하거나
- 2.) base 모델을 거친 후 [SDEdit](https://arxiv.org/abs/2108.01073) 방법으로 단순하게 refiner를 실행시킬 수 있습니다.
**참고**: SD-XL base와 refiner를 앙상블로 사용하는 아이디어는 커뮤니티 기여자들이 처음으로 제안했으며, 이는 다음과 같은 `diffusers`를 구현하는 데도 도움을 주셨습니다.
base와 refiner 모델을 denoiser의 앙상블로 사용할 때, base 모델은 고주파 diffusion 단계를 위한 전문가의 역할을 해야하고, refiner는 낮은 노이즈 diffusion 단계를 위한 전문가의 역할을 해야 합니다.
2.)에 비해 1.)의 장점은 전체적으로 denoising 단계가 덜 필요하므로 속도가 훨씬 더 빨라집니다. 단점은 base 모델의 결과를 검사할 수 없다는 것입니다. 즉, 여전히 노이즈가 심하게 제거됩니다.
base 모델과 refiner를 denoiser의 앙상블로 사용하기 위해 각각 고노이즈(high-nosise) (*즉* base 모델)와 저노이즈 (*즉* refiner 모델)의 노이즈를 제거하는 단계를 거쳐야하는 타임스텝의 기간을 정의해야 합니다.
base 모델의 [`denoising_end`](https://huggingface.co/docs/diffusers/main/en/api/pipelines/stable_diffusion/stable_diffusion_xl#diffusers.StableDiffusionXLPipeline.__call__.denoising_end)와 refiner 모델의 [`denoising_start`](https://huggingface.co/docs/diffusers/main/en/api/pipelines/stable_diffusion/stable_diffusion_xl#diffusers.StableDiffusionXLImg2ImgPipeline.__call__.denoising_start)를 사용해 간격을 정합니다.
`denoising_end`와 `denoising_start` 모두 0과 1사이의 실수 값으로 전달되어야 합니다.
전달되면 노이즈 제거의 끝과 시작은 모델 스케줄에 의해 정의된 이산적(discrete) 시간 간격의 비율로 정의됩니다.
노이즈 제거 단계의 수는 모델이 학습된 불연속적인 시간 간격과 선언된 fractional cutoff에 의해 결정되므로 '강도' 또한 선언된 경우 이 값이 '강도'를 재정의합니다.
예시를 들어보겠습니다.
우선, 두 개의 파이프라인을 가져옵니다. 텍스트 인코더와 variational autoencoder는 동일하므로 refiner를 위해 다시 불러오지 않아도 됩니다.
이제 추론 단계의 수와 고노이즈에서 노이즈를 제거하는 단계(*즉* base 모델)를 거쳐 실행되는 지점을 정의합니다.
```py
n_steps=40
high_noise_frac=0.8
```
Stable Diffusion XL base 모델은 타임스텝 0-999에 학습되며 Stable Diffusion XL refiner는 포괄적인 낮은 노이즈 타임스텝인 0-199에 base 모델로 부터 파인튜닝되어, 첫 800 타임스텝 (높은 노이즈)에 base 모델을 사용하고 마지막 200 타입스텝 (낮은 노이즈)에서 refiner가 사용됩니다. 따라서, `high_noise_frac`는 0.8로 설정하고, 모든 200-999 스텝(노이즈 제거 타임스텝의 첫 80%)은 base 모델에 의해 수행되며 0-199 스텝(노이즈 제거 타임스텝의 마지막 20%)은 refiner 모델에 의해 수행됩니다.
동일한 40 단계에서 base 모델을 실행한다면, 이미지의 디테일(예: 사자의 눈과 코)이 떨어졌을 것입니다:
<Tip>
앙상블 방식은 사용 가능한 모든 스케줄러에서 잘 작동합니다!
</Tip>
#### 2.) 노이즈가 완전히 제거된 기본 이미지에서 이미지 출력을 정제하기
일반적인 [`StableDiffusionImg2ImgPipeline`] 방식에서, 기본 모델에서 생성된 완전히 노이즈가 제거된 이미지는 [refiner checkpoint](huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0)를 사용해 더 향상시킬 수 있습니다.
이를 위해, 보통의 "base" text-to-image 파이프라인을 수행 후에 image-to-image 파이프라인으로써 refiner를 실행시킬 수 있습니다. base 모델의 출력을 잠재 공간에 남겨둘 수 있습니다.
Stable Diffusion XL는 두 개의 텍스트 인코더에 학습되었습니다. 기본 동작은 각 프롬프트에 동일한 프롬프트를 전달하는 것입니다. 그러나 [일부 사용자](https://github.com/huggingface/diffusers/issues/4004#issuecomment-1627764201)가 품질을 향상시킬 수 있다고 지적한 것처럼 텍스트 인코더마다 다른 프롬프트를 전달할 수 있습니다. 그렇게 하려면, `prompt_2`와 `negative_prompt_2`를 `prompt`와 `negative_prompt`에 전달해야 합니다. 그렇게 함으로써, 원래의 프롬프트들(`prompt`)과 부정 프롬프트들(`negative_prompt`)를 `텍스트 인코더`에 전달할 것입니다.(공식 SDXL 0.9/1.0의 [OpenAI CLIP-ViT/L-14](https://huggingface.co/openai/clip-vit-large-patch14)에서 볼 수 있습니다.) 그리고 `prompt_2`와 `negative_prompt_2`는 `text_encoder_2`에 전달됩니다.(공식 SDXL 0.9/1.0의 [OpenCLIP-ViT/bigG-14](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)에서 볼 수 있습니다.)
🤗 Diffusers는 이미지, 오디오, 심지어 분자의 3D 구조를 생성하기 위한 최첨단 사전 훈련된 diffusion 모델을 위한 라이브러리입니다. 간단한 추론 솔루션을 찾고 있든, 자체 diffusion 모델을 훈련하고 싶든, 🤗 Diffusers는 두 가지 모두를 지원하는 모듈식 툴박스입니다. 저희 라이브러리는 [성능보다 사용성](conceptual/philosophy#usability-over-performance), [간편함보다 단순함](conceptual/philosophy#simple-over-easy), 그리고 [추상화보다 사용자 지정 가능성](conceptual/philosophy#tweakable-contributorfriendly-over-abstraction)에 중점을 두고 설계되었습니다.
이 라이브러리에는 세 가지 주요 구성 요소가 있습니다:
- 몇 줄의 코드만으로 추론할 수 있는 최첨단 [diffusion 파이프라인](api/pipelines/overview).
- 생성 속도와 품질 간의 균형을 맞추기 위해 상호교환적으로 사용할 수 있는 [노이즈 스케줄러](api/schedulers/overview).
- 빌딩 블록으로 사용할 수 있고 스케줄러와 결합하여 자체적인 end-to-end diffusion 시스템을 만들 수 있는 사전 학습된 [모델](api/models).
<pclass="text-gray-700">파이프라인, 모델, 스케줄러를 로드하는 데 도움이 되는 실용적인 가이드입니다. 또한 특정 작업에 파이프라인을 사용하고, 출력 생성 방식을 제어하고, 추론 속도에 맞게 최적화하고, 다양한 학습 기법을 사용하는 방법도 배울 수 있습니다.</p>
<pclass="text-gray-700">🤗 Diffusers 클래스 및 메서드의 작동 방식에 대한 기술 설명.</p>
</a>
</div>
</div>
## Supported pipelines
| Pipeline | Paper/Repository | Tasks |
|---|---|:---:|
| [alt_diffusion](./api/pipelines/alt_diffusion) | [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) | Image-to-Image Text-Guided Generation |
| [stochastic_karras_ve](./api/pipelines/stochastic_karras_ve) | [Elucidating the Design Space of Diffusion-Based Generative Models](https://arxiv.org/abs/2206.00364) | Unconditional Image Generation |
| [text_to_video_sd](./api/pipelines/text_to_video) | [Modelscope's Text-to-video-synthesis Model in Open Domain](https://modelscope.cn/models/damo/text-to-video-synthesis/summary) | Text-to-Video Generation |
| [unclip](./api/pipelines/unclip) | [Hierarchical Text-Conditional Image Generation with CLIP Latents](https://arxiv.org/abs/2204.06125)(implementation by [kakaobrain](https://github.com/kakaobrain/karlo)) | Text-to-Image Generation |
| [versatile_diffusion](./api/pipelines/versatile_diffusion) | [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) | Text-to-Image Generation |
| [versatile_diffusion](./api/pipelines/versatile_diffusion) | [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) | Image Variations Generation |
| [versatile_diffusion](./api/pipelines/versatile_diffusion) | [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) | Dual Image and Text Guided Generation |
| [vq_diffusion](./api/pipelines/vq_diffusion) | [Vector Quantized Diffusion Model for Text-to-Image Synthesis](https://arxiv.org/abs/2111.14822) | Text-to-Image Generation |
<!--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.
-->
# Core ML로 Stable Diffusion을 실행하는 방법
[Core ML](https://developer.apple.com/documentation/coreml)은 Apple 프레임워크에서 지원하는 모델 형식 및 머신 러닝 라이브러리입니다. macOS 또는 iOS/iPadOS 앱 내에서 Stable Diffusion 모델을 실행하는 데 관심이 있는 경우, 이 가이드에서는 기존 PyTorch 체크포인트를 Core ML 형식으로 변환하고 이를 Python 또는 Swift로 추론에 사용하는 방법을 설명합니다.
Core ML 모델은 Apple 기기에서 사용할 수 있는 모든 컴퓨팅 엔진들, 즉 CPU, GPU, Apple Neural Engine(또는 Apple Silicon Mac 및 최신 iPhone/iPad에서 사용할 수 있는 텐서 최적화 가속기인 ANE)을 활용할 수 있습니다. 모델과 실행 중인 기기에 따라 Core ML은 컴퓨팅 엔진도 혼합하여 사용할 수 있으므로, 예를 들어 모델의 일부가 CPU에서 실행되는 반면 다른 부분은 GPU에서 실행될 수 있습니다.
<Tip>
PyTorch에 내장된 `mps` 가속기를 사용하여 Apple Silicon Macs에서 `diffusers` Python 코드베이스를 실행할 수도 있습니다. 이 방법은 [mps 가이드]에 자세히 설명되어 있지만 네이티브 앱과 호환되지 않습니다.
</Tip>
## Stable Diffusion Core ML 체크포인트
Stable Diffusion 가중치(또는 체크포인트)는 PyTorch 형식으로 저장되기 때문에 네이티브 앱에서 사용하기 위해서는 Core ML 형식으로 변환해야 합니다.
다행히도 Apple 엔지니어들이 `diffusers`를 기반으로 한 [변환 툴](https://github.com/apple/ml-stable-diffusion#-converting-models-to-core-ml)을 개발하여 PyTorch 체크포인트를 Core ML로 변환할 수 있습니다.
모델을 변환하기 전에 잠시 시간을 내어 Hugging Face Hub를 살펴보세요. 관심 있는 모델이 이미 Core ML 형식으로 제공되고 있을 가능성이 높습니다:
-[Apple](https://huggingface.co/apple) organization에는 Stable Diffusion 버전 1.4, 1.5, 2.0 base 및 2.1 base가 포함되어 있습니다.
-[coreml](https://huggingface.co/coreml) organization에는 커스텀 DreamBooth가 적용되거나, 파인튜닝된 모델이 포함되어 있습니다.
- 이 [필터](https://huggingface.co/models?pipeline_tag=text-to-image&library=coreml&p=2&sort=likes)를 사용하여 사용 가능한 모든 Core ML 체크포인트들을 반환합니다.
원하는 모델을 찾을 수 없는 경우 Apple의 [모델을 Core ML로 변환하기](https://github.com/apple/ml-stable-diffusion#-converting-models-to-core-ml) 지침을 따르는 것이 좋습니다.
## 사용할 Core ML 변형(Variant) 선택하기
Stable Diffusion 모델은 다양한 목적에 따라 다른 Core ML 변형으로 변환할 수 있습니다:
- 사용되는 어텐션 블록 유형. 어텐션 연산은 이미지 표현의 여러 영역 간의 관계에 '주의를 기울이고' 이미지와 텍스트 표현이 어떻게 연관되어 있는지 이해하는 데 사용됩니다. 어텐션 연산은 컴퓨팅 및 메모리 집약적이므로 다양한 장치의 하드웨어 특성을 고려한 다양한 구현이 존재합니다. Core ML Stable Diffusion 모델의 경우 두 가지 주의 변형이 있습니다:
*`split_einsum` ([Apple에서 도입](https://machinelearning.apple.com/research/neural-engine-transformers)은 최신 iPhone, iPad 및 M 시리즈 컴퓨터에서 사용할 수 있는 ANE 장치에 최적화되어 있습니다.
* "원본" 어텐션(`diffusers`에 사용되는 기본 구현)는 CPU/GPU와만 호환되며 ANE와는 호환되지 않습니다. "원본" 어텐션을 사용하여 CPU + GPU에서 모델을 실행하는 것이 ANE보다 *더* 빠를 수 있습니다. 자세한 내용은 [이 성능 벤치마크](https://huggingface.co/blog/fast-mac-diffusers#performance-benchmarks)와 커뮤니티에서 제공하는 일부 [추가 측정](https://github.com/huggingface/swift-coreml-diffusers/issues/31)을 참조하십시오.
- 지원되는 추론 프레임워크
*`packages`는 Python 추론에 적합합니다. 네이티브 앱에 통합하기 전에 변환된 Core ML 모델을 테스트하거나, Core ML 성능을 알고 싶지만 네이티브 앱을 지원할 필요는 없는 경우에 사용할 수 있습니다. 예를 들어, 웹 UI가 있는 애플리케이션은 Python Core ML 백엔드를 완벽하게 사용할 수 있습니다.
* Swift 코드에는 `컴파일된` 모델이 필요합니다. Hub의 `컴파일된` 모델은 iOS 및 iPadOS 기기와의 호환성을 위해 큰 UNet 모델 가중치를 여러 파일로 분할합니다. 이는 [`--chunk-unet` 변환 옵션](https://github.com/apple/ml-stable-diffusion#-converting-models-to-core-ml)에 해당합니다. 네이티브 앱을 지원하려면 `컴파일된` 변형을 선택해야 합니다.
공식 Core ML Stable Diffusion [모델](https://huggingface.co/apple/coreml-stable-diffusion-v1-4/tree/main)에는 이러한 변형이 포함되어 있지만 커뮤니티 버전은 다를 수 있습니다:
모델의 snapshot을 다운로드한 후에는 Apple의 Python 스크립트를 사용하여 테스트할 수 있습니다.
```shell
python -m python_coreml_stable_diffusion.pipeline --prompt"a photo of an astronaut riding a horse on mars"-i models/coreml-stable-diffusion-v1-4_original_packages -o </path/to/output/image> --compute-unit CPU_AND_GPU --seed 93
```
`<output-mlpackages-directory>`는 위 단계에서 다운로드한 체크포인트를 가리켜야 하며, `--compute-unit`은 추론을 허용할 하드웨어를 나타냅니다. 이는 다음 옵션 중 하나이어야 합니다: `ALL`, `CPU_AND_GPU`, `CPU_ONLY`, `CPU_AND_NE`. 선택적 출력 경로와 재현성을 위한 시드를 제공할 수도 있습니다.
추론 스크립트에서는 Stable Diffusion 모델의 원래 버전인 `CompVis/stable-diffusion-v1-4`를 사용한다고 가정합니다. 다른 모델을 사용하는 경우 추론 명령줄에서 `--model-version` 옵션을 사용하여 해당 허브 ID를 *지정*해야 합니다. 이는 이미 지원되는 모델과 사용자가 직접 학습하거나 파인튜닝한 사용자 지정 모델에 적용됩니다.
예를 들어, [`runwayml/stable-diffusion-v1-5`](https://huggingface.co/runwayml/stable-diffusion-v1-5)를 사용하려는 경우입니다:
```shell
python -m python_coreml_stable_diffusion.pipeline --prompt"a photo of an astronaut riding a horse on mars"--compute-unit ALL -o output --seed 93 -i models/coreml-stable-diffusion-v1-5_original_packages --model-version runwayml/stable-diffusion-v1-5
```
## Swift에서 Core ML 추론하기
Swift에서 추론을 실행하는 것은 모델이 이미 `mlmodelc` 형식으로 컴파일되어 있기 때문에 Python보다 약간 빠릅니다. 이는 앱이 시작될 때 모델이 불러와지는 것이 눈에 띄지만, 이후 여러 번 실행하면 눈에 띄지 않을 것입니다.
### 다운로드
Mac에서 Swift에서 추론을 실행하려면 `컴파일된` 체크포인트 버전 중 하나가 필요합니다. 이전 예제와 유사하지만 `컴파일된` 변형 중 하나를 사용하여 Python 코드를 로컬로 다운로드하는 것이 좋습니다:
그 다음 Apple의 명령어 도구인 [Swift 패키지 관리자](https://www.swift.org/package-manager/#)를 사용합니다:
```bash
swift run StableDiffusionSample --resource-path models/coreml-stable-diffusion-v1-4_original_compiled --compute-units all "a photo of an astronaut riding a horse on mars"
```
`--resource-path`에 이전 단계에서 다운로드한 체크포인트 중 하나를 지정해야 하므로 확장자가 `.mlmodelc`인 컴파일된 Core ML 번들이 포함되어 있는지 확인하시기 바랍니다. `--compute-units`는 다음 값 중 하나이어야 합니다: `all`, `cpuOnly`, `cpuAndGPU`, `cpuAndNeuralEngine`.
자세한 내용은 [Apple의 리포지토리 안의 지침](https://github.com/apple/ml-stable-diffusion)을 참고하시기 바랍니다.
## 지원되는 Diffusers 기능
Core ML 모델과 추론 코드는 🧨 Diffusers의 많은 기능, 옵션 및 유연성을 지원하지 않습니다. 다음은 유의해야 할 몇 가지 제한 사항입니다:
- Core ML 모델은 추론에만 적합합니다. 학습이나 파인튜닝에는 사용할 수 없습니다.
- Swift에 포팅된 스케줄러는 Stable Diffusion에서 사용하는 기본 스케줄러와 `diffusers` 구현에서 Swift로 포팅한 `DPMSolverMultistepScheduler` 두 개뿐입니다. 이들 중 약 절반의 스텝으로 동일한 품질을 생성하는 `DPMSolverMultistepScheduler`를 사용하는 것이 좋습니다.
- 추론 코드에서 네거티브 프롬프트, classifier-free guidance scale 및 image-to-image 작업을 사용할 수 있습니다. depth guidance, ControlNet, latent upscalers와 같은 고급 기능은 아직 사용할 수 없습니다.
Apple의 [변환 및 추론 리포지토리](https://github.com/apple/ml-stable-diffusion)와 자체 [swift-coreml-diffusers](https://github.com/huggingface/swift-coreml-diffusers) 리포지토리는 다른 개발자들이 구축할 수 있는 기술적인 데모입니다.
누락된 기능이 있다고 생각되면 언제든지 기능을 요청하거나, 더 좋은 방법은 기여 PR을 열어주세요. :)
## 네이티브 Diffusers Swift 앱
자체 Apple 하드웨어에서 Stable Diffusion을 실행하는 쉬운 방법 중 하나는 `diffusers`와 Apple의 변환 및 추론 리포지토리를 기반으로 하는 [자체 오픈 소스 Swift 리포지토리](https://github.com/huggingface/swift-coreml-diffusers)를 사용하는 것입니다. 코드를 공부하고 [Xcode](https://developer.apple.com/xcode/)로 컴파일하여 필요에 맞게 조정할 수 있습니다. 편의를 위해 앱스토어에 [독립형 Mac 앱](https://apps.apple.com/app/diffusers/id1666309574)도 있으므로 코드나 IDE를 다루지 않고도 사용할 수 있습니다. 개발자로서 Core ML이 Stable Diffusion 앱을 구축하는 데 가장 적합한 솔루션이라고 판단했다면, 이 가이드의 나머지 부분을 사용하여 프로젝트를 시작할 수 있습니다. 여러분이 무엇을 빌드할지 기대됩니다. :)
<!--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.
-->
# 메모리와 속도
메모리 또는 속도에 대해 🤗 Diffusers *추론*을 최적화하기 위한 몇 가지 기술과 아이디어를 제시합니다.
일반적으로, memory-efficient attention을 위해 [xFormers](https://github.com/facebookresearch/xformers) 사용을 추천하기 때문에, 추천하는 [설치 방법](xformers)을 보고 설치해 보세요.
다음 설정이 성능과 메모리에 미치는 영향에 대해 설명합니다.
| | 지연시간 | 속도 향상 |
| ---------------- | ------- | ------- |
| 별도 설정 없음 | 9.50s | x1 |
| cuDNN auto-tuner | 9.37s | x1.01 |
| fp16 | 3.61s | x2.63 |
| Channels Last 메모리 형식 | 3.30s | x2.88 |
| traced UNet | 3.21s | x2.96 |
| memory-efficient attention | 2.63s | x3.61 |
<em>
NVIDIA TITAN RTX에서 50 DDIM 스텝의 "a photo of an astronaut riding a horse on mars" 프롬프트로 512x512 크기의 단일 이미지를 생성하였습니다.
</em>
## cuDNN auto-tuner 활성화하기
[NVIDIA cuDNN](https://developer.nvidia.com/cudnn)은 컨볼루션을 계산하는 많은 알고리즘을 지원합니다. Autotuner는 짧은 벤치마크를 실행하고 주어진 입력 크기에 대해 주어진 하드웨어에서 최고의 성능을 가진 커널을 선택합니다.
**컨볼루션 네트워크**를 활용하고 있기 때문에 (다른 유형들은 현재 지원되지 않음), 다음 설정을 통해 추론 전에 cuDNN autotuner를 활성화할 수 있습니다:
```python
importtorch
torch.backends.cudnn.benchmark=True
```
### fp32 대신 tf32 사용하기 (Ampere 및 이후 CUDA 장치들에서)
Ampere 및 이후 CUDA 장치에서 행렬곱 및 컨볼루션은 TensorFloat32(TF32) 모드를 사용하여 더 빠르지만 약간 덜 정확할 수 있습니다.
기본적으로 PyTorch는 컨볼루션에 대해 TF32 모드를 활성화하지만 행렬 곱셈은 활성화하지 않습니다.
네트워크에 완전한 float32 정밀도가 필요한 경우가 아니면 행렬 곱셈에 대해서도 이 설정을 활성화하는 것이 좋습니다.
이는 일반적으로 무시할 수 있는 수치의 정확도 손실이 있지만, 계산 속도를 크게 높일 수 있습니다.
그것에 대해 [여기](https://huggingface.co/docs/transformers/v4.18.0/en/performance#tf32)서 더 읽을 수 있습니다.
추론하기 전에 다음을 추가하기만 하면 됩니다:
```python
importtorch
torch.backends.cuda.matmul.allow_tf32=True
```
## 반정밀도 가중치
더 많은 GPU 메모리를 절약하고 더 빠른 속도를 얻기 위해 모델 가중치를 반정밀도(half precision)로 직접 불러오고 실행할 수 있습니다.
여기에는 `fp16`이라는 브랜치에 저장된 float16 버전의 가중치를 불러오고, 그 때 `float16` 유형을 사용하도록 PyTorch에 지시하는 작업이 포함됩니다.
```Python
pipe = StableDiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
torch_dtype=torch.float16,
)
pipe = pipe.to("cuda")
prompt = "a photo of an astronaut riding a horse on mars"
image = pipe(prompt).images[0]
```
<Tipwarning={true}>
어떤 파이프라인에서도 [`torch.autocast`](https://pytorch.org/docs/stable/amp.html#torch.autocast) 를 사용하는 것은 검은색 이미지를 생성할 수 있고, 순수한 float16 정밀도를 사용하는 것보다 항상 느리기 때문에 사용하지 않는 것이 좋습니다.
</Tip>
## 추가 메모리 절약을 위한 슬라이스 어텐션
추가 메모리 절약을 위해, 한 번에 모두 계산하는 대신 단계적으로 계산을 수행하는 슬라이스 버전의 어텐션(attention)을 사용할 수 있습니다.
<Tip>
Attention slicing은 모델이 하나 이상의 어텐션 헤드를 사용하는 한, 배치 크기가 1인 경우에도 유용합니다.
하나 이상의 어텐션 헤드가 있는 경우 *QK^T* 어텐션 매트릭스는 상당한 양의 메모리를 절약할 수 있는 각 헤드에 대해 순차적으로 계산될 수 있습니다.
</Tip>
각 헤드에 대해 순차적으로 어텐션 계산을 수행하려면, 다음과 같이 추론 전에 파이프라인에서 [`~StableDiffusionPipeline.enable_attention_slicing`]를 호출하면 됩니다:
```Python
import torch
from diffusers import StableDiffusionPipeline
pipe = StableDiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
torch_dtype=torch.float16,
)
pipe = pipe.to("cuda")
prompt = "a photo of an astronaut riding a horse on mars"
pipe.enable_attention_slicing()
image = pipe(prompt).images[0]
```
추론 시간이 약 10% 느려지는 약간의 성능 저하가 있지만 이 방법을 사용하면 3.2GB 정도의 작은 VRAM으로도 Stable Diffusion을 사용할 수 있습니다!
## 더 큰 배치를 위한 sliced VAE 디코드
제한된 VRAM에서 대규모 이미지 배치를 디코딩하거나 32개 이상의 이미지가 포함된 배치를 활성화하기 위해, 배치의 latent 이미지를 한 번에 하나씩 디코딩하는 슬라이스 VAE 디코드를 사용할 수 있습니다.
이를 [`~StableDiffusionPipeline.enable_attention_slicing`] 또는 [`~StableDiffusionPipeline.enable_xformers_memory_efficient_attention`]과 결합하여 메모리 사용을 추가로 최소화할 수 있습니다.
VAE 디코드를 한 번에 하나씩 수행하려면 추론 전에 파이프라인에서 [`~StableDiffusionPipeline.enable_vae_slicing`]을 호출합니다. 예를 들어:
```Python
import torch
from diffusers import StableDiffusionPipeline
pipe = StableDiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
torch_dtype=torch.float16,
)
pipe = pipe.to("cuda")
prompt = "a photo of an astronaut riding a horse on mars"
pipe.enable_vae_slicing()
images = pipe([prompt] * 32).images
```
다중 이미지 배치에서 VAE 디코드가 약간의 성능 향상이 이루어집니다. 단일 이미지 배치에서는 성능 영향은 없습니다.
<aname="sequential_offloading"></a>
## 메모리 절약을 위해 가속 기능을 사용하여 CPU로 오프로딩
추가 메모리 절약을 위해 가중치를 CPU로 오프로드하고 순방향 전달을 수행할 때만 GPU로 로드할 수 있습니다.
CPU 오프로딩을 수행하려면 [`~StableDiffusionPipeline.enable_sequential_cpu_offload`]를 호출하기만 하면 됩니다:
```Python
import torch
from diffusers import StableDiffusionPipeline
pipe = StableDiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
torch_dtype=torch.float16,
)
prompt = "a photo of an astronaut riding a horse on mars"
pipe.enable_sequential_cpu_offload()
image = pipe(prompt).images[0]
```
그러면 메모리 소비를 3GB 미만으로 줄일 수 있습니다.
참고로 이 방법은 전체 모델이 아닌 서브모듈 수준에서 작동합니다. 이는 메모리 소비를 최소화하는 가장 좋은 방법이지만 프로세스의 반복적 특성으로 인해 추론 속도가 훨씬 느립니다. 파이프라인의 UNet 구성 요소는 여러 번 실행됩니다('num_inference_steps' 만큼). 매번 UNet의 서로 다른 서브모듈이 순차적으로 온로드된 다음 필요에 따라 오프로드되므로 메모리 이동 횟수가 많습니다.
<Tip>
또 다른 최적화 방법인 <ahref="#model_offloading">모델 오프로딩</a>을 사용하는 것을 고려하십시오. 이는 훨씬 빠르지만 메모리 절약이 크지는 않습니다.
</Tip>
또한 ttention slicing과 연결해서 최소 메모리(< 2GB)로도 동작할 수 있습니다.
```Python
import torch
from diffusers import StableDiffusionPipeline
pipe = StableDiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
torch_dtype=torch.float16,
)
prompt = "a photo of an astronaut riding a horse on mars"
pipe.enable_sequential_cpu_offload()
pipe.enable_attention_slicing(1)
image = pipe(prompt).images[0]
```
**참고**: 'enable_sequential_cpu_offload()'를 사용할 때, 미리 파이프라인을 CUDA로 이동하지 **않는** 것이 중요합니다.그렇지 않으면 메모리 소비의 이득이 최소화됩니다. 더 많은 정보를 위해 [이 이슈](https://github.com/huggingface/diffusers/issues/1934)를 보세요.
<aname="model_offloading"></a>
## 빠른 추론과 메모리 메모리 절약을 위한 모델 오프로딩
[순차적 CPU 오프로딩](#sequential_offloading)은 이전 섹션에서 설명한 것처럼 많은 메모리를 보존하지만 필요에 따라 서브모듈을 GPU로 이동하고 새 모듈이 실행될 때 즉시 CPU로 반환되기 때문에 추론 속도가 느려집니다.
전체 모델 오프로딩은 각 모델의 구성 요소인 _modules_을 처리하는 대신, 전체 모델을 GPU로 이동하는 대안입니다. 이로 인해 추론 시간에 미치는 영향은 미미하지만(파이프라인을 'cuda'로 이동하는 것과 비교하여) 여전히 약간의 메모리를 절약할 수 있습니다.
이 시나리오에서는 파이프라인의 주요 구성 요소 중 하나만(일반적으로 텍스트 인코더, unet 및 vae) GPU에 있고, 나머지는 CPU에서 대기할 것입니다.
여러 반복을 위해 실행되는 UNet과 같은 구성 요소는 더 이상 필요하지 않을 때까지 GPU에 남아 있습니다.
이 기능은 아래와 같이 파이프라인에서 `enable_model_cpu_offload()`를 호출하여 활성화할 수 있습니다.
```Python
import torch
from diffusers import StableDiffusionPipeline
pipe = StableDiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
torch_dtype=torch.float16,
)
prompt = "a photo of an astronaut riding a horse on mars"
pipe.enable_model_cpu_offload()
image = pipe(prompt).images[0]
```
이는 추가적인 메모리 절약을 위한 attention slicing과도 호환됩니다.
```Python
import torch
from diffusers import StableDiffusionPipeline
pipe = StableDiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
torch_dtype=torch.float16,
)
prompt = "a photo of an astronaut riding a horse on mars"
pipe.enable_model_cpu_offload()
pipe.enable_attention_slicing(1)
image = pipe(prompt).images[0]
```
<Tip>
이 기능을 사용하려면 'accelerate' 버전 0.17.0 이상이 필요합니다.
</Tip>
## Channels Last 메모리 형식 사용하기
Channels Last 메모리 형식은 차원 순서를 보존하는 메모리에서 NCHW 텐서 배열을 대체하는 방법입니다.
Channels Last 텐서는 채널이 가장 조밀한 차원이 되는 방식으로 정렬됩니다(일명 픽셀당 이미지를 저장).
현재 모든 연산자 Channels Last 형식을 지원하는 것은 아니라 성능이 저하될 수 있으므로, 사용해보고 모델에 잘 작동하는지 확인하는 것이 좋습니다.
예를 들어 파이프라인의 UNet 모델이 channels Last 형식을 사용하도록 설정하려면 다음을 사용할 수 있습니다:
<!--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.
-->
# Habana Gaudi에서 Stable Diffusion을 사용하는 방법
🤗 Diffusers는 🤗 [Optimum Habana](https://huggingface.co/docs/optimum/habana/usage_guides/stable_diffusion)를 통해서 Habana Gaudi와 호환됩니다.
## 요구 사항
- Optimum Habana 1.4 또는 이후, [여기](https://huggingface.co/docs/optimum/habana/installation)에 설치하는 방법이 있습니다.
- SynapseAI 1.8.
## 추론 파이프라인
Gaudi에서 Stable Diffusion 1 및 2로 이미지를 생성하려면 두 인스턴스를 인스턴스화해야 합니다:
-[`GaudiStableDiffusionPipeline`](https://huggingface.co/docs/optimum/habana/package_reference/stable_diffusion_pipeline)이 포함된 파이프라인. 이 파이프라인은 *텍스트-이미지 생성*을 지원합니다.
-[`GaudiDDIMScheduler`](https://huggingface.co/docs/optimum/habana/package_reference/stable_diffusion_pipeline#optimum.habana.diffusers.GaudiDDIMScheduler)이 포함된 스케줄러. 이 스케줄러는 Habana Gaudi에 최적화되어 있습니다.
파이프라인을 초기화할 때, HPU에 배포하기 위해 `use_habana=True`를 지정해야 합니다.
또한 가능한 가장 빠른 생성을 위해 `use_hpu_graphs=True`로 **HPU 그래프**를 활성화해야 합니다.
마지막으로, [Hugging Face Hub](https://huggingface.co/Habana)에서 다운로드할 수 있는 [Gaudi configuration](https://huggingface.co/docs/optimum/habana/package_reference/gaudi_config)을 지정해야 합니다.
"High quality photo of an astronaut riding a horse in space",
"Face of a yellow cat, high resolution, sitting on a park bench",
],
num_images_per_prompt=10,
batch_size=4,
)
```
더 많은 정보를 얻기 위해, Optimum Habana의 [문서](https://huggingface.co/docs/optimum/habana/usage_guides/stable_diffusion)와 공식 Github 저장소에 제공된 [예시](https://github.com/huggingface/optimum-habana/tree/main/examples/stable-diffusion)를 확인하세요.
## 벤치마크
다음은 [Habana/stable-diffusion](https://huggingface.co/Habana/stable-diffusion) Gaudi 구성(혼합 정밀도 bf16/fp32)을 사용하는 Habana first-generation Gaudi 및 Gaudi2의 지연 시간입니다:
| | Latency (배치 크기 = 1) | Throughput (배치 크기 = 8) |