Pipelines provide a simple way to run state-of-the-art diffusion models in inference. Most diffusion systems consist of multiple independently-trained models and highly adaptable scheduler components - all of which are needed to have a functioning end-to-end diffusion system.
As an example, Stable Diffusion has three independently trained models:
- Conditional Unet
- CLIP text encoder
- a scheduler component, scheduler,
- a CLIPFeatureExtractor,
- as well as a safety checker. All of these components are necessary to run stable diffusion in inference even though they were trained or created independently from each other.
To that end, we strive to offer all open-sourced, state-of-the-art diffusion system under a unified API. More specifically, we strive to provide pipelines that
- can load the officially published weights and yield 1-to-1 the same outputs as the original implementation according to the corresponding paper (e.g. LatentDiffusionPipeline, uses the officially released weights of High-Resolution Image Synthesis with Latent Diffusion Models),
- have a simple user interface to run the model in inference (see the Pipelines API section),
- are easy to understand with code that is self-explanatory and can be read along-side the official paper (see Pipelines summary),
- can easily be contributed by the community (see the Contribution section).
Note that pipelines do not (and should not) offer any training functionality. If you are looking for official training examples, please have a look at examples.
The following table summarizes all officially supported pipelines, their corresponding paper, and if available a colab notebook to directly try them out.
Note: Pipelines are simple examples of how to play around with the diffusion systems as described in the corresponding papers.
However, most of them can be adapted to use different scheduler components or even different model components. Some pipeline examples are shown in the Examples below.
Diffusion models often consist of multiple independently-trained models or other previously existing components.
Each model has been trained independently on a different task and the scheduler can easily be swapped out and replaced with a different one. During inference, we however want to be able to easily load all components and use them in inference - even if one component, e.g. CLIP’s text encoder, originates from a different library, such as Transformers. To that end, all pipelines provide the following functionality:
from_pretrainedmethod that accepts a Hugging Face Hub repository id, e.g. CompVis/stable-diffusion-v1-4 or a path to a local directory, e.g. ”./stable-diffusion”. To correctly retrieve which models and components should be loaded, one has to provide a
model_index.jsonfile, e.g. CompVis/stable-diffusion-v1-4/model_index.json, which defines all components that should be loaded into the pipelines. More specifically, for each model/component one needs to define the format
<name>: ["<library>", "<class name>"].
<name>is the attribute name given to the loaded instance of
<class name>which can be found in the library or pipeline folder called
save_pretrainedthat accepts a local path, e.g.
./stable-diffusionunder which all models/components of the pipeline will be saved. For each component/model a folder is created inside the local path that is named after the given attribute name, e.g.
./stable_diffusion/unet. In addition, a
model_index.jsonfile is created at the root of the local path, e.g.
./stable_diffusion/model_index.jsonso that the complete pipeline can again be instantiated from the local path.
towhich accepts a
torch.deviceto move all models that are of type
torch.nn.Moduleto the passed device. The behavior is fully analogous to PyTorch’s
__call__method to use the pipeline in inference.
__call__defines inference logic of the pipeline and should ideally encompass all aspects of it, from pre-processing to forwarding tensors to the different models and schedulers, as well as post-processing. The API of the
__call__method can strongly vary from pipeline to pipeline. E.g. a text-to-image pipeline, such as
StableDiffusionPipelineshould accept among other things the text prompt to generate the image. A pure image generation pipeline, such as DDPMPipeline on the other hand can be run without providing any inputs. To better understand what inputs can be adapted for each pipeline, one should look directly into the respective pipeline.
Note: All pipelines have PyTorch’s autograd disabled by decorating the
__call__ method with a
torch.no_grad decorator because pipelines should
not be used for training. If you want to store the gradients during the forward pass, we recommend writing your own pipeline, see also our community-examples
We are more than happy about any contribution to the officially supported pipelines 🤗. We aspire all of our pipelines to be self-contained, easy-to-tweak, beginner-friendly and for one-purpose-only.
- Self-contained: A pipeline shall be as self-contained as possible. More specifically, this means that all functionality should be either directly defined in the pipeline file iteslf, should be inherited from (and only from) the
DiffusionPipelineclass or be directly attached to the model and scheduler components of the pipeline.
- Easy-to-use: Pipelines should be extremely easy to use - one should be able to load the pipeline and
use it for its designated task, e.g. text-to-image generation, in just a couple of lines of code. Most
logic including pre-processing, an unrolled diffusion loop, and post-processing should all happen inside the
- Easy-to-tweak: Certain pipelines will not be able to handle all use cases and tasks that you might like them to. If you want to use a certain pipeline for a specific use case that is not yet supported, you might have to copy the pipeline file and tweak the code to your needs. We try to make the pipeline code as readable as possible so that each part –from pre-processing to diffusing to post-processing– can easily be adapted. If you would like the community to benefit from your customized pipeline, we would love to see a contribution to our community-examples. If you feel that an important pipeline should be part of the official pipelines but isn’t, a contribution to the official pipelines would be even better.
- One-purpose-only: Pipelines should be used for one task and one task only. Even if two tasks are very similar from a modeling point of view, e.g. image2image translation and in-painting, pipelines shall be used for one task only to keep them easy-to-tweak and readable.
# make sure you're logged in with `huggingface-cli login` from torch import autocast from diffusers import StableDiffusionPipeline, LMSDiscreteScheduler pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", use_auth_token=True) pipe = pipe.to("cuda") prompt = "a photo of an astronaut riding a horse on mars" with autocast("cuda"): image = pipe(prompt).images image.save("astronaut_rides_horse.png")
StableDiffusionImg2ImgPipeline lets you pass a text prompt and an initial image to condition the generation of new images.
from torch import autocast import requests from PIL import Image from io import BytesIO from diffusers import StableDiffusionImg2ImgPipeline # load the pipeline device = "cuda" pipe = StableDiffusionImg2ImgPipeline.from_pretrained( "CompVis/stable-diffusion-v1-4", revision="fp16", torch_dtype=torch.float16, use_auth_token=True ).to(device) # let's download an initial image url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" response = requests.get(url) init_image = Image.open(BytesIO(response.content)).convert("RGB") init_image = init_image.resize((768, 512)) prompt = "A fantasy landscape, trending on artstation" with autocast("cuda"): images = pipe(prompt=prompt, init_image=init_image, strength=0.75, guidance_scale=7.5).images images.save("fantasy_landscape.png")
You can generate your own latents to reproduce results, or tweak your prompt on a specific result you liked. This notebook shows how to do it step by step. You can also run it in Google Colab .
StableDiffusionInpaintPipeline lets you edit specific parts of an image by providing a mask and text prompt.
from io import BytesIO from torch import autocast import requests import PIL from diffusers import StableDiffusionInpaintPipeline def download_image(url): response = requests.get(url) return PIL.Image.open(BytesIO(response.content)).convert("RGB") img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" init_image = download_image(img_url).resize((512, 512)) mask_image = download_image(mask_url).resize((512, 512)) device = "cuda" pipe = StableDiffusionInpaintPipeline.from_pretrained( "CompVis/stable-diffusion-v1-4", revision="fp16", torch_dtype=torch.float16, use_auth_token=True ).to(device) prompt = "a cat sitting on a bench" with autocast("cuda"): images = pipe(prompt=prompt, init_image=init_image, mask_image=mask_image, strength=0.75).images images.save("cat_on_bench.png")