SDXL consists of an ensemble of experts pipeline for latent diffusion:
In a first step, the base model is used to generate (noisy) latents,
which are then further processed with a refinement model (available here: https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/) specialized for the final denoising steps.
Note that the base model can be used as a standalone module.
Alternatively, we can use a two-stage pipeline as follows:
First, the base model is used to generate latents of the desired output size.
In the second step, we use a specialized high-resolution model and apply a technique called SDEdit (https://arxiv.org/abs/2108.01073, also known as "img2img")
to the latents generated in the first step, using the same prompt. This technique is slightly slower than the first one, as it requires more function evaluations.
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 that uses two fixed, pretrained text encoders (OpenCLIP-ViT/G and CLIP-ViT/L).
For research purposes, we recommend our generative-models Github repository (https://github.com/Stability-AI/generative-models), which implements the most popular diffusion frameworks (both training and inference) and for which new functionalities like distillation will be added over time.
Clipdrop provides free SDXL inference.
The chart above evaluates user preference for SDXL (with and without refinement) over SDXL 0.9 and Stable Diffusion 1.5 and 2.1.
The SDXL base model performs significantly better than the previous variants, and the model combined with the refinement module achieves the best overall performance.
🧨 Diffusers
Make sure to upgrade diffusers to >= 0.19.0:
pip install diffusers --upgrade
In addition make sure to install transformers, safetensors, accelerate as well as the invisible watermark:
1from diffusers import DiffusionPipeline
2import torch
34pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16, use_safetensors=True, variant="fp16")5pipe.to("cuda")67# if using torch < 2.08# pipe.enable_xformers_memory_efficient_attention()910prompt ="An astronaut riding a green horse"1112images = pipe(prompt=prompt).images[0]
To use the whole base + refiner pipeline as an ensemble of experts you can run:
py
1from diffusers import DiffusionPipeline
2import torch
34# load both base & refiner5base = DiffusionPipeline.from_pretrained(6"stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16, variant="fp16", use_safetensors=True7)8base.to("cuda")9refiner = DiffusionPipeline.from_pretrained(10"stabilityai/stable-diffusion-xl-refiner-1.0",11 text_encoder_2=base.text_encoder_2,12 vae=base.vae,13 torch_dtype=torch.float16,14 use_safetensors=True,15 variant="fp16",16)17refiner.to("cuda")1819# Define how many steps and what % of steps to be run on each experts (80/20) here20n_steps =4021high_noise_frac =0.82223prompt ="A majestic lion jumping from a big stone at night"2425# run both experts26image = base(27 prompt=prompt,28 num_inference_steps=n_steps,29 denoising_end=high_noise_frac,30 output_type="latent",31).images
32image = refiner(33 prompt=prompt,34 num_inference_steps=n_steps,35 denoising_start=high_noise_frac,36 image=image,37).images[0]
When using torch >= 2.0, you can improve the inference speed by 20-30% with torch.compile. Simple wrap the unet with torch compile before running the pipeline:
To install Optimum with the dependencies required for OpenVINO :
pip install optimum[openvino]
To load an OpenVINO model and run inference with OpenVINO Runtime, you need to replace StableDiffusionXLPipeline with Optimum OVStableDiffusionXLPipeline. In case you want to load a PyTorch model and convert it to the OpenVINO format on-the-fly, you can set export=True.
diff
1- from diffusers import StableDiffusionPipeline
2+ from optimum.intel import OVStableDiffusionPipeline
34model_id = "stabilityai/stable-diffusion-xl-base-1.0"
5- pipeline = StableDiffusionPipeline.from_pretrained(model_id)
6+ pipeline = OVStableDiffusionPipeline.from_pretrained(model_id)
7prompt = "A majestic lion jumping from a big stone at night"
8image = pipeline(prompt).images[0]
You can find more examples (such as static reshaping and model compilation) in optimum documentation.
ONNX
To install Optimum with the dependencies required for ONNX Runtime inference :
pip install optimum[onnxruntime]
To load an ONNX model and run inference with ONNX Runtime, you need to replace StableDiffusionXLPipeline with Optimum ORTStableDiffusionXLPipeline. In case you want to load a PyTorch model and convert it to the ONNX format on-the-fly, you can set export=True.
diff
1- from diffusers import StableDiffusionPipeline
2+ from optimum.onnxruntime import ORTStableDiffusionPipeline
34model_id = "stabilityai/stable-diffusion-xl-base-1.0"
5- pipeline = StableDiffusionPipeline.from_pretrained(model_id)
6+ pipeline = ORTStableDiffusionPipeline.from_pretrained(model_id)
7prompt = "A majestic lion jumping from a big stone at night"
8image = pipeline(prompt).images[0]
You can find more examples in optimum documentation.
Uses
Direct Use
The model is intended for research purposes only. Possible research areas and tasks include
Generation of artworks and use in design and other artistic processes.
Applications in educational or creative tools.
Research on generative models.
Safe deployment of models which have the potential to generate harmful content.
Probing and understanding the limitations and biases of generative models.
Excluded uses are described below.
Out-of-Scope Use
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.
Limitations and Bias
Limitations
The model does not achieve perfect photorealism
The model cannot render legible text
The model struggles with more difficult tasks which involve compositionality, such as rendering an image corresponding to “A red cube on top of a blue sphere”
Faces and people in general may not be generated properly.
The autoencoding part of the model is lossy.
Bias
While the capabilities of image generation models are impressive, they can also reinforce or exacerbate social biases.
