README.md
| 1 | --- |
| 2 | pipeline_tag: text-to-image |
| 3 | inference: false |
| 4 | license: other |
| 5 | license_name: sai-nc-community |
| 6 | license_link: https://huggingface.co/stabilityai/sdxl-turbo/blob/main/LICENSE.md |
| 7 | --- |
| 8 | |
| 9 | # SDXL-Turbo Model Card |
| 10 | |
| 11 | <!-- Provide a quick summary of what the model is/does. --> |
| 12 |  |
| 13 | SDXL-Turbo is a fast generative text-to-image model that can synthesize photorealistic images from a text prompt in a single network evaluation. |
| 14 | A real-time demo is available here: http://clipdrop.co/stable-diffusion-turbo |
| 15 | |
| 16 | Please note: For commercial use, please refer to https://stability.ai/license. |
| 17 | |
| 18 | ## Model Details |
| 19 | |
| 20 | ### Model Description |
| 21 | SDXL-Turbo is a distilled version of [SDXL 1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0), trained for real-time synthesis. |
| 22 | SDXL-Turbo is based on a novel training method called Adversarial Diffusion Distillation (ADD) (see the [technical report](https://stability.ai/research/adversarial-diffusion-distillation)), which allows sampling large-scale foundational |
| 23 | image diffusion models in 1 to 4 steps at high image quality. |
| 24 | This approach uses score distillation to leverage large-scale off-the-shelf image diffusion models as a teacher signal and combines this with an |
| 25 | adversarial loss to ensure high image fidelity even in the low-step regime of one or two sampling steps. |
| 26 | |
| 27 | - **Developed by:** Stability AI |
| 28 | - **Funded by:** Stability AI |
| 29 | - **Model type:** Generative text-to-image model |
| 30 | - **Finetuned from model:** [SDXL 1.0 Base](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) |
| 31 | |
| 32 | ### Model Sources |
| 33 | |
| 34 | For research purposes, we recommend our `generative-models` Github repository (https://github.com/Stability-AI/generative-models), |
| 35 | which implements the most popular diffusion frameworks (both training and inference). |
| 36 | |
| 37 | - **Repository:** https://github.com/Stability-AI/generative-models |
| 38 | - **Paper:** https://stability.ai/research/adversarial-diffusion-distillation |
| 39 | - **Demo:** http://clipdrop.co/stable-diffusion-turbo |
| 40 | |
| 41 | |
| 42 | ## Evaluation |
| 43 |  |
| 44 |  |
| 45 | The charts above evaluate user preference for SDXL-Turbo over other single- and multi-step models. |
| 46 | SDXL-Turbo evaluated at a single step is preferred by human voters in terms of image quality and prompt following over LCM-XL evaluated at four (or fewer) steps. |
| 47 | In addition, we see that using four steps for SDXL-Turbo further improves performance. |
| 48 | For details on the user study, we refer to the [research paper](https://stability.ai/research/adversarial-diffusion-distillation). |
| 49 | |
| 50 | |
| 51 | ## Uses |
| 52 | |
| 53 | ### Direct Use |
| 54 | |
| 55 | The model is intended for both non-commercial and commercial usage. You can use this model for non-commercial or research purposes under this [license](https://huggingface.co/stabilityai/sdxl-turbo/blob/main/LICENSE.md). Possible research areas and tasks include |
| 56 | |
| 57 | - Research on generative models. |
| 58 | - Research on real-time applications of generative models. |
| 59 | - Research on the impact of real-time generative models. |
| 60 | - Safe deployment of models which have the potential to generate harmful content. |
| 61 | - Probing and understanding the limitations and biases of generative models. |
| 62 | - Generation of artworks and use in design and other artistic processes. |
| 63 | - Applications in educational or creative tools. |
| 64 | |
| 65 | For commercial use, please refer to https://stability.ai/membership. |
| 66 | |
| 67 | Excluded uses are described below. |
| 68 | |
| 69 | ### Diffusers |
| 70 | |
| 71 | ``` |
| 72 | pip install diffusers transformers accelerate --upgrade |
| 73 | ``` |
| 74 | |
| 75 | - **Text-to-image**: |
| 76 | |
| 77 | SDXL-Turbo does not make use of `guidance_scale` or `negative_prompt`, we disable it with `guidance_scale=0.0`. |
| 78 | Preferably, the model generates images of size 512x512 but higher image sizes work as well. |
| 79 | A **single step** is enough to generate high quality images. |
| 80 | |
| 81 | ```py |
| 82 | from diffusers import AutoPipelineForText2Image |
| 83 | import torch |
| 84 | |
| 85 | pipe = AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16") |
| 86 | pipe.to("cuda") |
| 87 | |
| 88 | prompt = "A cinematic shot of a baby racoon wearing an intricate italian priest robe." |
| 89 | |
| 90 | image = pipe(prompt=prompt, num_inference_steps=1, guidance_scale=0.0).images[0] |
| 91 | ``` |
| 92 | |
| 93 | - **Image-to-image**: |
| 94 | |
| 95 | When using SDXL-Turbo for image-to-image generation, make sure that `num_inference_steps` * `strength` is larger or equal |
| 96 | to 1. The image-to-image pipeline will run for `int(num_inference_steps * strength)` steps, *e.g.* 0.5 * 2.0 = 1 step in our example |
| 97 | below. |
| 98 | |
| 99 | ```py |
| 100 | from diffusers import AutoPipelineForImage2Image |
| 101 | from diffusers.utils import load_image |
| 102 | import torch |
| 103 | |
| 104 | pipe = AutoPipelineForImage2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16") |
| 105 | pipe.to("cuda") |
| 106 | |
| 107 | init_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cat.png").resize((512, 512)) |
| 108 | |
| 109 | prompt = "cat wizard, gandalf, lord of the rings, detailed, fantasy, cute, adorable, Pixar, Disney, 8k" |
| 110 | |
| 111 | image = pipe(prompt, image=init_image, num_inference_steps=2, strength=0.5, guidance_scale=0.0).images[0] |
| 112 | ``` |
| 113 | |
| 114 | ### Out-of-Scope Use |
| 115 | |
| 116 | The model was not trained to be factual or true representations of people or events, |
| 117 | and therefore using the model to generate such content is out-of-scope for the abilities of this model. |
| 118 | The model should not be used in any way that violates Stability AI's [Acceptable Use Policy](https://stability.ai/use-policy). |
| 119 | |
| 120 | ## Limitations and Bias |
| 121 | |
| 122 | ### Limitations |
| 123 | - The generated images are of a fixed resolution (512x512 pix), and the model does not achieve perfect photorealism. |
| 124 | - The model cannot render legible text. |
| 125 | - Faces and people in general may not be generated properly. |
| 126 | - The autoencoding part of the model is lossy. |
| 127 | |
| 128 | |
| 129 | ### Recommendations |
| 130 | |
| 131 | The model is intended for both non-commercial and commercial usage. |
| 132 | |
| 133 | ## How to Get Started with the Model |
| 134 | |
| 135 | Check out https://github.com/Stability-AI/generative-models |