Delete README.md
Browse files
README.md
DELETED
@@ -1,123 +0,0 @@
|
|
1 |
-
---
|
2 |
-
license: openrail++
|
3 |
-
tags:
|
4 |
-
- stable-diffusion
|
5 |
-
- image-to-image
|
6 |
-
---
|
7 |
-
# SD-XL 1.0-refiner Model Card
|
8 |
-
![row01](01.png)
|
9 |
-
|
10 |
-
## Model
|
11 |
-
|
12 |
-
![pipeline](pipeline.png)
|
13 |
-
|
14 |
-
[SDXL](https://arxiv.org/abs/2307.01952) consists of an [ensemble of experts](https://arxiv.org/abs/2211.01324) pipeline for latent diffusion:
|
15 |
-
In a first step, the base model (available here: https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) is used to generate (noisy) latents,
|
16 |
-
which are then further processed with a refinement model specialized for the final denoising steps.
|
17 |
-
Note that the base model can be used as a standalone module.
|
18 |
-
|
19 |
-
Alternatively, we can use a two-stage pipeline as follows:
|
20 |
-
First, the base model is used to generate latents of the desired output size.
|
21 |
-
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")
|
22 |
-
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.
|
23 |
-
|
24 |
-
Source code is available at https://github.com/Stability-AI/generative-models .
|
25 |
-
|
26 |
-
### Model Description
|
27 |
-
|
28 |
-
- **Developed by:** Stability AI
|
29 |
-
- **Model type:** Diffusion-based text-to-image generative model
|
30 |
-
- **License:** [CreativeML Open RAIL++-M License](https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/blob/main/LICENSE.md)
|
31 |
-
- **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 two fixed, pretrained text encoders ([OpenCLIP-ViT/G](https://github.com/mlfoundations/open_clip) and [CLIP-ViT/L](https://github.com/openai/CLIP/tree/main)).
|
32 |
-
- **Resources for more information:** Check out our [GitHub Repository](https://github.com/Stability-AI/generative-models) and the [SDXL report on arXiv](https://arxiv.org/abs/2307.01952).
|
33 |
-
|
34 |
-
### Model Sources
|
35 |
-
|
36 |
-
For research purposes, we recommned our `generative-models` Github repository (https://github.com/Stability-AI/generative-models), which implements the most popoular diffusion frameworks (both training and inference) and for which new functionalities like distillation will be added over time.
|
37 |
-
[Clipdrop](https://clipdrop.co/stable-diffusion) provides free SDXL inference.
|
38 |
-
|
39 |
-
- **Repository:** https://github.com/Stability-AI/generative-models
|
40 |
-
- **Demo:** https://clipdrop.co/stable-diffusion
|
41 |
-
|
42 |
-
|
43 |
-
## Evaluation
|
44 |
-
![comparison](comparison.png)
|
45 |
-
The chart above evaluates user preference for SDXL (with and without refinement) over SDXL 0.9 and Stable Diffusion 1.5 and 2.1.
|
46 |
-
The SDXL base model performs significantly better than the previous variants, and the model combined with the refinement module achieves the best overall performance.
|
47 |
-
|
48 |
-
|
49 |
-
### 🧨 Diffusers
|
50 |
-
|
51 |
-
Make sure to upgrade diffusers to >= 0.18.0:
|
52 |
-
```
|
53 |
-
pip install diffusers --upgrade
|
54 |
-
```
|
55 |
-
|
56 |
-
In addition make sure to install `transformers`, `safetensors`, `accelerate` as well as the invisible watermark:
|
57 |
-
```
|
58 |
-
pip install invisible_watermark transformers accelerate safetensors
|
59 |
-
```
|
60 |
-
|
61 |
-
Yon can then use the refiner to improve images.
|
62 |
-
|
63 |
-
```py
|
64 |
-
import torch
|
65 |
-
from diffusers import StableDiffusionXLImg2ImgPipeline
|
66 |
-
from diffusers.utils import load_image
|
67 |
-
|
68 |
-
pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained(
|
69 |
-
"stabilityai/stable-diffusion-xl-refiner-1.0", torch_dtype=torch.float16, variant="fp16", use_safetensors=True
|
70 |
-
)
|
71 |
-
pipe = pipe.to("cuda")
|
72 |
-
url = "https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/aa_xl/000000009.png"
|
73 |
-
|
74 |
-
init_image = load_image(url).convert("RGB")
|
75 |
-
prompt = "a photo of an astronaut riding a horse on mars"
|
76 |
-
image = pipe(prompt, image=init_image).images
|
77 |
-
```
|
78 |
-
|
79 |
-
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:
|
80 |
-
```py
|
81 |
-
pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
|
82 |
-
```
|
83 |
-
|
84 |
-
If you are limited by GPU VRAM, you can enable *cpu offloading* by calling `pipe.enable_model_cpu_offload`
|
85 |
-
instead of `.to("cuda")`:
|
86 |
-
|
87 |
-
```diff
|
88 |
-
- pipe.to("cuda")
|
89 |
-
+ pipe.enable_model_cpu_offload()
|
90 |
-
```
|
91 |
-
|
92 |
-
For more advanced use cases, please have a look at [the docs](https://huggingface.co/docs/diffusers/main/en/api/pipelines/stable_diffusion/stable_diffusion_xl).
|
93 |
-
|
94 |
-
## Uses
|
95 |
-
|
96 |
-
### Direct Use
|
97 |
-
|
98 |
-
The model is intended for research purposes only. Possible research areas and tasks include
|
99 |
-
|
100 |
-
- Generation of artworks and use in design and other artistic processes.
|
101 |
-
- Applications in educational or creative tools.
|
102 |
-
- Research on generative models.
|
103 |
-
- Safe deployment of models which have the potential to generate harmful content.
|
104 |
-
- Probing and understanding the limitations and biases of generative models.
|
105 |
-
|
106 |
-
Excluded uses are described below.
|
107 |
-
|
108 |
-
### Out-of-Scope Use
|
109 |
-
|
110 |
-
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.
|
111 |
-
|
112 |
-
## Limitations and Bias
|
113 |
-
|
114 |
-
### Limitations
|
115 |
-
|
116 |
-
- The model does not achieve perfect photorealism
|
117 |
-
- The model cannot render legible text
|
118 |
-
- 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”
|
119 |
-
- Faces and people in general may not be generated properly.
|
120 |
-
- The autoencoding part of the model is lossy.
|
121 |
-
|
122 |
-
### Bias
|
123 |
-
While the capabilities of image generation models are impressive, they can also reinforce or exacerbate social biases.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|