File size: 5,083 Bytes
ef25943 e677eaf ef25943 55978e9 ef25943 55978e9 e677eaf 55978e9 e677eaf e3473c0 e677eaf 55978e9 e677eaf e3473c0 e677eaf 55978e9 e677eaf 55978e9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 |
---
license: apache-2.0
library_name: diffusers
tags:
- stable-diffusion-xl
- stable-diffusion-xl-diffusers
- text-to-image
- diffusers
- controlnet
- diffusers-training
---
# SDXL ControlNet: DWPose
Here are ControlNet weights trained on stabilityai/stable-diffusion-xl-base-1.0 with [DWPose](https://github.com/IDEA-Research/DWPose) conditioning.
## Using in 🧨 diffusers
First, install all the libraries:
```bash
pip install -q easy-dwpose transformers accelerate
pip install -q git+https://github.com/huggingface/diffusers
```
### Example 1
To generate a realistic DJ with the following image driving the pose:
![Pose Image 1](./images/pose_image_1.png)
Run the following code:
```python
from diffusers import ControlNetModel, StableDiffusionXLControlNetPipeline
import torch
from diffusers.utils import load_image
from easy_dwpose import DWposeDetector
pose_image = load_image("./pose_image_1.png")
# Load detector
device = "cuda:0" if torch.cuda.is_available() else "cpu"
dwpose = DWposeDetector(device=device)
# Compute DWpose conditioning image.
skeleton = dwpose(
pose_image,
detect_resolution=pose_image.width,
output_type="pil",
include_hands=True,
include_face=True,
)
# Initialize ControlNet pipeline.
controlnet = ControlNetModel.from_pretrained(
"dimitribarbot/controlnet-dwpose-sdxl-1.0",
torch_dtype=torch.float16,
)
pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
controlnet=controlnet,
torch_dtype=torch.float16,
variant="fp16",
).to(device)
# Infer.
prompt = "DJ in a party, shallow depth of field, highly detailed, high budget, gorgeous"
negative_prompt = "bad quality, blur, anime, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured"
image = pipe(
prompt,
negative_prompt=negative_prompt,
num_inference_steps=50,
guidance_scale=5,
image=skeleton,
generator=torch.manual_seed(97),
).images[0]
```
Generated pose is:
![Pose 1](./images/dwpose_1.png)
Image generated by SDXL is:
![Pose 1](./images/dwpose_image_1.png)
### Example 2
To generate a anime version of a woman sitting on a bench with the following image driving the pose:
![Pose Image 2](./images/pose_image_2.png)
Run the following code:
```python
from diffusers import ControlNetModel, StableDiffusionXLControlNetPipeline
import torch
from diffusers.utils import load_image
from easy_dwpose import DWposeDetector
pose_image = load_image("./pose_image_2.png")
# Load detector
device = "cuda:0" if torch.cuda.is_available() else "cpu"
dwpose = DWposeDetector(device=device)
# Compute DWpose conditioning image.
skeleton = dwpose(
pose_image,
detect_resolution=pose_image.width,
output_type="pil",
include_hands=True,
include_face=True,
)
# Initialize ControlNet pipeline.
controlnet = ControlNetModel.from_pretrained(
"dimitribarbot/controlnet-dwpose-sdxl-1.0",
torch_dtype=torch.float16,
)
pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
controlnet=controlnet,
torch_dtype=torch.float16,
variant="fp16",
)
if torch.cuda.is_available():
pipe.to(torch.device("cuda"))
# Infer.
prompt = "Anime girl sitting on a bench, highly detailed, noon, ambiant light"
negative_prompt = "bad quality, blur, anime, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured"
image = pipe(
prompt,
negative_prompt=negative_prompt,
num_inference_steps=25,
guidance_scale=18,
image=skeleton,
generator=torch.manual_seed(79),
).images[0]
```
Generated pose is:
![Pose 2](./images/dwpose_2.png)
Image generated by SDXL is:
![Pose 2](./images/dwpose_image_2.png)
## Training
The [training script](https://github.com/huggingface/diffusers/blob/main/examples/controlnet/README_sdxl.md) by HF🤗 was used.
#### Training data
This checkpoint was trained for 15,000 steps on the [dimitribarbot/dw_pose_controlnet](https://huggingface.co/datasets/dimitribarbot/dw_pose_controlnet) dataset with a resolution of 1024.
#### Compute
One 1xA40 machine (during 48 hours)
#### Batch size
Data parallel with a single GPU batch size of 2 with gradient accumulation 8.
#### Hyper Parameters
Constant learning rate of 8e-5
#### Mixed precision
fp16
## Thanks
- [raulc0399](https://huggingface.co/raulc0399): for highly inspiring me with the creation of the [DWpose dataset](https://huggingface.co/datasets/dimitribarbot/dw_pose_controlnet) based on the [Openpose one](https://huggingface.co/datasets/raulc0399/open_pose_controlnet).
- [thibaud](https://huggingface.co/thibaud): for highly inspiring me with the hyper parameters of the HF training script, based on the [Openpose ControlNet](https://huggingface.co/thibaud/controlnet-openpose-sdxl-1.0).
- [RedHash](https://huggingface.co/RedHash): for the [easy_dwpose](https://github.com/reallyigor/easy_dwpose) module, which highly simplifies the DWPose inference and which I used in the examples above.
- [Hugging Face](https://huggingface.co/): for the ControlNet training script 🤗. |