# SD3 Controlnet softedge
The softedge controlnet is finetuned based on SD3-medium. It is trained using 12M open source and internal e-commerce dataset, and achieve good performance on both general and e-commerce image generation. It supports preprocessors such as pidinet, hed as well as their safe mode.


## Examples
From left to right: pidinet preprocessor, ours with pidinet, hed preprocessor, ours with hed.

`pidinet`   |`controlnet`|`hed`       |`controlnet`
:--:|:--:|:--:|:--:
![images)](./images/im1_1.webp) | ![images)](./images/im1_2.webp) | ![images)](./images/im1_3.webp) | ![images)](./images/im1_4.webp)
![images)](./images/im2_1.webp) | ![images)](./images/im2_2.webp) | ![images)](./images/im2_3.webp) | ![images)](./images/im2_4.webp)
![images)](./images/im3_1.webp) | ![images)](./images/im3_2.webp) | ![images)](./images/im3_3.webp) | ![images)](./images/im3_4.webp)
![images)](./images/im4_1.webp) | ![images)](./images/im4_2.webp) | ![images)](./images/im4_3.webp) | ![images)](./images/im4_4.webp)
![images)](./images/im5_1.webp) | ![images)](./images/im5_2.webp) | ![images)](./images/im5_3.webp) | ![images)](./images/im5_4.webp)

			
			
## Usage with Diffusers
```python
import torch
from diffusers.utils import load_image, check_min_version
from diffusers.models import SD3ControlNetModel
from diffusers import StableDiffusion3ControlNetPipeline
from controlnet_aux import PidiNetDetector

controlnet = SD3ControlNetModel.from_pretrained(
    "alimama-creative/SD3-Controlnet-Softedge",torch_dtype=torch.float16
)
pipe = StableDiffusion3ControlNetPipeline.from_pretrained(
    "stabilityai/stable-diffusion-3-medium-diffusers",
    controlnet=controlnet,
    variant="fp16",
    torch_dtype=torch.float16,
)
pipe.text_encoder.to(torch.float16)
pipe.controlnet.to(torch.float16)
pipe.to("cuda")

image = load_image(
    "https://huggingface.co/alimama-creative/SD3-Controlnet-Softedge/resolve/main/images/im1_0.png"
)
prompt = "A dog sitting on a park bench."
width = 1024
height = 1024

edge_processor = PidiNetDetector.from_pretrained('lllyasviel/Annotators')
edge_image = edge_processor(image, detect_resolution=width, image_resolution=width)

res_image = pipe(
    prompt=prompt,
    negative_prompt="deformed, distorted, disfigured, poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, mutated hands and fingers, disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation, NSFW",
    height=height,
    width=width,
    control_image=edge_image,
    num_inference_steps=25,
    controlnet_conditioning_scale=0.95,
    guidance_scale=5,
).images[0]
res_image.save("sd3.png")

```

## Training Detail
The model was trained on 12M laion2B and internal sources images with aesthetic 6+ for 20k steps at resolution 1024x1024. ControlNet with 6, 12 and 23 layers have been explored, and the 12-layer model achieves a good balance between performance and model size, so we release the 12-layer model.

Mixed precision : FP16<br/>
Learning rate : 1e-4<br/>
Batch size : 256<br/>
Timestep sampling mode : 'logit_normal'<br/>
Loss : Flow Matching<br/>

## LICENSE
The model is based on SD3 finetuning; therefore, the license follows the original SD3 license.