Canny - Use a Canny edge map to guide the structure of the generated image. This is especially useful for illustrations, but works with all styles.
Depth - use a depth map, generated by DepthFM, to guide generation. Some example use cases include generating architectural renderings, or texturing 3D assets.
Blur - can be used to perform extremely high fidelity upscaling. A common use case is to tile an input image, apply the ControlNet to each tile, and merge the tiles to produce a higher resolution image. A more in-depth description of this use case is here.
We recommend tiling the image at a tile size between 128 and 512.
All currently released ControlNets are compatible only with Stable Diffusion 3.5 Large (8b).
Additional ControlNet models, including 2B versions of the variants above, and multiple other control types, will be added to this repository in the future.
Free for non-commercial use: Individuals and organizations can use the model free of charge for non-commercial use, including scientific research.
Free for commercial use (up to $1M in annual revenue): Startups, small to medium-sized businesses, and creators can use the model for commercial purposes at no cost, as long as their total annual revenue is less than $1M.
Ownership of outputs: Retain ownership of the media generated without restrictive licensing implications.
For organizations with annual revenue more than $1M, please contact us here to inquire about an Enterprise License.
Usage
For local or self-hosted use, we recommend ComfyUI for node-based UI inference, or the standalone SD3.5 repo for programmatic use.
The conditioning image should already be preprocessed before being used as input to the standalone repo; sd3.5 does not implement the preprocessing code below.
Usage in Diffusers 🧨
Make sure you upgrade to the latest version of diffusers: pip install -U diffusers. And then you can run:
python
1import torch
2from diffusers import StableDiffusion3ControlNetPipeline,SD3ControlNetModel
3from diffusers.utils import load_image
45controlnet = SD3ControlNetModel.from_pretrained("diffusers-internal-dev/sd35-controlnet-blur-8b", torch_dtype=torch.float16)6pipe = StableDiffusion3ControlNetPipeline.from_pretrained(7"stabilityai/stable-diffusion-3.5-large",8 controlnet=controlnet,9 torch_dtype=torch.float16
10).to("cuda")1112control_image = load_image("https://huggingface.co/datasets/diffusers/diffusers-images-docs/resolve/main/blur.png")13prompt ="generated ai art, a tiny, lost rubber ducky in an action shot close-up, surfing the humongous waves, inside the tube, in the style of Kelly Slater"1415generator = torch.Generator(device="cpu").manual_seed(0)16image = pipe(17 prompt,18 control_image=control_image,19 guidance_scale=3.5,20 num_inference_steps=60,21 generator=generator,22 max_sequence_length=77,23).images[0]24image.save('blur-8b.jpg')
Preprocessing
Below are code snippets for preprocessing the various control image types.
Canny
python
1import torchvision.transforms.functional as F
2# assuming img is a PIL image3img = F.to_tensor(img)4img = cv2.cvtColor(img.transpose(1,2,0), cv2.COLOR_RGB2GRAY)5img = cv2.Canny(img,100,200)
Blur
python
1import torchvision.transforms as transforms
2# assuming img is a PIL image3gaussian_blur = transforms.GaussianBlur(kernel_size=50)4blurred_image = gaussian_blur(img)
Depth
python
1# install depthfm from https://github.com/CompVis/depth-fm2import torchvision.transforms as transforms
3from depthfm.dfm import DepthFM
4depthfm_model = DepthFM(ckpt_path=checkpoint_path)5depthfm_model.eval()67# assuming img is a PIL image8img = F.to_tensor(img)9c, h, w = img.shape
10img = F.interpolate(img,(512,512), mode='bilinear', align_corners=False)11with torch.no_grad():12 img = self.depthfm_model(img, num_steps=2, ensemble_size=4)13img = F.interpolate(img,(h, w), mode='bilinear', align_corners=False)
Tips
We recommend starting with a ControlNet strength of 0.7-0.8, and adjusting as needed.
Euler sampler and a slightly higher step count (50-60) gives best results, especially with Canny.
Pass --text_encoder_device <device_name> to load the text encoders directly to VRAM, which can speed up the full inference loop at the cost of extra VRAM usage.
The model was not trained to be factual or true representations of people or events. As such, using the model to generate such content is out-of-scope of the abilities of this model.
Training Data and Strategy
These models were trained on a wide variety of data, including synthetic data and filtered publicly available data.
Safety
We believe in safe, responsible AI practices and take deliberate measures to ensure Integrity starts at the early stages of development. This means we have taken and continue to take reasonable steps to prevent the misuse of Stable Diffusion 3.5 by bad actors. For more information about our approach to Safety please visit our Safety page.
Integrity Evaluation
Our integrity evaluation methods include structured evaluations and red-teaming testing for certain harms. Testing was conducted primarily in English and may not cover all possible harms.
Risks identified and mitigations:
Harmful content: We have implemented safeguards that attempt to strike the right balance between usefulness and preventing harm. However, this does not guarantee that all possible harmful content has been removed. All developers and deployers should exercise caution and implement content safety guardrails based on their specific product policies and application use cases.
Misuse: Technical limitations and developer and end-user education can help mitigate against malicious applications of models. All users are required to adhere to our Acceptable Use Policy, including when applying fine-tuning and prompt engineering mechanisms. Please reference the Stability AI Acceptable Use Policy for information on violative uses of our products.
Privacy violations: Developers and deployers are encouraged to adhere to privacy regulations with techniques that respect data privacy.
Acknowledgements
Lvmin Zhang, Anyi Rao, and Maneesh Agrawala, authors of the original ControlNet paper.
Lvmin Zhang, who also developed the Tile ControlNet, which inspired the Blur ControlNet.
Diffusers library authors, whose code was referenced during development.
InstantX team, whose Flux and SD3 ControlNets were also referenced during training.
All early testers and raters of the models, and the Stability AI team.
Contact
Please report any issues with the model or contact us:
