ControlNet is a neural network structure to control diffusion models by adding extra conditions.
img
It copys the weights of neural network blocks into a "locked" copy and a "trainable" copy.
The "trainable" one learns your condition. The "locked" one preserves your model.
Thanks to this, training with small dataset of image pairs will not destroy the production-ready diffusion models.
The "zero convolution" is 1×1 convolution with both weight and bias initialized as zeros.
Before training, all zero convolutions output zeros, and ControlNet will not cause any distortion.
No layer is trained from scratch. You are still fine-tuning. Your original model is safe.
This allows training on small-scale or even personal devices.
This is also friendly to merge/replacement/offsetting of models/weights/blocks/layers.
FAQ
Q: But wait, if the weight of a conv layer is zero, the gradient will also be zero, and the network will not learn anything. Why "zero convolution" works?
By repeating the above simple structure 14 times, we can control stable diffusion in this way:
img
Note that the way we connect layers is computational efficient. The original SD encoder does not need to store gradients (the locked original SD Encoder Block 1234 and Middle). The required GPU memory is not much larger than original SD, although many layers are added. Great!
2023/02/11 - Low VRAM mode is added. Please use this mode if you are using 8GB GPU(s) or if you want larger batch size.
Production-Ready Pretrained Models
First create a new conda environment
conda env create -f environment.yaml
conda activate control
All models and detectors can be downloaded from our Hugging Face page. Make sure that SD models are put in "ControlNet/models" and detectors are put in "ControlNet/annotator/ckpts". Make sure that you download all necessary pretrained weights and detector models from that Hugging Face page, including HED edge detection model, Midas depth estimation model, Openpose, and so on.
We provide 9 Gradio apps with these models.
All test images can be found at the folder "test_imgs".
Note that the UI is based on Gradio, and Gradio is somewhat difficult to customize. Right now you need to draw scribbles outside the UI (using your favorite drawing software, for example, MS Paint) and then import the scribble image to Gradio.
Prompt: "turtle"
p
Prompt: "hot air balloon"
p
Interactive Interface
We actually provide an interactive interface
python gradio_scribble2image_interactive.py
However, because gradio is very buggy and difficult to customize, right now, user need to first set canvas width and heights and then click "Open drawing canvas" to get a drawing area. Please do not upload image to that drawing canvas. Also, the drawing area is very small; it should be bigger. But I failed to find out how to make it larger. Again, gradio is really buggy. (Now fixed, will update asap)
The below dog sketch is drawn by me. Perhaps we should draw a better dog for showcase.
Sometimes we are lazy, and we do not want to draw scribbles. This script use the exactly same scribble-based model but use a simple algorithm to synthesize scribbles from input images.
Prompt: "bag"
p
Prompt: "shose" (Note that "shose" is a typo; it should be "shoes". But it still seems to work.)
p
ControlNet with Human Pose
Stable Diffusion 1.5 + ControlNet (using human pose)
python gradio_pose2image.py
Apparently, this model deserves a better UI to directly manipulate pose skeleton. However, again, Gradio is somewhat difficult to customize. Right now you need to input an image and then the Openpose will detect the pose for you.
This model use ADE20K's segmentation protocol. Again, this model deserves a better UI to directly draw the segmentations. However, again, Gradio is somewhat difficult to customize. Right now you need to input an image and then a model called Uniformer will detect the segmentations for you. Just try it for more details.
Great! Now SD 1.5 also have a depth control. FINALLY. So many possibilities (considering SD1.5 has much more community models than SD2).
Note that different from Stability's model, the ControlNet receive the full 512×512 depth map, rather than 64×64 depth. Note that Stability's SD2 depth model use 64*64 depth maps. This means that the ControlNet will preserve more details in the depth map.
This is always a strength because if users do not want to preserve more details, they can simply use another SD to post-process an i2i. But if they want to preserve more details, ControlNet becomes their only choice. Again, SD2 uses 64×64 depth, we use 512×512.
Prompt: "Stormtrooper's lecture"
p
ControlNet with Normal Map
Stable Diffusion 1.5 + ControlNet (using normal map)
python gradio_normal2image.py
This model use normal map. Rightnow in the APP, the normal is computed from the midas depth map and a user threshold (to determine how many area is background with identity normal face to viewer, tune the "Normal background threshold" in the gradio app to get a feeling).
Prompt: "Cute toy"
p
Prompt: "Plaster statue of Abraham Lincoln"
p
Compared to depth model, this model seems to be a bit better at preserving the geometry. This is intuitive: minor details are not salient in depth maps, but are salient in normal maps. Below is the depth result with same inputs. You can see that the hairstyle of the man in the input image is modified by depth model, but preserved by the normal model.
Prompt: "Plaster statue of Abraham Lincoln"
p
ControlNet with Anime Line Drawing
We also trained a relatively simple ControlNet for anime line drawings. This tool may be useful for artistic creations. (Although the image details in the results is a bit modified, since it still diffuse latent images.)
You need to manually check the "Guess Mode" toggle to enable this mode.
In this mode, the ControlNet encoder will try best to recognize the content of the input control map, like depth map, edge map, scribbles, etc, even if you remove all prompts.
Let's have fun with some very challenging experimental settings!
No prompts. No "positive" prompts. No "negative" prompts. No extra caption detector. One single diffusion loop.
For this mode, we recommend to use 50 steps and guidance scale between 3 and 5.
p
No prompts:
p
Note that the below example is 768×768. No prompts. No "positive" prompts. No "negative" prompts.
p
By tuning the parameters, you can get some very intereting results like below:
p
Because no prompt is available, the ControlNet encoder will "guess" what is in the control map. Sometimes the guess result is really interesting. Because diffusion algorithm can essentially give multiple results, the ControlNet seems able to give multiple guesses, like this:
p
Without prompt, the HED seems good at generating images look like paintings when the control strength is relatively low:
No prompts. Default WebUI parameters. Pure random results with the seed being 12345. Standard SD1.5. Input scribble is in "test_imgs" folder to reproduce.
p
Below is another challenging example:
p
No prompts. Default WebUI parameters. Pure random results with the seed being 12345. Standard SD1.5. Input scribble is in "test_imgs" folder to reproduce.
p
Note that in the guess mode, you will still be able to input prompts. The only difference is that the model will "try harder" to guess what is in the control map even if you do not provide the prompt. Just try it yourself!
Besides, if you write some scripts (like BLIP) to generate image captions from the "guess mode" images, and then use the generated captions as prompts to diffuse again, you will get a SOTA pipeline for fully automatic conditional image generating.
Combining Multiple ControlNets
ControlNets are composable: more than one ControlNet can be easily composed to multi-condition control.
As long as the models are controlling the same SD, the "boundary" between different research projects does not even exist. This plugin also allows different methods to work together!
