This program is an addon for AUTOMATIC1111's Stable Diffusion WebUI that creates depth maps. Using either generated or custom depth maps, it can also create 3D stereo image pairs (side-by-side or anaglyph), normalmaps and 3D meshes. The outputs of the script can be viewed directly or used as an asset for a 3D engine. Please see wiki to learn more. The program has integration with Rembg. It also supports batch processing, processing of videos, and can also be run in standalone mode, without Stable Diffusion WebUI.
To generate realistic depth maps from individual images, this script uses code and models from the MiDaS and ZoeDepth repositories by Intel ISL, or LeReS from the AdelaiDepth repository by Advanced Intelligent Machines. Multi-resolution merging as implemented by BoostingMonocularDepth is used to generate high resolution depth maps.
Stereoscopic images are created using a custom-written algorithm.
3D Photography using Context-aware Layered Depth Inpainting by Virginia Tech Vision and Learning Lab, or 3D-Photo-Inpainting is used to generate a 3D inpainted mesh and render videos from said mesh.
Rembg uses U-2-Net and IS-Net.
video by @graemeniedermayer, more examples here
images generated by @semjon00 from CC0 photos, more examples here.
The script can be installed directly from WebUI. Please navigate to Extensions tab, then click Available, Load from and then install the Depth Maps extension. Alternatively, the extension can be installed from the URL: https://github.com/thygate/stable-diffusion-webui-depthmap-script.
In the WebUI, in the Extensions tab, in the Installed subtab, click Check for Updates and then Apply and restart UI.
Clone the repository, install the requirements from requirements.txt, launch using main.py.
Model weights will be downloaded automatically on their first use and saved to /models/midas, /models/leres and /models/pix2pix. Zoedepth models are stored in the torch cache folder.
Select the "DepthMap" script from the script selection box in either txt2img or img2img, or go to the Depth tab when using existing images.
The models can Compute on GPU and CPU, use CPU if low on VRAM.
There are ten models available from the Model dropdown. For the first model, res101, see AdelaiDepth/LeReS for more info. The others are the midas models: dpt_beit_large_512, dpt_beit_large_384, dpt_large_384, dpt_hybrid_384, midas_v21, and midas_v21_small. See the MiDaS repository for more info. The newest dpt_beit_large_512 model was trained on a 512x512 dataset but is VERY VRAM hungry. The last three models are ZoeDepth models.
Net size can be set with net width and net height, or will be the same as the input image when Match input size is enabled. There is a trade-off between structural consistency and high-frequency details with respect to net size (see observations).
Boost will enable multi-resolution merging as implemented by BoostingMonocularDepth and will significantly improve the results, mitigating the observations mentioned above, at the cost of much larger compute time. Best results with res101.
Clip and renormalize allows for clipping the depthmap on the near and far side, the values in between will be renormalized to fit the available range. Set both values equal to get a b&w mask of a single depth plane at that value. This option works on the 16-bit depthmap and allows for 1000 steps to select the clip values.
When enabled, Invert DepthMap will result in a depthmap with black near and white far.
Regardless of global settings, Save DepthMap will always save the depthmap in the default txt2img or img2img directory with the filename suffix '_depth'. Generation parameters are saved with the image if enabled in settings. Files generated from the Depth tab are saved in the default extras-images directory.
To see the generated output in the webui Show DepthMap should be enabled. When using Batch img2img this option should also be enabled.
When Combine into one image is enabled, the depthmap will be combined with the original image, the orientation can be selected with Combine axis. When disabled, the depthmap will be saved as a 16 bit single channel PNG as opposed to a three channel (RGB), 8 bit per channel image when the option is enabled.
When either Generate Stereo or Generate anaglyph is enabled, a stereo image pair will be generated. Divergence sets the amount of 3D effect that is desired. Balance between eyes determines where the (inevitable) distortion from filling up gaps will end up, -1 Left, +1 Right, and 0 balanced.
The different Gap fill technique options are : none (no gaps are filled),
naive (the original method), naive_interpolating (the original method with interpolation), polylines_soft and polylines_sharp are the latest technique, the last one being best quality and slowest. Note: All stereo image generation is done on CPU.
To generate the mesh required to generate videos, enable Generate 3D inpainted mesh. This can be a lengthy process, from a few minutes for small images to an hour for very large images. This option is only available on the Depth tab. When enabled, the mesh in ply format and four demo video are generated. All files are saved to the extras directory.
Videos can be generated from the PLY mesh on the Depth Tab.
It requires the mesh created by this extension, files created elsewhere might not work corectly, as some extra info is stored in the file (required value for dolly). Most options are self-explanatory, like Number of frames and Framerate. Two output formats are supported: mp4 and webm. Supersampling Anti-Aliasing (SSAA) can be used to get rid of jagged edges and flickering. The render size is scaled by this factor and then downsampled.
There are three trajectories to choose from : circle, straight-line, double-straight-line, to translate in three dimensions. The border can be cropped on four sides, and the Dolly option adjusts the FOV so the center subject will stay approximately the same size, like the dolly-zoom.
Settings on WebUI Settings tab :
Maximum wholesize for boost sets the r_max value from the BoostingMonocularDepth paper, it relates to the max size that is chosen to render at internally, and directly influences the max amount of VRAM that could be used. The default value for this from the paper is 3000, I have lowered the value to 1600 so it will work more often with 8GB VRAM GPU's.
If you often get out of memory errors when computing a depthmap on GPU while using Boost, you can try lowering this value. Note the 'wholeImage being processed in : xxxx' output when using boost, this number will never be greater than the r_max, but can be larger with a larger r_max. See the paper for more details.
💡 Saving as any format other than PNG always produces an 8 bit, 3 channel RGB image. A single channel 16 bit image is only supported when saving as PNG.
Can I use this on existing images ?- Yes, you can use the Depth tab to easily process existing images.
- Another way of doing this would be to use img2img with denoising strength to 0. This will effectively skip stable diffusion and use the input image. You will still have to set the correct size, and need to select
Crop and resizeinstead ofJust resizewhen the input image resolution does not match the set size perfectly.
Can I run this on Google Colab?- You can run the MiDaS network on their colab linked here https://pytorch.org/hub/intelisl_midas_v2/
- You can run BoostingMonocularDepth on their colab linked here : https://colab.research.google.com/github/compphoto/BoostingMonocularDepth/blob/main/Boostmonoculardepth.ipynb
- Running this program on Colab is not officially supported, but it may work. Please look for more suitable ways of running this. If you still decide to try, standalone installation may be easier to manage.
What other depth-related projects could I check out?- Several scripts by @Extraltodeus using depth maps.
- geo-11 and Depth3D for playing existing games in 3D.
How can I know what changed in the new version of the script?- You can see the git history log or refer to the
CHANGELOG.mdfile.
- You can see the git history log or refer to the
Feel free to comment and share in the discussions! Submitting issues and merge requests is heavilly appreciated!
This project relies on code and information from following papers :
MiDaS :
@article {Ranftl2022,
author = "Ren\'{e} Ranftl and Katrin Lasinger and David Hafner and Konrad Schindler and Vladlen Koltun",
title = "Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer",
journal = "IEEE Transactions on Pattern Analysis and Machine Intelligence",
year = "2022",
volume = "44",
number = "3"
}
Dense Prediction Transformers, DPT-based model :
@article{Ranftl2021,
author = {Ren\'{e} Ranftl and Alexey Bochkovskiy and Vladlen Koltun},
title = {Vision Transformers for Dense Prediction},
journal = {ICCV},
year = {2021},
}
AdelaiDepth/LeReS :
@article{yin2022towards,
title={Towards Accurate Reconstruction of 3D Scene Shape from A Single Monocular Image},
author={Yin, Wei and Zhang, Jianming and Wang, Oliver and Niklaus, Simon and Chen, Simon and Liu, Yifan and Shen, Chunhua},
journal={TPAMI},
year={2022}
}
@inproceedings{Wei2021CVPR,
title = {Learning to Recover 3D Scene Shape from a Single Image},
author = {Wei Yin and Jianming Zhang and Oliver Wang and Simon Niklaus and Long Mai and Simon Chen and Chunhua Shen},
booktitle = {Proc. IEEE Conf. Comp. Vis. Patt. Recogn. (CVPR)},
year = {2021}
}
Boosting Monocular Depth Estimation Models to High-Resolution via Content-Adaptive Multi-Resolution Merging :
@inproceedings{Miangoleh2021Boosting,
title={Boosting Monocular Depth Estimation Models to High-Resolution via Content-Adaptive Multi-Resolution Merging},
author={S. Mahdi H. Miangoleh and Sebastian Dille and Long Mai and Sylvain Paris and Ya\u{g}{\i}z Aksoy},
journal={Proc. CVPR},
year={2021},
}
3D Photography using Context-aware Layered Depth Inpainting :
@inproceedings{Shih3DP20,
author = {Shih, Meng-Li and Su, Shih-Yang and Kopf, Johannes and Huang, Jia-Bin},
title = {3D Photography using Context-aware Layered Depth Inpainting},
booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2020}
}
U2-Net:
@InProceedings{Qin_2020_PR,
title = {U2-Net: Going Deeper with Nested U-Structure for Salient Object Detection},
author = {Qin, Xuebin and Zhang, Zichen and Huang, Chenyang and Dehghan, Masood and Zaiane, Osmar and Jagersand, Martin},
journal = {Pattern Recognition},
volume = {106},
pages = {107404},
year = {2020}
}
IS-Net:
@InProceedings{qin2022,
author={Xuebin Qin and Hang Dai and Xiaobin Hu and Deng-Ping Fan and Ling Shao and Luc Van Gool},
title={Highly Accurate Dichotomous Image Segmentation},
booktitle={ECCV},
year={2022}
}
ZoeDepth :
@misc{https://doi.org/10.48550/arxiv.2302.12288,
doi = {10.48550/ARXIV.2302.12288},
url = {https://arxiv.org/abs/2302.12288},
author = {Bhat, Shariq Farooq and Birkl, Reiner and Wofk, Diana and Wonka, Peter and Müller, Matthias},
keywords = {Computer Vision and Pattern Recognition (cs.CV), FOS: Computer and information sciences, FOS: Computer and information sciences},
title = {ZoeDepth: Zero-shot Transfer by Combining Relative and Metric Depth},
publisher = {arXiv},
year = {2023},
copyright = {arXiv.org perpetual, non-exclusive license}
}