This script is an addon for AUTOMATIC1111's Stable Diffusion WebUI that creates depth maps, and now also 3D stereo image pairs as side-by-side or anaglyph from a single image. The result can be viewed on 3D or holographic devices like VR headsets or Looking Glass displays, used in Render- or Game- Engines on a plane with a displacement modifier, and maybe even 3D printed.
To generate realistic depth maps from a single image, this script uses code and models from the MiDaS repository 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.
3D stereo, and red/cyan anaglyph images are generated using code from the stereo-image-generation repository. Thanks to @sina-masoud-ansari for the tip! Discussion here. Improved techniques for generating stereo images and balancing distortion between eyes by @semjon00, see here and here.
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 by @DanielGatis support added by @graemeniedermayer, using U-2-Net by @xuebinqin to remove backgrounds.
video by @graemeniedermayer, more examples here
images generated by @semjon00 from CC0 photos, more examples here.
- v0.3.8 bugfix
- bugfix in remove background path
- v0.3.7 new features
- rembg Remove Background PR by @graemeniedermayer merged
- setting to flip Left/Right SBS images
- added missing parameter for 3d inpainting (repeat_inpaint_edge)
- option to generate demo videos with mesh
- v0.3.6 new feature
- implemented binary ply file format for the inpainted 3D mesh, big reduction in filesize and save/load times.
- added progress indicators to the inpainting process
- v0.3.5 bugfix
- create path to 3dphoto models before download (see issue)
- v0.3.4 new featues
- depth clipping option (original idea by @Extraltodeus)
- by popular demand, 3D-Photo-Inpainting is now implemented
- generate inpainted 3D mesh (PLY) and videos of said mesh
- v0.3.3 bugfix and new midas models
- updated to midas 3.1, bringing 2 new depth models (the 512 one eats VRAM for breakfast!)
- fix Next-ViT dependency issue for new installs
- extension no longer clones repositories, all dependencies are now contained in the extension
- v0.3.2 new feature and bugfixes
- v0.3.1 bugfix
- small speed increase for anaglyph creation
- clone midas repo before midas 3.1 to fix issue (see here)
- v0.3.0 improved stereo image generation
- v0.2.9 new feature
- 3D Stereo (side-by-side) and red/cyan anaglyph image generation.
(Thanks to @sina-masoud-ansari for the tip! Discussion here)
- 3D Stereo (side-by-side) and red/cyan anaglyph image generation.
- v0.2.8 bugfix
- boost (pix2pix) now also able to compute on cpu
- res101 able to compute on cpu
- v0.2.7 separate tab
- Depth Tab now available for easier stand-alone (batch) processing
- v0.2.6 ui layout and settings
- added link to repo so more people find their way to the instructions.
- boost rmax setting
- v0.2.5 bugfix
- error checking on model download (now with progressbar)
- v0.2.4 high resolution depthmaps
- multi-resolution merging is now implemented, significantly improving results!
- res101 can now also compute on CPU
- v0.2.3 bugfix
- path error on linux fixed
- v0.2.2 new features
- added (experimental) support for AdelaiDepth/LeReS (GPU Only!)
- new option to view depthmap as heatmap
- optimised ui layout
- v0.2.1 bugfix
- Correct seed is now used in filename and pnginfo when running batches. (see issue)
- v0.2.0 upgrade
- the script is now an extension, enabling auto installation.
- v0.1.9 bugfixes
- sd model moved to system memory while computing depthmap
- memory leak/fragmentation issue fixed
- recover from out of memory error
- v0.1.8 new options
- net size can now be set as width and height, option to match input size, sliders now have the same range as generation parameters. (see usage below)
- better error handling
- v0.1.7 bugfixes
- batch img2img now works (see issue)
- generation parameters now only saved when enabled in settings
- model memory freed explicitly at end of script
- v0.1.6 new option
- option to invert depthmap (black=near, white=far), as required by some viewers.
- v0.1.5 bugfix
- saving as any format other than PNG now always produces an 8 bit, 3 channel RGB image. A single channel 16 bit image is only supported when saving as PNG. (see issue)
- v0.1.4 update
- added support for
--no-half. Now also works with cards that don't support half precision like GTX 16xx. (verified)
- added support for
- v0.1.3 bugfix
- bugfix where some controls where not visible (see issue)
- v0.1.2 new option
- network size slider. higher resolution depth maps (see usage below)
- v0.1.1 bugfixes
- overflow issue (see here for details and examples of artifacts)
- when not combining, depthmap is now saved as single channel 16 bit
The script is now also available to install from the Available subtab under the Extensions tab in the WebUI.
In the WebUI, in the Extensions tab, in the Installed subtab, click Check for Updates and then Apply and restart UI.
In the WebUI, in the Extensions tab, in the Install from URL subtab, enter this repository
https://github.com/thygate/stable-diffusion-webui-depthmap-script
and click install and restart.
Model
weightswill be downloaded automatically on first use and saved to /models/midas, /models/leres and /models/pix2pix
Select the "DepthMap vX.X.X" 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 seven 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.
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. Net size is ignored when enabled. 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.
To make the depthmap easier to analyze for human eyes, Show HeatMap shows an extra image in the WebUI that has a color gradient applied. It is not saved.
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 now use the Depth tab to easily process existing images.
- Yes, in img2img, set 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
-
There is the excellent depthy by Rafał Lindemann. LIVE link : https://depthy.stamina.pl/ (Instructions: Drag the rgb image into the window, then select Load depthmap, and drag the depthmap into the dialog inside the window.) Generates GIF and video.
-
The depth-player by @spite can load rgb and depthmap images and export a Wavefront OBJ file of a displaced plane mesh with the rgb image applied as texture. LIVE link : https://depthplayer.ugocapeto.com/ Thanks to @AugmentedRealityCat for the tip.
-
Simple interactive depthmap viewer using three (source). LIVE link : https://thygate.github.io/depthmap-viewer-three (Instructions: Drag a combined-rgb-and-depth-horizontally image into the window to view it)
-
SBS Stereo images can easily be viewed in 3D on VR devices, even cheap ones that use a smartphone like Google Cardboard. To view an SBS image, you may simply display it on the phone screen and then insert the phone into the headset. A more convenient option may be to stream the picture from the computer screen to the phone using Sunshine. You may want to change resolution to match phone's aspect ratio. If you decide to buy a headset, pay attention to the lens' size - usually headsets with larger lenses work the best.
-
Simple interactive depthmap viewer for Looking Glass using three. LIVE link : https://thygate.github.io/depthmap-viewer-three-lookingglass (Instructions: Drag a combined-rgb-and-depth-horizontally image into the window to view it)
-
Unity3D project to view the depthmaps on Looking Glass in realtime as images are generated. Leave a message in the discussion section if you want me to publish it too.
-
Blender depthmap import addon by @Ladypoly (comment).
Download the addon here : importdepthmap_1.0.3.zip (Blender 3.3.0 or newer)
Demonstration videos : (1) https://www.youtube.com/watch?v=vfu5yzs_2EU , (2) https://www.youtube.com/watch?v=AeDngG9kQNI -
To view the 3D-inpainted mesh in blender:
- import the PLY file
- Set camera to origin (0, 0, 0) and pointing up to align it with the mesh
- adjust camera FOV to match the mesh
- Add a 'Color Attribute' Node and connect it to the shader color input
-
Generate normal maps from depth maps : stable-diffusion-webui-normalmap-script by @graemeniedermayer, also check his clothseg extension.
-
Several scripts by @Extraltodeus using depth maps : https://github.com/Extraltodeus?tab=repositories
This project uses 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}
}