Solvent

AI-Assisted Texture Generation Toolkit in Blender

Features

Text-to-Image Texture Generation
- Generate textures, concept art, and background assets by using Stable Diffusion
- Generate multiple texture images, all readily accessible in Blender
- Generate seamless and tileable textures
- Experiment with various adjustable configuration parameters
- Enable optimizations by default to save GPU VRAM and time taken for generation
Image-to-Image Texture Generation
- Create variations on pre-existing input textures, concept art, and background assets
- Provide mask images to indicate which sections of input textures to keep or modify

Prerequisites

Ensure that you already have the following software installed, preferably the latest versions:

If you are on MacOS and Linux, you would also need to install Rust. This is because the transformers Python library uses the tokenizers library's wheel built in Rust.

Additionally, you would also need to have a Hugging Face account to download the pre-trained Stable Diffusion model checkpoint weights. You would need to agree to the Terms and Conditions stipulated by the CreativeML Open RAIL-M License here before you are able to download their pre-trained model weights.

Optionally, you might want to create a GitHub account as well to raise issues or pull requests to this repository.

Instructions

Note that if you are on Windows, you would need to run Blender (and hence, the add-on) as administrator with Internet connection in order to install all of the necessary Python packages. This only needs to be done once.

Otherwise, you would need to install Blender by visiting the official site here instead of using other methods (i.e., such as using a package manager or snap):

Repositories used by package managers might not have the latest version of Blender.

More details about the issues of a snap-based installation can be found here.

Do also ensure that you possess the appropriate permissions to access the directory that you install Blender to.

Download the latest version of the add-on files from this repository:
```
git clone --recurse-submodules https://github.com/jamestiotio/solvent
```
Note that the size can be quite large as it includes the pre-trained Stable Diffusion model checkpoint weights. Depending on your Internet speed, you can go get a cup of coffee while waiting.
Zip up the solvent folder that has been cloned. This would create a solvent.zip file.
Install the add-on by opening Blender and following the instructions on the official Blender manual here. Follow along the instructions to also enable the add-on.
Since the add-on will need to download and install additional Python packages, you can trigger the installation manually by going to Edit > Preferences > Add-ons and click the Install Python packages button. Don't worry if Blender seems to hang and not respond for a while. You can monitor the progress of the installation in Blender's Console Window. If you are on Windows, it will be automatically opened when the aforementioned button is pressed. Otherwise, you might need to run Blender from a terminal and view the Console Window output there instead. Depending on your Internet speed, you can go get another cup of coffee while waiting. After the installation has finished, you would need to restart Blender.
Enable the Sidebar (go to View > Sidebar or click the N key) if it is not enabled yet. You can then go to Sidebar > Solvent and input your parameters to generate the texture that you want. After setting the parameters to your heart's content, press the Generate Texture button and wait. This might take a while on a CUDA-compatible GPU and even longer on a CPU. Depending on your CPU/GPU's speed and the parameters that you have chosen, you can go get yet another cup of coffee while waiting.
If you have selected an object/mesh before generating the texture, the add-on will also attempt to automatically apply the texture (or the first generated texture, if you have indicated to generate multiple texture images) to the currently-active material (as the Base Color of the Principled BSDF node). Otherwise, you can find the texture image in the folder that you have specified and you can apply it to the material yourself.
???
Profit!!!

Requirements

On top of Blender's requirements specified here, you would also need:

Depending on what your operating system is, around 5-6 GB of space if you use the CPU-only version of PyTorch or around 7-8 GB of space if you use the CUDA GPU version of PyTorch.
Around 16 GB of RAM. More is highly recommended.
Preferably either an Apple Mac with the M1 chip (which supports MPS) or a device with a CUDA-compatible NVIDIA GPU as listed here. This is because PyTorch currently only works with either MPS or CUDA-compatible NVIDIA GPUs for GPU acceleration. You could try using PyTorch with CPU (which would run more slowly) if you do not have an M1 Mac or a CUDA-compatible NVIDIA GPU. The GPU should also possess sufficient dedicated GPU memory, preferably 6.5 GB or more. Otherwise, if the GPU possesses insufficient dedicated memory, Blender might intermittently and randomly crash.

Known Issues

If you are on Windows, you might sometimes encounter errors if you have not enabled long file and directory paths. To fix this, simply execute the attached PowerShell script as an administrator.

Fair warning: the script will modify your Windows Registry, which might break your system if anything goes wrong. If you do not trust the script or if you do not understand what it is doing, then do not run it. For this very reason, the script is not automatically executed by this add-on during setup. You have been warned.
As mentioned here, MacOS binaries of PyTorch do not support CUDA yet. Thus, building and installing PyTorch from source would be required if you would like to use CUDA with it. You can follow the instructions outlined here to do that and ensure that Blender's Python interpreter can detect it.
For some reason, if you are unlucky, the PyTorch package download speed can be quite slow as mentioned here. Until the root cause and a permanent solution can be found, this issue will remain.

TODOs

Add image upscaler module.
Add material map generator module (normal map, displacement map, roughness map).
Add prompt engineering to assist beginners with experimenting with Stable Diffusion.
Store the history of prompts created by a particular user.
Implement asynchronous Python package downloading and texture generation to allow Blender to stay interactive during processing.
Add support to use MPS acceleration on M1 Mac once the relevant PyTorch version becomes stable.
Ensure cross-compatibility across Windows, Mac, and Linux.
Check the minimum Blender version that this add-on can support.

If you encounter any problems or have any suggestions, feel free to raise an issue or a pull request!

Licenses

This add-on is licensed under the GNU General Public License v2.0 as attached in this repository here, the same license that Blender uses.

This project depends on the following packages:

Package	License
PyTorch	BSD 3-Clause "New" or "Revised" License
Diffusers	Apache License 2.0
Transformers	Apache License 2.0
SciPy	BSD 3-Clause "New" or "Revised" License
ftfy	MIT License
spaCy	MIT License

Furthermore, this project also depends on the saved checkpoint weights of the following pre-trained models:

Model	License
Stable Diffusion v1-4	CreativeML Open RAIL-M License

Any generated texture images are subject to the Creative Commons CC0 1.0 Universal Public Domain Dedication License specified here.

Use of this add-on implies that you agree with all of the terms and conditions mentioned in all of the relevant licenses, which would include whichever model(s) you used to produce the image outputs.

jamestiotio/solvent