Tiled Diffusion & VAE Extension for SD-WebUI
The "Tiled Diffusion & VAE extension" is a powerful tool created for sd-webui, designed to help users generate or upscale large images, specifically those with dimensions of 2K and above, while using limited VRAM, typically 6GB or less. The extension is particularly valuable for artists and developers who wish to maximize their graphical output without hefty hardware investments.
Key Technologies
The following techniques are integral to the extension's functionality:
- State-of-the-art Tiled Diffusion Methods: The extension reproduces leading-edge diffusion techniques such as:
- Mixture of Diffusers
- MultiDiffusion
- Demofusion
- Original Tiled VAE (Variational Autoencoder) Method: A proprietary method developed by the creators to enhance image processing.
- Original Tiled Noise Inversion Method: Another in-house innovation for image refinement.
Core Features
- Tiled VAE: Enables efficient processing of large images, ensuring high-quality outputs.
- Tiled Diffusion for Txt2Img: This feature allows for the generation of ultra-large images from text prompts, making it ideal for creative individuals looking to visualize ideas on a grand scale.
- Tiled Diffusion for Img2Img Upscaling: Enhances image details, allowing users to improve image resolution and clarity significantly.
- Regional Prompt Control: Offers users the ability to apply specific prompts to different image regions for customized outputs.
- Tiled Noise Inversion: Facilitates better noise management ensuring smoother results.
Advanced Features
In addition to core capabilities, the extension supports several advanced features such as:
- ControlNet Compatibility: Integrates with ControlNet for additional functionality.
- StableSR Support: Works seamlessly with the StableSR extension.
- SDXL Compatibility: Although in the experimental phase, this feature is supported.
- Demofusion Support: Another experimental yet powerful addition that enhances the extension's capabilities.
Examples and Usability
The extension is designed with practicality in mind, demonstrated in several examples:
- Txt2Img: Generates large, high-quality images from text, perfect for envisioning cityscapes or scenery.
- Img2Img Upscaling: Shows its prowess in significantly enhancing image detail, taking a low-resolution image and transforming it into a higher-resolution version.
- Regional Prompt Control: Users can see how specific prompts affect different areas of an image, useful in creating complex, multi-character scenes.
Licensing and Community
This extension is licensed under the Creative Commons BY-NC-SA 4.0, allowing free access, use, modification, and redistribution of the tool with the same license. However, from March 28, 2023, versions beyond this date are not to be used for commercial sales, applying specifically to the code and not to derived artworks.
The community is encouraged to contribute, and users can find further documentation and tutorials to aid them in making the most of the features via the project's Wiki.
This project invites users to give it a star if they find it valuable and even support its development through platforms like Ko-fi.