| Property | Value |
|---|---|
| Parameter Count | 57,266,643 |
| License | MIT |
| Library | Diffusers |
| PSNR Score | 31.166 |
| LPIPS Score | 0.0198 |
What is vae-kl-f8-d16?
vae-kl-f8-d16 is a lightweight variational autoencoder (VAE) designed for efficient image processing. It features 16 channels with an 8x downsample, trained from scratch on a diverse dataset including photos, artistic works, text, cartoons, and vector images. This model achieves performance metrics comparable to larger VAEs while using significantly fewer parameters.
Implementation Details
The model employs a streamlined architecture with 57.2M parameters, considerably less than standard VAEs like SD3 (83.8M parameters). Despite its smaller size, it achieves impressive metrics with a PSNR of 31.166 and LPIPS of 0.0198, nearly matching SD3's performance.
- Optimized 16-channel architecture
- 8x downsampling capability
- Compatible with Diffusers library
- Requires adapter implementation for SD1.5 usage
Core Capabilities
- Efficient VRAM usage due to lightweight design
- High-quality image reconstruction comparable to larger models
- Versatile application across various image types
- Full compatibility with diffusers pipeline
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for achieving near SD3-level performance metrics while using 31% fewer parameters, resulting in faster processing and lower VRAM usage. Its MIT license also allows unrestricted usage.
Q: What are the recommended use cases?
The model is primarily intended for developers looking to implement a lightweight VAE in their projects. It requires training into a network before practical use and is particularly suitable for SD 1.5, SDXL, and potentially pixart implementations.