AI BriefingHigh-Compression-Based Würstchen Released
Key point
Würstchen, a new diffusion model that dramatically reduces training and inference costs through a high spatial compression ratio of 42x, has been released.
Details
Würstchen is a diffusion model whose text-conditional component operates in a highly compressed latent space.
While existing models use 4-8x spatial compression, Würstchen achieves 42x spatial compression, drastically reducing computational costs. The model architecture consists of three stages as follows.
- Stage A & B (Decoder): Combines VQGAN and a Diffusion Autoencoder to restore compressed images back to pixel space.
- Stage C (Prior): A model trained in the highly compressed latent space.
Key Features and Advantages
- Dramatic cost reduction: Würstchen v1 required 9,000 GPU hours for 512x512 resolution training, making it about 16x cheaper than Stable Diffusion 1.4 (150,000 GPU hours).
- Fast inference and low memory: Enables image generation much faster and with less memory than SDXL, resulting in fewer hardware constraints.
- Resolution scalability: Supports training at 1024x1024 and 1536x1536 resolutions, with very low fine-tuning costs for new resolutions.
The model is currently available on Hugging Face Hub and is integrated into the Diffusers library, ready for immediate use.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.