Würstchen hits 42x image compression, slashing AI training costs 16-fold

Würstchen, a new diffusion model from the team behind Stable Diffusion, takes image compression to an extreme that researchers previously considered impossible. While most models compress images by 4x to 8x spatially, Würstchen achieves a 42x spatial compression using a two-stage decoder design. Stage A, a VQGAN, and Stage B, a Diffusion Autoencoder, work together to decode a highly compressed latent space back into pixel space. A third model, Stage C, operates as the Prior and learns in that ultra-compressed space. The result: training at 1024×1024 resolution costs a fraction of what it would otherwise.

The efficiency numbers are stark. Würstchen v1 required only 9,000 GPU hours to train at 512×512 resolution, compared to 150,000 GPU hours for Stable Diffusion 1.4. That’s a 16x reduction. Würstchen v2, which scales up to 1536×1536, needed 24,602 GPU hours — still 6x cheaper than SD1.4, which never went beyond 512×512. For researchers without A100s lying around, the model also runs faster and uses less memory than Stable Diffusion XL, especially as batch sizes grow.

Integration with Hugging Face’s Diffusers library means Würstchen works with familiar tooling. It supports PyTorch 2’s SDPA accelerated attention, xFormers flash attention for PyTorch 1.x users, model offloading to CPU, sequential CPU offload for low-memory situations, Apple Silicon MPS devices, and prompt weighting via Compel. The model ships with three Prior checkpoints and one Decoder checkpoint on the Hugging Face Hub.

What makes this genuinely notable is the training cost implications, not just the inference speed. Cutting training from 150,000 to 9,000 GPU hours doesn’t just save money for big labs — it changes who can afford to build custom text-to-image models. A research group or startup that could never justify training a full Stable Diffusion-scale model might now fine-tune Würstchen’s Prior at new resolutions like 2048×2048 without breaking the bank. The Prior adapts to new resolutions extremely fast, according to the team, which opens the door to more specialized image generation work. The catch, as with any extreme compression approach, will be whether image quality holds up across diverse prompts and edge cases. The team acknowledges outputs are best between 1024×1024 and 1536×1536, with some good results at 1024×2048. Beyond that, you’re in experimental territory.