Wuerstchen v2
A research model featuring a unique 3-stage cascade architecture with 42x latent compression. CLIP-G text encoder feeds into Prior → Decoder → VQ-GAN stages. Developed in 2023, Wuerstchen is no longer actively maintained — its authors went on to create Stable Cascade (also discontinued).
- Developer: Wuerstchen Team
- License: MIT
- HuggingFace: warp-ai/wuerstchen
Recommendation: For most use cases, Flux.2 Klein produces significantly better images at similar or lower VRAM usage in fewer steps. Wuerstchen is best suited for users interested in the cascade architecture or who prefer its natural painterly aesthetic.
Variants
| Model | Steps | Size | Notes |
|---|---|---|---|
wuerstchen-v2:fp16 | 30 | 5.6 GB | Full cascade pipeline |
No quantized (GGUF) variants are available for this model.
Defaults
- Resolution: 1024x1024
- Guidance: 4.0
- Steps: 30
Recommended Dimensions
| Width | Height | Aspect Ratio |
|---|---|---|
| 1024 | 1024 | 1:1 (native) |
Using non-recommended dimensions will trigger a warning. All values must be multiples of 16.
Notes
Wuerstchen produces softer, painterly images compared to FLUX or SDXL. Output quality is lower than other model families — expect less fine detail and occasional anatomical inconsistencies. There is no community ecosystem of LoRA adapters, fine-tunes, or ControlNet support for this model.
Wuerstchen includes a default negative prompt. Negative prompts are supported and effective with this model. The 42x latent compression means the diffusion process operates in a very compact space (24x24 for 1024x1024 output).
img2img and inpainting are supported. ControlNet is not available for this model.
Example
Wuerstchen v2 FP16 — 30 steps, seed 42:
mold run wuerstchen-v2:fp16 \
"A lighthouse on a rocky coast during a dramatic sunset, \
oil painting style, vibrant orange and purple sky" \
--seed 42