Stable Cascade¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
stable-cascade |
stabilityai/stable-cascade | a89f66d |
stable-cascade-nc-community | Verified |
stable-cascade-lite |
stabilityai/stable-cascade | a89f66d |
stable-cascade-nc-community | Verified |
Catalog: wuerstchen ·
Models: stable-cascade, stable-cascade-lite · Task: t2i
Architecture. Three stages work at very different resolutions. Stage C (the prior) denoises a small 16-channel image embedding that is about 42 times smaller than the output, conditioned on the CLIP ViT-bigG text states and pooled text embedding. Stage B (the decoder) denoises a 4-channel latent about 10.7 times smaller than the output, conditioned on the stage C embedding and the same pooled text embedding. Stage A, a Paella VQGAN, decodes that latent to pixels. Stages B and C use the same U-Net module with a different block layout, and both use a DDPM Wuerstchen scheduler with a cosine schedule.
| Model | Stage C (prior) | Stage B (decoder) |
|---|---|---|
stable-cascade |
prior, 3.6B parameters |
decoder, 1.5B parameters |
stable-cascade-lite |
prior_lite, 1B parameters |
decoder_lite, 700M parameters |
Usage.
from imagehub import pipeline
with pipeline(
"t2i",
backend="stable-cascade-lite",
model="checkpoints/stable-cascade-lite",
device="mps",
) as pipe:
result = pipe("a cat, oil painting", seed=42)
result.images[0].save("cat.png")
Settings. The prior defaults to 60 steps and guidance 4.0; the decoder defaults to 12 steps,
guidance 0.0 and 512×512 output. The supported options are steps, guidance, prior_steps,
prior_guidance, width, height, seed and negative_prompt. Width and height must be
multiples of 8. Decoder guidance above 1 needs the negative embeddings, so it also requires
prior_guidance above 1. Unrecognized settings raise an error.
Memory. In float32 the weights take about 14.4 GB for the stage C prior (4.1 GB for lite), 6.3 GB for the stage B decoder (2.8 GB for lite), 2.8 GB for the text encoder and 0.07 GB for stage A. The full pair therefore needs about 24 GB, and the lite pair about 10 GB.
Seed compatibility. On CPU, the same seed and settings give the same image as diffusers
StableCascadeCombinedPipeline with generator=torch.Generator("cpu").manual_seed(seed)
(validation). Both stages draw from the one generator. MPS and CUDA runs are smoke
tests only; there is no claim of exact equality on these devices.
Out of scope. Only text-to-image is supported. Image variation (the prior's CLIP image encoder), ControlNet, the stage A encoder, the previewer and fp16/bf16 parity are not implemented. The decoder repository's own text encoder is not used: the combined pipeline passes the prior's text embeddings to the decoder, and so does imagehub.
License. The weights use the Stable Cascade Non-Commercial Community License
(stable-cascade-nc-community), which prohibits commercial use; see the license. imagehub does not
distribute weights.