Stable Cascade

Stable-Casdade

Original repo: https://github.com/Stability-AI/StableCascade

Install: Experimental only

Note: Those steps will be removed for the release version

• install sdnext as usual and start it so all requirements are installed
• Install diffusers dev:

  pip uninstall diffusers pip install git+https://github.com/huggingface/diffusers/

• use SD.Next from dev branch

  git checkout dev

• start with --experimental command line flag

  ./webui --experimental --debug

Use

1. Set your compute precision in Settings -> Compute -> Precision
   to either BF16 (if supported) or FP32 (if not supported)
2. Enable model offloading in Settings -> Diffusers -> Model CPU offload
   without this, stable cascade will use ~16GB of VRAM
3. Select model from Networks -> Models -> Reference
   and it will automatically be downloaded on first use and loaded into SD.Next
   attempting to load a manually downloaded safetensors files is not supported as model requires special handling

Resources

With correct tuning, it is possible to run Stable Cascade on a 8GB VRAM GPU But any real performance is to be desired - 1024x1024 on RTX4090 using BF16 is barely reaching 4 it/s

Params

• Prompt & Negative prompt: as usual
• Width & Height: as usual
• CFG scale: used to condition the prior model, reference value is ~4
• Secondary CFG scale: used to condition decoder model, reference value is ~1
• Steps: used to control number of steps of the prior model
• Refiner steps: used to control number of steps of the decoder model
• Sampler: set to Default before loading a model
  Stable Cascade has its own sampler and results with standard samplers will look suboptimal
  Built-in sampler is DDIM/DDPM based, so if you want to experiment at least use similar sampler

Notes

• Default variation that will be downloaded and loaded is FULL model with BF16 precision
• If model download fails, simply retry it, it will continue from where it left off
• Model consists out of 3 stages split into 2 pipelines which are exected as C -> B -> A:
  • Prior pipeline: 8.9 GB total = 1.3GB TextEncoder + 0.5 ImageEncoder + 7GB Stage C UNet
  • Decoder pipeline: 4.4 GB total = 3.0GB Stage B Decoder + 1.4GB Stage A VQGan VAE

Variations

Overview

Note this is included as reference only as loading different variations is currently not supported

Stable cascade is a 3-stage model split into two pipelines (so-called prior and decoder) and comes into two main variations: Full and Lite
You can select which one to use from Networks -> Models -> Reference

Additionally, each variation comes in 3 different precisions: FP32, BF16, and FP16
Note: FP16 is an unofficial version by @KohakuBlueleaf of the model fixed to work with FP16 and may result in slightly different output

Which precision is going to get loaded depends on:

• your user preference in Settings -> Compute -> Precision
• and GPU compatibility as not all GPUs support all precision types

Sizes

Stage A and auxiliary models sizes are fixed and noted above
Stage B and Stage C models are dependent on the variation and precision used

Variation	Precision	Stage B	Stage C
Full	FP32	6.2GB	14GB
Full	BF16	3.1GB	7GB
Full	FP16	N/A	7GB
Lite	FP32	2.8GB	4GB
Lite	BF16	1.4GB	2GB
Lite	FP16	N/A	N/A