Tutorial for using stable-cascade
Install the latest version of Comyfui
Place the stage_b and stage_c models from https://huggingface.co/stabilityai/stable-cascade/tree/main under ComfyUI/models/unet
Place the stage_a model from https://huggingface.co/stabilityai/stable-cascade/tree/main
Place the clip model from https://huggingface.co/stabilityai/stable-cascade/tree/main/text_encoder into ComfyUI/models/clip
Explanation:
Stage_b and stage_c can be combined differently depending on the VRAM available. The combinations below are listed from the most to the least VRAM consumption:
- stage_b.safetensors + stage_c.safetensors
- stage_b_bf16.safetensors + stage_c_bf16.safetensors
- stage_b_lite.safetensors + stage_c_lite.safetensors
- stage_b_lite_bf16.safetensors + stage_c_lite_bf16.safetensors
Comyfui Workflow
stable_cascade_workflow_test.json
Training tutorial for stable-cascade
Kohya_ss now supports early training of stable-cascade
https://github.com/bmaltais/kohya_ss/tree/stable-cascade
Training example:
https://github.com/bmaltais/kohya_ss/tree/stable-cascade/examples/stable_cascade
Introduction to stable-cascade
Stable Cascade is an innovative text-to-image model built upon the Würstchen architecture. Its notable feature is the three-stage approach, which not only achieves new heights in image quality, flexibility, and fine-tuning capabilities but also significantly lowers hardware requirements. This makes training and fine-tuning on standard consumer-grade hardware straightforward and accessible.