Catalog / LLM routing & guardrails
3-model local image pipeline: Qwen prompt enhancer → Laya style router → Qwen-Image
spoof.sh's ComfyUI workflow hands VRAM between a Qwen3.5-4B prompt enhancer, a Laya style router and Qwen-Image-2.1 on a 12 GB 3080 Ti. ~37 s per image, zero OOM.