Use bf16 text encoder + bf16 stock fallback: fp8 weights bypass ZeroGPU packing (comfy_kitchen QuantizedTensor) and produce NaN

Fix conditioning debug indexing

Add conditioning NaN diagnostics

Fix black output: drop inference_mode (fp8 quantized weights crash on device moves), NaN-safe conversion, debug stats, grad disabled

Fix gr.Examples: 14 values/inputs vs generate_image's 13 args (resolution is UI-only)

load_models_gpu: current UNETLoader returns ModelPatcher directly, use getattr fallback

Fix hf_hub_download local_dir nesting: files were landing in models/<sub>/<sub>/

Add torchvision/torchaudio (import-time deps of comfy.sd); drop platform-pinned packages

Krea 2 Turbo Gradio app: ComfyUI backend, CivitAI PornMaster UNet (V2.5->V2 ladder), ZeroGPU ready

initial commit