Multimodal — LLM (local first: GGUF + optional transformers)

**GGUF:** set **GGUF main + mmproj** (.gguf files under models/llm). **Use vision** ON + IMAGE → vision run; **OFF** → text-only (linked image is ignored — Comfy cannot hide the socket, but the model never sees pixels). **Image-only:** vision ON + IMAGE + empty text fields — GGUF sends image only; HF VLM may use a single `.` token only so the HF API accepts the call (not a written instruction). **combined_model** = transformers only (HF id / snapshot folder). Leave **empty** for GGUF-only (including text-only on GGUF). **load_source** applies only to **combined_model**.