I haven't used llama.cpp for image generation either but I recall an issue about it. Unfortunately I can't pinpoint it now and there is the older closed issue https://github.com/ggml-org/llama.cpp/issues/4408 so unless mtmd supports also multimodal outputs out of the box safe to assume output is still limited to text.