MinusZone - deprecated - ImageInterrogator(LLava)

deprecated - ImageInterrogator(LLava) is the v1 vision node of this pack: feed it an image, a local LLaVA model describes it as a text prompt, and with a clip wired in you get conditioning out - the pack's CLIP-Interrogator-style feature, before the current CLIPTextEncode(ImageInterrogator) redesign. It still loads in ComfyUI, it still works, and you should still use the new one.

The v1 framing shows in the inputs: the vision model is a hardcoded dropdown of classic LLaVA GGUFs - llava-1.6-mistral-7b (default), llava-v1.6-vicuna-13b, ggml_llava-v1.5-7b, ggml_bakllava-1 - plus a matched mmproj_model dropdown and a download_source selector (none / modelscope / hf-mirror.com). No "use whatever GGUF I have" option, no auto mmproj matching; you pick both files by hand from the preset list. resolution (default 512) controls the resize before the model sees the image, and sd_format (default none, option v1) is a vestigial SD 1.5-era flag the README screenshots show enabled.

The inputs

llama_cpp_model, mmproj_model, download_source, resolution, sd_format, keep_device, seed. Optional: image, clip, llama_cpp_options, customize_instruct, captioner_config. Outputs: text and conditioning.

Why it's deprecated

The v1 design - preset-only models, manual mmproj pairing, no automatic download - is strictly worse than the current stack. The modern CLIPTextEncode(ImageInterrogator) + ModelConfigManualSelect(ImageInterrogator) / ModelConfigDownloaderSelect(ImageInterrogator) nodes give you auto mmproj matching by SHA256, a broader curated list (including int4 and Llama-3 vision models), and post_processing control. Nothing here is worth inheriting into a new workflow. Legacy graphs import fine; new ones shouldn't use it.

Installing

It's in the pack's v1 subpackage, loaded on startup:

cd ComfyUI/custom_nodes
git clone https://github.com/MinusZoneAI/ComfyUI-Prompt-MZ

Restart. Needs llama-cpp-python + diskcache; selected vision models download into ComfyUI/models/gguf on first run.

Troubleshooting

The FAQ trio dominates: upgrade llama-cpp-python if LLAMA_SPLIT_MODE_LAYER is missing, use CUDA 12.1 PyTorch if LLama.dll fails to load, and fix the network if the pip wheel build fails. The v1-specific trap is model pairing: pick a base model and its matching mmproj as a set (e.g. both llava-1.6-mistral-7b files), because a mismatched projector gives garbage descriptions rather than an error. And quality expectations: Q5-K-M 7B models describe images competently but tersely; if the descriptions feel thin, that's the model tier, not a bug - the current nodes let you reach for better or lighter options.