Llama.cpp Media Diagnostics

When a llama.cpp generation quietly ignores your image, the hardest part isn't the fix - it's proving the image wasn't ingested. Llama.cpp Media Diagnostics is the node for that exact moment. The llama.cpp Generate nodes emit a typed media diagnostics output that records whether the fork's multimodal ingestion actually accepted your media; this node expands that opaque receipt into booleans, counts, and text you can read at a glance or wire into logic.

The input is a single socket, media_diagnostics, which comes from the media diagnostics output on Llama.cpp Generate, Profiled Generate, or Sequential Generate. Feed it in and you get nine outputs:

• All Media Evaluated (BOOLEAN) - did everything you connected get evaluated, or was something skipped?
• Vision Available / Audio Available / Video Available (BOOLEANs) - capability flags for the installed wheel and projector combination. This is where you learn that your wheel was built without MTMD_VIDEO support, or that your mmproj doesn't do audio.
• AUDIO_COUNT, IMAGE_COUNT, VIDEO_COUNT (INTs) - how many of each item actually made it into the request.
• JSON (STRING) - the raw receipt.
• FormattedText (STRING) - the human-readable version, for a text output node.

The failure modes this surfaces are the classic "it didn't do anything and I don't know why" ones. If Vision Available is false but you connected images, no amount of prompt tuning will help - the wheel or projector can't see. If the count is zero but you connected a list, the flattening didn't do what you expected. If All Media Evaluated is false, something was deliberately skipped rather than mis-handled, and the formatted text tells you what. This is the "read the receipt" node that turns a mystery into a checklist.

Wiring is trivial: generate → media diagnostics → this node → a ShowText or a few conditionals. It's a debug utility, not a pipeline component, so it earns its keep on the day things break, not on the day they work.

Install is the pack standard:

cd ComfyUI/custom_nodes
git clone https://github.com/craftingmod/ComfyUI-Ollama-ImageList

Restart ComfyUI. Same shared prerequisites as every llama.cpp node here: the JamePeng multimodal llama-cpp-python wheel (with MTMD_VIDEO for video) in ComfyUI's Python environment, and ComfyUI 0.19.3+.

Honest verdict: you don't need this node most of the time, and you'll be very glad it exists the other 5% of the time. The count and capability outputs double as a sanity check when you're setting up a new mmproj - connect it once after a successful run, confirm the numbers match what you fed in, and you'll never wonder again whether your 12 images actually went over the wire or got silently flattened into one.