CivitAI Meta Checker

Every image ComfyUI saves carries its whole recipe with it - the full node graph gets baked right into the PNG's metadata, which is genuinely one of Comfy's best features: drag the image back in later and you get the exact workflow back. But that also means you're publishing whatever's sitting in there, sight unseen, the moment you hit save. CivitAI Meta Checker is the "let me actually look at that first" node - a plain diagnostic that dumps the exact prompt graph and extra metadata ComfyUI is about to embed, straight into the node itself, so you can read it before you commit to it.

What it's actually for

Despite the name, it doesn't call CivitAI, doesn't validate anything against CivitAI's API, and there's no pass/fail. The node's own docstring is honest about the ambition - it aims to help you verify that CivitAI might find the expected information, nothing stronger - and the implementation backs that up: it's a dump, not a checker. Wire it in, run your graph, and the node's body fills in with the raw prompt (the API-format node graph for this run) and extra_pnginfo (the extra workflow data ComfyUI tucks in alongside it), rendered as text. You read it and judge for yourself whether the resource hashes, prompt text, and settings you expect are actually in there.

That's most useful paired with RK_CivitAIAddHashes from the same pack: run AddHashes first to inject checkpoint/LoRA/VAE hashes into the graph, then run this one downstream to confirm the hashes actually landed before you upload anywhere. It's also just handy on its own if you're cautious about what's riding along in extra_pnginfo - notes, absolute file paths, anything typed into a text node - before you post an image publicly.

The inputs and outputs that matter

One input, no outputs. image is required - feed it whatever's about to get saved. There's nothing to route onward: this node returns nothing at all, so it's a dead end in your graph, not a passthrough. Treat it like a probe you tap off your image line in parallel with your real Save Image node, not something you chain in series with everything downstream.

How it works

Same trick as its sibling RK_CivitAIAddHashes: ComfyUI hands every node a hidden prompt and extra_pnginfo object for the current run, if the node asks for them. This one asks, and instead of modifying anything, just hands both straight back as UI output - ComfyUI renders whatever a node returns under its UI payload directly in the node's body on the canvas. So what you're looking at is exactly the object that would get serialized into the PNG's metadata if a Save Image node ran right now with that same data.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/rklaffehn/rk-comfy-nodes

Restart ComfyUI - that's it. Zero dependencies beyond the Python standard library and ComfyUI's own internals, and no models to fetch. One thing worth knowing before you go looking: as of writing, the pack's GitHub repo doesn't resolve anymore, and it's not currently showing up in ComfyUI Manager's node list either. The author's other public repos are still there, just not this one - it looks pulled or abandoned rather than renamed. If you can't find or clone it, that's the likely reason, not something wrong on your end.

Where people get tripped up

The name oversells it a little - there's no actual check happening, no red/green indicator, no comparison against what CivitAI would really parse on upload. It just shows you the raw JSON; you're the one doing the checking. Go in expecting a pass/fail badge and you'll be disappointed.

Second: because it returns nothing, don't try to chain anything after it or use it to gate the rest of your workflow. It has to sit somewhere your image already flows through - wire the same image you're about to save into this node alongside your real Save Image node, run once, read the dump, then save for real once it looks right. And remember it's strictly read-only: if the hashes you expected aren't in the dump, this node isn't what's missing them - go check whether RK_CivitAIAddHashes (or whatever's supposed to inject them) actually ran first and succeeded.