Evaluater Node (Creepybits)

Evaluater Node is one of those ComfyUI nodes where the name is a promise the node itself can't keep. It doesn't evaluate anything. It's a glorified text loader - a dropdown that pulls a pre-written, very opinionated "AI image critic" instruction file out of the pack and hands it to you as a string. The actual judging happens later, when you feed that string to an LLM node like the pack's Gemini 2.5 Unified API.

That sounds underwhelming, but it's the whole design. Creepybits calls these "Creative Council" nodes: little specialist personas you slot into a prompt-generation workflow. You use Master Key to capture a raw idea, pipe it into Gemini, and let Gemini write the final prompt. The Evaluater sits on the critique side - you send it an image you generated, and its canned instructions turn the LLM into a hard-nosed art director who scores your image 1-10 across aesthetic and technical criteria.

How it works

At startup the node reads evaluate_img.json in the pack's assets folder, which lists the available prompt files. Your prompt_file dropdown is populated from that list:

• evaluate_img.txt - the standard critic
• evaluate_img_long.txt - a longer, pickier brief
• assistant_prompt.txt and summary.txt - other personas you can drop in

When you run it, it loads the selected .txt file from assets/prompts and outputs it as text. There's also an optional_input wildcard socket, but the shipped code accepts it and then ignores it - don't wire anything in expecting it to change the output.

The two inputs that matter

Honestly there's only one: prompt_file. Pick your critic, run the node, and take the text output straight into the system_prompt input of your Gemini node. Add your actual image to Gemini's image input, and you've got a self-rating loop: generate, evaluate, adjust the prompt, regenerate.

Where people get burned: because this node outputs only a prompt string, people expect it to produce a score by itself. It won't. If you connect Evaluater's output straight to a KSampler, you've just written a long essay into your positive prompt. The whole thing only makes sense with an LLM downstream - no LLM, no evaluation.

Installing it

This ships in the Creepybits pack, so install is the same as every other node in it:

• ComfyUI Manager: search for "Creepy_nodes" (or "Creepybits") and install.
• Manual:

  cd ComfyUI/custom_nodes
  git clone https://github.com/Creepybits/ComfyUI-Creepy_nodes.git
  

Restart ComfyUI. No model downloads, no heavy deps for this node specifically - the only thing that needs installing is google-genai, which your Gemini node pulls in via the pack's requirements.txt.

Troubleshooting

The two failure modes are both "the file isn't where I expected." If your dropdown comes up empty, evaluate_img.json didn't load - check it's still in the pack's assets folder. If you see ERROR: Prompt file not found. in the output, you've deleted or renamed a prompt file. Both are fixable by restoring the file from the repo. And if you want a different flavor of critic, you can just edit the .txt files - that's the actual point of loading them from disk instead of hardcoding the text.