Jev Interpret

What it actually does

Almost every LLM node in ComfyUI does the same trick: take your rough idea, hand back a longer version, feed that to the encoder. You still accept whatever comes out.

Jev Interpret does the opposite half of the job. It doesn't write - it decides. Give it a pile of candidate prompts and an instruction like "choose the one that best fits a warm, approachable luxury brand" and it hands one of your candidates back. Or ask a yes/no question about a string, score something from 0 to 1, or pull the number out of "we shot at 7.5 stops". Not "make my prompt better" but "pick between these four", or "branch if the caption mentions a cat". Diffusion hands you options cheaply; something still has to choose.

How it works

You don't write a JSON schema. The node compiles state, instructions and task into a private question set and posts it to TypeSafe's SystemOne endpoint - Jev isn't a chat model, it's a judgment model that answers structured questions with choices and probabilities. provider defaults to typesafe; switch it to openrouter and the same request goes to OpenRouter's Jev decisions path instead (literally an /alpha/ URL, so don't build a business on it).

The neat part is the binding. For choice, your candidates become the criteria, the API picks one, and the node returns the original text of that candidate, byte for byte - not a paraphrase, not a JSON wrapper, your string. That's why result drops straight into a CLIP Text Encode. Same idea for JSON candidates: the model compares description, and you get the value back.

The inputs that matter

• state - what's being judged. It doesn't take images; caption an image first and paste that in.
• instructions - the judgment itself, in plain language. Write it as a test, not a vibe.
• task - picks the mode. choice returns one candidate's full text; multi_choice returns every candidate that passes, as a JSON array; boolean returns the string true or false and needs no candidates; score treats your candidates as an ordered low-to-high scale and returns 0–1; extract pulls a number out of state and returns its original text; suggest is the two-pass mode below.
• candidates - an autogrow input, one candidate per connection, up to 100. Each candidate is a link, not a text box: STRING outputs from Text nodes.
• candidates_json - the fast path. Connect OpenRouter Text's candidates-mode text here, or paste a JSON array of description / content / value records. Wired candidates come first, then generated ones.
• threshold (0.5), refresh, model, provider, api_key. The advanced knobs matter only for suggest: shortlist_size (3) is how many candidates get their full content read, max_selections (1) caps what comes back, and gate_threshold (0.3) filters out requests that never needed guidance. suggest can legitimately return [], and that's a feature.

Outputs

result is the STRING you'll wire - into CLIP Text Encode, Preview as Text, Compare Text, a Switch, or a Convert Number node when task is score or extract. details carries the judgment plus model and usage, and response_json is the raw API response. Both are inspection outputs; nothing downstream consumes them in a normal graph.

Installing it

Search ComfyUI-Jev in ComfyUI Manager and restart. By hand:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/hndrr/ComfyUI-Jev.git

The pack needs ComfyUI 0.36.0 or later and Python 3.10+. Its dependencies, aiohttp and PyYAML, ship with a standard ComfyUI install, but if the import fails:

python -m pip install "aiohttp>=3.11.8" PyYAML

Windows portable users, use the bundled interpreter:

.\python_embeded\python.exe -m pip install "aiohttp>=3.11.8" PyYAML

Then set a TypeSafe key and start ComfyUI from that terminal, so it inherits the variable:

export TYPESAFE_API_KEY="your-typesafe-key"

$env:TYPESAFE_API_KEY="your-typesafe-key"

Leave api_key empty to use it. Pasting the key into the node works too, but node values are saved into the workflow JSON and the execution history, so strip it before sharing a file.

Common issues

• The nodes don't show up. Usually an older ComfyUI - this pack uses the current node API. Restart and check the startup log for import failed.
• "Provide at least one candidate". You set task = choice with nothing connected. Candidates arrive over links: Text nodes, or a candidates_json array.
• extract errors instead of returning empty. By design - nothing in state answering the instruction is an error, not a blank.
• Running again gives the same answer. ComfyUI cached it. Changing the seed or resolution doesn't invalidate text nodes; bump refresh for a fresh judgment.
• jev-preview fails over OpenRouter. There's no verified alias for it there (only jev-latest and jev-1.13.0 are mapped). Use TypeSafe, or custom with a real model ID.