ComfyUI-TypedDecision
ComfyUI nodes for Jev-style noul, choice and score decisions over images and text using a local imajev adapter on Qwen3.5.
Overview
ComfyUI-TypedDecision brings Jev-style typed decisions into ComfyUI without a hosted API. A multimodal model does not write text; it answers one typed question about up to two images plus optional state, returning calibrated probabilities for a noul (true/false), a choice among options, or a score on ordered levels.
It reuses ComfyUI's own multimodal inference and starts with the open imajev-4b adapter on a Qwen3.5 4B backbone. Outputs include the value, label, index, a pass flag against a threshold, and an abstained flag when the inputs are not enough to answer. An optional debias mode averages over several option orders. Example workflows cover counting people, picking the main subject, scoring training-image usability, checking whether two images show the same person, sorting a dataset, and choosing an aspect ratio from a prompt before generation.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Are many people in the image? | true · false |
| Choice | What is the main subject? | — |
| Score | How usable is the photo for training? | — |
How it works
- Load the Qwen3.5 backbone and imajev-4b adapter
- Pick mode (noul/choice/score) and write the question and options
- Pass up to two images and optional state
- Use value/label/pass/abstained outputs to route or save files
Strengths
- Runs locally inside ComfyUI with no API key
- Abstain output when inputs are insufficient
- Optional option-order debiasing
- Multimodal decisions over images
Caveats
- Large model download (several GB backbone)
- Questions must be written in English
- Calibration depends on the imajev adapter
Getting started
Install ComfyUI-TypedDecision from ComfyUI Manager and place the Qwen3.5 text encoder and imajev-4b files under ComfyUI/models.At a glance
- Type
- Project
- Status
- beta
- Language
- Python
- Runtime
- ComfyUI custom nodes
- Has code
- yes
Use cases and integrations
How curated lists describe it
Typed decisions (noul / choice / score) with calibrated probabilities in…
wh000wh000/awesome-jev-live
Frequently asked questions
Does this need TypeSafe?
No. It runs a local open adapter (imajev-4b) on ComfyUI's own Qwen3.5 inference.
What question types are supported?
noul (true/false), choice (one of your options) and score (a position on ordered levels).
Can it decline to answer?
Yes, it reports abstained when the images and state are not enough to answer.
Related projects
Free hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaResearch code for an image-and-text System One model that reads option-letter logits from a multimodal LLM, with training and calibration.
ProjectresearchPythonResearch MVP pairing an LLM-authored, revisable ontology with Jev classification, where low confidence triggers ontology revisions.
ServiceresearchPythonIndependent open model on Molmo2-ER that answers spatial questions by choosing from fixed options with one shared classifier head.
ProjectresearchPythonOpen project training a vision-enabled Jev-like model (0.8B, 9B) that judges dynamic candidate sets over images with Choice, Noul and Score.
ProjectresearchPythonIndependent Jev-style server on DiffusionGemma and SGLang that classifies doodles, flowers and messages with Choice, Score and Noul.
ServiceresearchPython