Custom nodes that let a multimodal LLM make a decision instead of writing text (Jev-style typed decisions). Ask a typed question about images, videos and text, and get the answer with calibrated probabilities.

noul: true or false
choice: one of your options
score: a position on your ordered levels

Inference runs on ComfyUI's own LLM engine (model implementations, loaders, memory management), so it works almost natively. No extra dependencies.

Supported models

| Model | Released | Base | int8 size | Images | Video | | --- | --- | --- | --- | --- | --- | | imajev-4b | 2026-09-24 | Qwen3.5-4B + LoRA | 5.76 GB + 490 MB | up to 2 | × | | Clef-flash | 2026-09-30 | Qwen3.5-9B | 12.2 GB | any number | ○ | | d3-flash | 2026-10-10 | Qwen3.5-9B | 9.88 GB | any number | ○ | | d3-mini | 2026-10-10 | Qwen3.5-4B | 5.52 GB | any number | ○ | | d3-nano | 2026-10-10 | Qwen3.5-2B | 3.06 GB | any number | ○ | | d3-lite | 2026-10-10 | Qwen3.5-0.8B | 1.21 GB | any number | ○ |

More on the blog: 日本語 · English · 中文

Installation

Search for ComfyUI-TypedDecision in ComfyUI Manager and install it.

Models

imajev

Uses Qwen3.5-4B and the imajev-4b folder.

| | int8 ConvRot | bf16 | | --- | --- | --- | | Qwen3.5-4B | qwen3.5_4b_int8_convrot (5.76 GB) | qwen3.5_4b_bf16 (9.32 GB) |

Put the files below from mohit67890/imajev-4b (490 MB) in an imajev-4b folder. Downloading the whole repository also works.

📂ComfyUI/
└── 📂models/
    ├── 📂text_encoders/
    │   └── qwen3.5_4b_int8_convrot.safetensors
    └── 📂typed_decision/
        └── 📂imajev-4b/
            ├── adapter_config.json
            ├── adapter_model.safetensors
            ├── calibration-rot4-modality.json
            ├── decision_readout.json
            └── decision_readout.safetensors

d3 / Clef-flash

One file holds the whole model, decision head included. Put the ones you use in text_encoders.

| | int8 ConvRot | bf16 | | --- | --- | --- | | d3-flash | d3_flash_int8_convrot (9.88 GB) | d3_flash_bf16 (16.8 GB) | | d3-mini | d3_mini_int8_convrot (5.52 GB) | d3_mini_bf16 (9.08 GB) | | d3-nano | d3_nano_int8_convrot (3.06 GB) | d3_nano_bf16 (4.43 GB) | | d3-lite | d3_lite_int8_convrot (1.21 GB) | d3_lite_bf16 (1.71 GB) | | Clef-flash | clef_flash_int8_convrot (12.2 GB) | clef_flash_bf16 (19.1 GB) |

📂ComfyUI/
└── 📂models/
    └── 📂text_encoders/
        ├── d3_mini_int8_convrot.safetensors
        └── clef_flash_int8_convrot.safetensors

Nodes

Load Typed Decision Model

Pick the model as backbone, then the matching adapter.

| Model | backbone | adapter | | --- | --- | --- | | imajev | Qwen3.5 | imajev-4b | | d3 | a d3 file | d3 | | Clef-flash | a Clef-flash file | clef |

The adapter shows that model's settings.

debias (imajev, d3): judge with the options in other orders and average. Slower.
- imajev: every mode, up to 4 orders
- d3: choice only, also in reverse order
fps (d3, Clef-flash): video frames read per second (default 2)
calibration (imajev): apply temperature calibration

Typed Decision

Judges one question. Write questions and options in English.

| Input | Meaning | | --- | --- | | mode | noul / choice / score | | instructions | The question | | criteria | choice options or score levels, one per line | | criteria_true / criteria_false | When noul answers yes / no (optional) | | images | Images to judge. With 2 on imajev, the first is the reference and the second is judged | | videos | Videos to judge (d3, Clef-flash) | | state | Facts to judge against, as text or JSON (optional) | | threshold | Threshold for pass |

| Output | Meaning | | --- | --- | | value | noul: probability of yes<br>choice: probability of the chosen option<br>score: expected level (first line = 0) | | label | The answer (not affected by threshold) | | index | noul: yes = 1<br>choice, score: line number of the answer, from 0 | | pass | true when value >= threshold and the model did not abstain | | abstained | true when the images and state are not enough to answer (imajev only) | | result | Full answer as JSON |

Videos

The whole video is read at the loader's fps and all frames are judged at once. A minute at 2 fps is 120 frames.
d3 slows down sharply on long videos (d3-mini int8, 12 GB: 20 s took 4 s, 70 s took 5 min). Lower fps for long videos.
Clef-flash shares one pixel budget across a video, so longer videos are read smaller.
result lists the frames read and their spacing.

Typed Decision Moment Search

Finds the intervals of a video where a condition holds (e.g. "A dog is visible."). Meant for conditions one frame shows; motion such as direction cannot be judged.

| Input | Meaning | | --- | --- | | instructions, criteria_true, criteria_false, state | As noul in Typed Decision | | images | Reference images (optional), to find the person or thing in them | | video | The video to search | | window | Frames judged together at first (default 8 = 4 seconds) | | threshold | Probability for a frame to pass |

| Output | Meaning | | --- | --- | | intervals | JSON of the intervals: start, end (seconds), start_frame, end_frame (source frames, end exclusive), score (highest probability in the interval) | | videos | The video trimmed to each interval. When nothing is found, the video nodes after it are skipped |

How it works

Reads the video at 2 fps and judges in three steps.

Judge window frames at a time and skip blocks that fail
Judge each frame of the remaining blocks
Check the source frames between passing and failing frames to set the edges

Notes

With reference images, ask about the video. "The video shows the same person as the image." works; "The person in the image is visible." is true of the reference itself, so every frame passes.
A larger window is faster but misses short events more easily (at 16, d3-mini and d3-flash can drop a half-second event). imajev takes up to 2, or 1 with a reference image.
Intervals follow the model's answers. Something half out of frame may already count as gone, so write the question for what you want to count.
Events shorter than half a second can fall between the 2 fps frames.

Sample workflows

noul: are many people in the image?
choice: what is the main subject?
score: how usable is the photo for training?
two images: do two images show the same person?
dataset sort: save photos with and without people under different names
resolution from prompt: pick the aspect ratio from the prompt, then generate with Krea 2
Moment Search: find where the person in the reference image appears with d3-mini, join and save the intervals

Credits

imajev and imajev-4b by mohit67890: model, prompt format and scoring
d3 (Decision 3.0) by the vLLM Semantic Router team: models, prompt format and readout
Clef-flash by Cloudflare: model, prompt format and joint schema head
Qwen3.5 by the Qwen team, repackaged for ComfyUI by Comfy-Org