ComfyUI Node
PreFlight Observe
Runs Qwen3.5-9B INT8 ConvRot (downloaded on first use) as a pure observation sensor. It describes what is visually present (garment, exposure, framing, pose, in-image text, ...) as a strict JSON object; it does NOT judge acceptability. Feed the result into PreFlight Report. Deterministic: greedy decoding, so the same image and settings give the same output run to run. Errors never break the graph: a problem comes back as {"error": ...} so Report can fail closed.
PreFlight Observe
- image
- observations_json
- raw_response
◄max_frames6►
◄keep_model_loadedtrue►
CategoryBCNodes/preflight
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| image | IMAGE | A single image or a video frame batch. | |
| max_frames | INT | 61–16 | For a video: how many frames to sample evenly from the batch. All of them go to the model in one call. |
| keep_model_loaded | BOOLEAN | true | On: the model stays loaded for the next run (one LM model is held at a time, shared with Qwen LM). Off: it is unloaded when the run ends. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| observations_json | STRING | — |
| raw_response | STRING | — |