Nodes/Eric_Qwen_Edit_Experiments/Eric Qwen Reference Describer (VL→Text)
ComfyUI Node

Eric Qwen Reference Describer (VL→Text)

A ComfyUI node in Eric/QwenImage with 10 inputs and 2 outputs.

By EricRollei·Created 5 months ago·Updated 3 months ago· 19
Eric Qwen Reference Describer (VL→Text)
  • ref_image_1
  • ref_image_2
  • ref_image_3
  • ref_image_4
  • combined_prompt
  • descriptions_log
grounded_prompt
vl_api_urlhttp://localhost:30000
model_name
description_focusauto
synthesis_modesynthesize
max_image_dim512
CategoryEric/QwenImage

Inputs (10)

NameTypeDefaultDescription
grounded_promptSTRINGGrounded prompt from EricGenSearcherNode.
ref_image_1IMAGEPrimary reference image.
vl_api_urlSTRINGhttp://localhost:30000URL of the SGLang VL server (Gen-Searcher-8B). This is the same server used by EricGenSearcherNode. Default: http://localhost:30000
ref_image_2optIMAGEOptional 2nd reference.
ref_image_3optIMAGEOptional 3rd reference.
ref_image_4optIMAGEOptional 4th reference.
model_nameoptSTRINGVL model override. Leave blank to auto-detect.
description_focusoptCOMBOautoWhat to focus on when describing reference images: auto: let the VL model decide what's most important (recommended) full_scene: subject + clothing + setting + lighting subject_appearance: face, hair, build only clothing_and_style: garments and accessories only setting_and_environment: location and lighting only
synthesis_modeoptCOMBOsynthesizesynthesize (recommended): ask the VL model to weave the descriptions into the grounded prompt as a single paragraph. Produces the most natural and coherent result. append: simply append the descriptions to the grounded prompt. Faster (one fewer API call) but less coherent.
max_image_dimoptINT512256–1024Max dimension when sending reference images to the VL model. 512px is sufficient for the model to extract visual details. Higher = slower API calls, diminishing returns.

Outputs (2)

NameTypeDescription
combined_promptSTRING
descriptions_logSTRING