Eric Reference Describer (VL→Text)

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

By EricRollei·Created 3 months ago·Updated 3 months ago· 0

◄grounded_prompt►

◄vl_api_urlhttp://localhost:30000►

◄model_name►

◄description_focusauto►

◄synthesis_modesynthesize►

◄max_image_dim512►

CategoryEric/VisualResearch

Inputs (10)

Name	Type	Default	Description
grounded_prompt	STRING		Base prompt - typically from EricGenSearcherNode.
ref_image_1	IMAGE		Primary reference image.
vl_api_url	STRING	http://localhost:30000	OpenAI-compatible API URL of a VL model server. Default: http://localhost:30000 (SGLang with Gen-Searcher-8B) Also works with: LM Studio, vLLM, or any OpenAI-compatible endpoint that supports vision/image inputs.
ref_image_2opt	IMAGE		Optional 2nd reference.
ref_image_3opt	IMAGE		Optional 3rd reference.
ref_image_4opt	IMAGE		Optional 4th reference.
model_nameopt	STRING		Model name override. Leave blank to auto-detect.
description_focusopt	COMBO	auto	What aspect of each reference image to focus on: auto - model decides 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_modeopt	COMBO	synthesize	synthesize (recommended): The VL model weaves descriptions into the grounded prompt as a single coherent paragraph. One extra API call. append: Descriptions are appended below the grounded prompt. Faster, no extra API call, but less coherent.
max_image_dimopt	INT	512256–1024	Max image dimension when sending to the VL model. 512px is sufficient - VL models don't need full resolution to understand content. Higher = slower, diminishing returns.

Name	Type	Description
combined_prompt	STRING	—
descriptions_log	STRING	—