Eric Qwen Reference Describer (VL→Text)

Most of the time when people say "use this image as a reference," they mean feeding it into a pipeline as pixel conditioning. Eric Qwen Reference Describer takes the opposite approach: it reads your reference images with a vision-language model, converts what it sees into text, and hands you a single prompt you can feed to the pure text-to-image pipeline. No pixels, no conditioning latents - just words.

Why bother? The node's own docstring makes the case: feeding low-res web thumbnails into the edit pipeline as reference images causes pixel contamination. The reference's visual knowledge gets mixed in as actual pixels, muddying the output. But if you extract that knowledge as a clean text description - "subject wearing a red coat, standing in a rain-soaked alley, cinematic lighting" - and let the text-only generation model work from that, you get the style and content guidance without the contamination. It's a translation step, and it's a legitimately smart use of a VLM.

Where it sits in a workflow

This is a specialized node for a specific pipeline, not a casual utility. It's designed to chain after EricGenSearcherNode (the pack's grounded-search node): that node produces a grounded_prompt and hands you reference images, and this node merges them into a final prompt for Eric Qwen-Image Generate. The wiring:

EricGenSearcherNode → grounded_prompt ──┐
                   → ref_image_1..4 ──→ Eric Qwen Reference Describer → combined_prompt → Generate

The required inputs are grounded_prompt, ref_image_1 (the primary reference), and vl_api_url - note the default is http://localhost:30000, the URL of a SGLang VL server running Gen-Searcher-8B (a Qwen3-VL-8B derivative). This isn't a normal Ollama endpoint; you're expected to have that server running, the same one EricGenSearcherNode uses. Up to three more optional references (ref_image_2 through ref_image_4) can be added.

Controls worth knowing

• description_focus - what the VLM prioritizes: auto, full_scene, subject_appearance, clothing_and_style, or setting_and_environment
• synthesis_mode - synthesize (default) has the VLM weave the reference descriptions into the grounded prompt as one coherent paragraph; append just tacks them on, faster but less natural
• model_name - override the VL model, blank = auto-detect
• max_image_dim - reference image size sent to the VLM (512 default is plenty)

Two outputs: combined_prompt (the generation-ready string) and descriptions_log (the raw per-image descriptions, handy for debugging what the VLM actually saw).

Install & gotchas

Install the pack via ComfyUI Manager (search "Eric Qwen-Edit") or clone https://github.com/EricRollei/Eric_Qwen_Edit_Experiments.git into custom_nodes/. The node itself downloads no model - but the SGLang VL server is a hard external dependency, and it's the part that bites. If you just want to describe a reference image without the whole Gen-Searcher stack, this node is overkill and the pack's simpler vision rewriters are friendlier. This one earns its keep only inside the grounded-search workflow where the server's already running.