小珠光qwenVL

Most of ComfyUI looks at an image and returns a different image. This node (小珠光 qwenVL) looks at an image - and your question about it - and returns text. It runs an actual Qwen3-VL instruct model on your GPU, through HuggingFace transformers, inside the graph. The jobs it's good at are the ones the KB's LLM-in-graph essay lays out: caption an image into a detailed description, analyze composition or style, answer "what's in this frame," then feed that text to a prompt encoder or a video-prompt node. It's a description/understanding model, not a generator.

Mechanically it's the standard transformers Qwen3-VL stack: the model loads from a local directory (the pack looks under ComfyUI/models/LLM/Qwen-VL/), and the node keeps it resident when keep_model_loaded is on so consecutive calls don't reload. You can run the FP8 variants (default) or the full-precision 4B/8B builds, with optional 4-bit/8-bit quantization on top via bitsandbytes for tight VRAM. Image pixels get fed through a min_pixels/max_pixels budget that controls how many vision tokens the image consumes - lower for speed and memory, higher for detail.

Inputs that matter:

• text - your instruction. For captioning: "describe this image in detail." For analysis: ask a question. This is a chat model; prompt it like one.
• model - the checkpoint (Qwen3-VL 4B/8B, Instruct or Thinking, FP8 or full). Thinking variants reason before answering and cost more tokens.
• quantization - none / 4bit / 8bit. This is the VRAM lever. Note the FP8 models are already 8-bit; adding 4bit on top of FP8 is unusual - mostly you pick the FP8 checkpoint as-is, or the full-precision model with a quant.
• keep_model_loaded - leave on unless VRAM is tight.
• temperature - sampling randomness, default 0.7.
• max_tokens - answer length cap, default 1024.
• min_pixels / max_pixels - image token budget (default 20 / 100). Bigger max = sharper vision, fatter VRAM.
• seed - -1 means no fixed seed.
• attention - sdpa (default), flash_attention_2, or eager. SDPA is the safe default; flash-attn is faster on Ampere+ if you've installed it.
• image - the optional IMAGE input. Leave it disconnected and it's a plain text LLM.

One output: result (STRING) - the model's answer.

Two honest caveats. First, the dependency weight: this needs transformers>=4.57.1, qwen-vl-utils, and bitsandbytes for quant - all in the pack's declared deps, and all heavy. It also loads models locally only; there's no HF auto-download path, so the checkpoint must actually sit in models/LLM/Qwen-VL/ or you'll get a load error. Second, the quality reality check from the KB: an 8B Qwen3-VL is a solid local captioner, but it can fumble multi-subject attribution - "who's wearing what" gets mixed up - and the Instruct builds carry the usual assistant-style refusals. Fine for SFW scene description and prompt seeds; don't bet a training dataset's accuracy on it.

Install the pack (ComfyUI Manager → ComfyUI-xiaozhuguang), restart, put a Qwen3-VL checkpoint in ComfyUI/models/LLM/Qwen-VL/, and you've got a private, offline vision model in your workflow. The pack's got a Bilibili tutorial if you want the author's worked example.