QwenVL (Advanced)

Same Qwen-VL engine as the standard node, same models, same image-in-text-out job - but with the generation dials exposed. If the plain QwenVL node is giving you captions that are too samey, too rambling, or you're feeding it video and want control over how many frames it looks at, this is the one to reach for. If you just want "describe this image" and don't care about the mechanics, stick with the standard node; the extra parameters here are rope you don't need for basic captioning.

Why the extra knobs matter

A vision-language model generates its answer token by token, and how it picks each token is a sampling choice. The standard node makes that choice for you. The Advanced node hands it over - which matters when you're doing this at scale (say, captioning a training set) and the default output isn't quite the register you want. Want tighter, more deterministic descriptions? Lower the temperature. Want it to stop repeating itself? Nudge the repetition penalty. Want the single most-likely reading instead of a sampled one? Switch to beam search. It's the difference between accepting the model's default voice and directing it.

The inputs that matter

Everything from the standard node carries over - model_name (the full Qwen3-VL / Qwen2.5-VL lineup), quantization, attention_mode, preset_prompt, custom_prompt, max_tokens, keep_model_loaded, seed, plus the image / video inputs and the RESPONSE output. The additions worth knowing:

• temperature (0.1–1.0, default 0.6) - randomness. Low = focused, repeatable, factual captions. Higher = more varied phrasing. For captioning you usually want it down, around 0.3–0.6, so descriptions stay literal.
• num_beams (1–8, default 1) - beam search. Anything above 1 makes the model explore several candidate continuations and keep the best overall, which trades speed for a more considered answer - and it disables temperature/top_p sampling while it's on. Leave it at 1 unless you specifically want that.
• repetition_penalty (0.5–2.0, default 1.2) - discourages the model from looping the same phrase. If your captions keep restating themselves, raise it a touch.
• frame_count (1–64, default 16) - for the video input, how many frames get sampled and fed to the model. More frames = more temporal detail but more VRAM and time. This is the lever for video summaries.

Two more: device (auto / cpu / mps) lets you pin execution - auto picks your GPU, mps is for Apple Silicon, cpu is the slow last resort. use_torch_compile (off by default) compiles the model for faster repeat inference; it adds a one-time compile cost up front, so it only pays off across many runs.

How to install it

It ships in the same pack as every other QwenVL node. ComfyUI Manager: search ComfyUI-QwenVL and install. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/1038lab/ComfyUI-QwenVL.git
cd ComfyUI-QwenVL
pip install -r requirements.txt

Restart ComfyUI. Models auto-download from HuggingFace to ComfyUI/models/LLM/Qwen-VL/ on first use, so budget for that first slow run.

Common issues

The advanced knobs give you a couple of new ways to shoot yourself in the foot. Beam search plus a big model plus lots of max_tokens multiplies memory and time - if the node crawls or OOMs, drop num_beams back to 1 first. On video, frame_count is a VRAM multiplier: 64 frames of a large model will blow past a consumer card that handles single images fine, so scale it down if things freeze. And the general Qwen-VL rule still holds - a real 4080 Super report had the 4B FP8 model working but anything larger locking up, so if you're stalling, the model is probably too big for your card before any of these settings are the problem. Cold first run erroring on a 401? That's a stale HuggingFace token, not the node - clear it and retry. And if flash_attention_2 throws, fall back to sdpa.