QwenVL (HF Alt)

"QwenVL (HF Alt)" is the pack's second take on the same trick: a Qwen3-VL vision-language model running through Hugging Face transformers, captioning images and video frames into text right inside ComfyUI. It shares the pack's auto-download into models/LLM/Qwen-VL, the same weight choices, the same local-and-free pitch. The "Alt" is the honest label - the pack ships two transformer-based implementations, and this one is the cleaner, more tooltipped sibling. If you bounced off the main node because you weren't sure what a parameter did, this variant spells it out on every widget.

What's different from the main QwenVL node

Two real differences, then a bunch of cosmetic ones. First, defaults: this node defaults to Qwen3-VL-2B-Instruct in full FP16, where the main node defaults to the 4B in 8-bit. That means it's the lighter, higher-precision starting point - ideal if your card is small and you want maximum quality per megabyte of VRAM. Second, output wiring: the HF node returns RESPONSE and is not an output node, so nothing appears in the UI on its own - you wire it into a ShowText, the pack's aistudynow_SaveText, or your CLIP/conditioning path. The main QwenVL node displays its text inline. Neither is better; they're different shapes for different habits.

And then there's the tooltips, which are the real reason a beginner should start here. Hover quantization and it tells you FP16 "gives the best quality if memory allows; 8-bit suits 8–16 GB GPUs; 4-bit fits 6 GB or lower but is slower." Hover attention_mode and it explains that auto tries flash-attn v2 and falls back to SDPA. Hover seed and it notes that the seed controls "sampling and frame picking; reuse it to reproduce results." The author wrote the docs into the node.

The inputs that matter

• model_name - 2B/4B/8B, Instruct or Thinking, plain or FP8 pre-quantized. The tooltip warns the first run downloads weights, so leave disk space.
• quantization - None (FP16) / 8-bit / 4-bit; match it to your card's VRAM.
• preset_prompt - 🖼️ Tags, Detailed/Ultra Detailed/Cinematic Description, 📹 Video Summary, 🪄 Prompt Refine & Expand, and more.
• custom_prompt - "when filled it completely replaces the preset template."
• max_tokens - 512 default, up to 2048; bigger answers cost more time and memory.
• keep_model_loaded - keeps the model resident so the next prompt skips loading.

Optional image and video sockets feed it media; output is one RESPONSE string.

Getting it

The whole pack installs in one move - ComfyUI Manager, search "ComfyUI-QwenVL", or:

cd ComfyUI/custom_nodes
git clone https://github.com/aistudynow/ComfyUI-QwenVL

then restart. First run pulls the selected weights into models/LLM/Qwen-VL; the 4/8-bit options need CUDA and bitsandbytes (the pack's requirements install it), and attention_mode will use flash-attn if it's present on Linux x86_64 and otherwise fall back to SDPA without complaint.

Where people get burned

The first-run download catches everyone - several GB, and on a slow connection the node sits there looking dead while weights stream in. That's normal; leave it. Second, remember RESPONSE goes nowhere by itself: the "no output" confusion is the top beginner trip on this variant, and it's a two-second fix with a ShowText node. Third, if captions refuse content, that's the Instruct model's fine-tuning talking - the pack's community thread on this exact pack noted it's "too censored" for some jobs, with abliterated builds as the workaround (test those before trusting a dataset to them). And if an 8B in FP16 OOMs your card, the tooltip's own advice applies: drop to 8-bit or 4-bit, or the 2B. This node is where I'd point anyone new to the pack - it's the same engine with the manual written on it.