QwenVL (HF Alt Advanced)

"QwenVL (HF Alt Advanced)" is the tooltipped HF variant of the pack with every parameter pulled out onto the widget. Same Qwen3-VL transformer engine, same auto-download into models/LLM/Qwen-VL, same "local captioning, no API key" pitch - plus temperature, top_p, num_beams, repetition_penalty, frame_count, device, use_torch_compile, and a max token budget that stretches to 4096. If you've decided the plain HF node is your captioner but its fixed internal settings keep missing, this is the version to run.

What the extra knobs do

The sampling row behaves exactly like the other Advanced nodes in this pack, only here the tooltips actually tell you the numbers. temperature: "0.2–0.4 is focused, 0.7+ is creative." top_p: "0.9–0.95 allows more variety." repetition_penalty: "1.1–1.3 penalize repeated phrases." And the sharp edge of the set: num_beams - "values >1 disable temperature/top_p and trade speed for more stable answers." So you can't have both beam stability and sampling creativity at once, which is the single most useful fact about this node, and it's written right there in the hover text.

The runtime row: device for auto/cpu/mps or an explicit GPU, use_torch_compile for a throughput boost after a one-time compile, and frame_count (1–64, default 16) for how many frames get sampled out of a video input before prompting. max_tokens here reaches 4096 - enough for genuinely long analyses - and the rest of the widget is the familiar set: model pick (2B/4B/8B, Instruct/Thinking, FP8 variants), quantization (FP16/8-bit/4-bit), the 🖼️ preset prompt menu, custom_prompt, keep_model_loaded, and seed (which, per the tooltip, governs both sampling and frame picking - reuse it to reproduce an exact run).

The inputs that matter

• num_beams - the fork in the road: 1 for sampled variety, 2–4 for stable answers (and it overrides temperature/top_p while on).
• temperature / top_p - 0.2–0.4 for disciplined training captions, 0.7+ for creative writing.
• frame_count - only for video; more frames = more context, more time.
• use_torch_compile - a real speedup on supported CUDA/Torch 2.1+ after the first compile pass.

Inputs are image and video; output is a single RESPONSE string, and - same as the plain HF node - it is not an output node, so you need a ShowText or the pack's aistudynow_SaveText to see the text in the UI.

Getting it

Identical to the rest of the pack - ComfyUI Manager, search "ComfyUI-QwenVL", or:

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

First run downloads the chosen weights into models/LLM/Qwen-VL; 4/8-bit quantization wants CUDA and bitsandbytes; attention_mode uses flash-attn when present on Linux x86_64 and falls back to SDPA otherwise.

Where people get burned

The beam-vs-sampling trap is #1: crank num_beams to 3, dial temperature to 0.9, get deterministic output, and conclude the node ignores you. It doesn't - the tooltip on num_beams says exactly what's happening. Second, use_torch_compile on its first run looks like a hang while it compiles; give it a minute. Third, video jobs with high frame_count plus a big model can OOM a 12 GB card - that's the moment to drop to 8-bit, the 2B, or fewer frames. And remember the Instruct-model refusal quirk if your captions come back sanitized; abliterated builds exist, though users of this pack report those can be inconsistent, so spot-check before captioning a whole dataset on one. For captioning a LoRA set with tight quality control, this node plus a low temperature and repetition_penalty around 1.2 is a setup you can actually trust.