APNext QwenVL Video Analyzer

You've got a video clip and you want to know what's actually in it - to caption it, to turn it into a text prompt, to build an image-to-video workflow that continues from its last frame, or just to index a folder of footage. Sending video to a cloud vision model means uploading frames and paying per call. This node keeps it local: it pulls frames out of the video, hands them to a Qwen3-VL vision model on your own GPU, and returns an analysis string plus the actual extracted_frames it looked at.

It's the video sibling of the QwenVL Vision Analyzer, and it slots into the same local-first story the QwenVL nodes tell: no API key, no per-frame billing, models auto-download to ComfyUI/models/LLM/Qwen-VL/ on first use. If you're already running one QwenVL node in the pack, this one is nearly free to add.

How it works

The node takes a VIDEO input (from a LoadVideo node) - or an IMAGE batch if that's what your pipeline hands you - and does two jobs. First, frame selection: video_section picks start or end of the clip, frame_window says how many frames of that section to consider, and fps_extract decides how many frames to actually pull out for analysis (default 3 - enough to understand a scene, cheap to run). Then the model analyzes those frames with your chosen qwen_model, temperature and max_tokens, and the node returns the write-up plus the frames it used.

Frame decoding is worth knowing about: the pack lists decord as optional-but-commented-out in requirements, because it's unmaintained and not numpy-2 safe. The video nodes fall back to OpenCV automatically - which is to say, it just works, and you don't need to uncomment anything unless you specifically want decord's decoder.

The inputs that matter

• video (optional) - the actual clip. Connect VIDEO from LoadVideo, or an IMAGE batch.
• video_section / frame_window / fps_extract - where in the clip to look, and how much of it to look at. frame_window has effectively no limit, so you can scan a whole feature if you're patient.
• qwen_model - Qwen3-VL 2B–8B, with Thinking and FP8 variants. 2B is fine for description; reach for 8B when you need the model to actually reason about what's happening.
• custom_prompt (optional) - overrides the default analysis instruction when you want something specific ("describe only the lighting and camera movement").
• keep_model_loaded / use_flash_attention / seed / randomize_each_run - the standard QwenVL set; flash attention is off by default and requires the flash-attn package.

Outputs are analysis (the description - wire it into a prompt, a text preview, or a save node) and extracted_frames (the exact frames it looked at, so you can verify it saw what you think it saw).

Install

Standard pack install - ComfyUI Manager (search "comfyui_dagthomas") or:

cd ComfyUI/custom_nodes
git clone https://github.com/dagthomas/comfyui_dagthomas
cd comfyui_dagthomas
pip install -r requirements.txt

Restart, and let the model download on first use. No API keys.

Common issues

• Nothing comes out of LoadVideo - check your loader's output type; the node accepts VIDEO or IMAGE, but the connection has to match one of them.
• Slow on long clips - you're loading and analyzing a model on your own GPU; keep fps_extract low (3–5) and use the 2B model unless you need 8B reasoning.
• Out of memory - FP8 variants exist for exactly this. And don't enable use_flash_attention without flash-attn installed, or the load may fail.