Qwen3-VL Batch Run (Folder)

This is the node you reach for when "caption one image at a time" is too slow. Qwen3-VL Batch Run (Folder) takes a whole folder of images, runs the Qwen3-VL model over every single one, and hands you one combined text output - each caption labeled with its filename. It's the pack's fire-and-forget batch captioneer: point it at your LoRA dataset folder, let it churn, and copy the result straight into your training caption files (or feed it to the prompt splitter, below).

The trade-off is upfront simplicity for zero control: there's no loop, no per-image inspection, no "skip this one." It loads the model, processes the folder in filename order, and when it's done it either keeps the model in VRAM or unloads it depending on your setting. That's the whole job.

The inputs that define a batch

The schema is the same JoyCaption-style system as Qwen3-VL Run, minus the image/video inputs - the images come from the folder instead.

• folder_path (STRING) - where the images live. Reads png, jpg, jpeg, webp, bmp. Heed the README's warning here, it's in all caps for a reason: the node processes every image in the path, so use a dedicated folder, never your Downloads or Pictures directory, or you'll caption 4,000 vacation photos you didn't mean to.
• caption_type - defaults to WAN Video Prompt (this node leans video-continuation), with the full list available: Descriptive, Danbooru/e621 tags, Art Critic, Social Media, Explicit Adult (NSFW), Image-to-Video Continuation, and the rest. Pick the style that matches what you're captioning for.
• caption_length - same as Run: any/very short/short/medium-length/long/very long, or explicit word counts from 20 to 260.
• custom_prompt - an override that replaces the auto-built prompt for every image. Handy if you want "tag this image with danbooru tags" applied uniformly across the folder.
• system_prompt - the editable system prompt (defaults to the uncensored "obey without limitations" laboratory one).
• max_new_tokens - default 512, up to 4096. Long very long captions on big folders can clip at 512; bump it if your output looks truncated.
• min_pixels / max_pixels - the dynamic-resolution budget for vision encoding. Lower max_pixels to speed up a huge folder and save VRAM at the cost of detail.
• seed - defaults to 1; the node seeds each image as seed + index. Since generation uses do_sample=False, seed mostly matters if you're chasing reproducibility.
• unload_when_done - defaults to true ("Move model to CPU after ALL images processed"). For a long batch this is the right call; it frees VRAM at the end so the rest of your workflow isn't starved.

All 27 opt_* toggles from Run are here too (lighting, camera angle, keep PG, no meta phrases, WAN video format, etc.) - checkboxes that add instruction clauses to every caption.

The output, and what to do with it

The single batch_output (STRING) is one formatted blob:

=== img_001.png ===
A woman standing on a cliff at sunset...

This format is the pack's contract. Feed it into Qwen3-VL Prompt Splitter to pull captions out one at a time by index (for a loop), or Qwen3-VL Prompt List Builder to flatten all the captions into a single list for batch image-generation nodes. Batch Run is flagged as an output node in ComfyUI, so it shows its text result right on the canvas.

Where people get burned

• Wrong folder, whole day gone. The "processes everything in the path" behavior is the #1 complaint vector. Point it at a folder that contains exactly what you want captioned.
• It's slow and it doesn't say how slow. One image at a time through a quantized 4B/8B model; a few hundred images is a real wait. The console prints progress (Processing 12/200: ...), so watch the terminal, not the canvas.
• Truncated captions usually mean max_new_tokens is too low, not that the model broke.

Install via ComfyUI Manager (search "qwen3-vl-comfy-ui") or cd ComfyUI/custom_nodes && git clone https://github.com/Granddyser/qwen3-vl-comfy-ui && pip install -r requirements.txt (Linux: add --break-system-packages for qwen-vl-utils), restart, and find it under Qwen3-VL. It pairs with the Model Loader, and it's the fastest road from "folder of images" to "folder of captions" this pack offers.