Qwen3-VL Model Loader

Every workflow in this pack starts the same way: drop in Qwen3-VL Model Loader, pick a model, wire its model output into a Run node, and hope you've got the VRAM. If you've used a LoRA loader or a checkpoint loader, the shape is familiar - this is the "which weights" node, and it's the only one in the pack that touches the disk for the actual model. It loads a Qwen3-VL vision-language model via Hugging Face Transformers, quantizes it to fit your card, and hands the whole thing downstream as a single QWEN3_VL_MODEL object that the Run, Batch Run, and Loop Run nodes all accept.

What Qwen3-VL is worth knowing before you click: it's Alibaba's open-weights vision-language model (Apache 2.0), and in this community it's basically infrastructure - the default local captioner for building LoRA and fine-tuning datasets, the model people run instead of paying for Gemini/GPT captions. It understands images and video and writes text; it does not generate images. So this loader is the "understanding" side of your stack, not a checkpoint for sampling.

The three inputs that matter

• model - a dropdown with 10 presets: the official Qwen Qwen3-VL-2B/4B/8B-Instruct and -Thinking builds, plus a set of prithivMLmods abliterated (uncensored) versions. The Instruct models carry some assistant-style refusals on explicit content; if your captioning is NSFW, you want the abliterated builds. For everything else, Qwen3-VL-4B-Instruct is the sane default - it runs on a modest card (roughly 8-9GB unquantized) and captions well.
• quantization - none, 4bit, or 8bit (default 8bit). This is BitsAndBytesConfig under the hood, so 8-bit cuts the model's VRAM roughly in half, and 4-bit cuts it further. Start at 8bit; drop to 4bit if you're tight, move to none if you have headroom and want max quality.
• attention - flash_attention_2, sdpa, or eager (default sdpa). sdpa is the safe modern choice and works out of the box. flash_attention_2 is faster but needs the flash-attn package installed separately or the load will fail; eager is the slow fallback.

The one output and where it goes

The single model output (type QWEN3_VL_MODEL) feeds straight into Qwen3-VL Run, Qwen3-VL Batch Run (Folder), or Qwen3-VL Loop Run. There's nothing else to it - one loader, wired to whatever flavor of Run you're using.

How the download actually works

Here's the part that trips people up. On first use, the node auto-downloads the model into ComfyUI/models/VLM/<model-name> via snapshot_download - no manual model fetching, but it's a several-GB download on your first queue, so don't panic when nothing happens for a while. The loader scans ComfyUI/models/VLM for local model directories at startup, which means you can also download a full model folder manually (the README says: put the whole model directory under ComfyUI/models/VLM) and it'll show up in the dropdown automatically.

Gotchas

• Thinking models add reasoning. The Thinking variants emit internal reasoning before answering; the pack strips the </think> block, and the README itself recommends the non-reasoning model for captioning. Unless you specifically want reasoning, pick the plain Instruct build.
• 8-bit needs bitsandbytes. BitsAndBytesConfig requires the bitsandbytes package; if your load errors on 4bit/8bit, that's the missing piece, and 8bit also needs accelerate for device_map="auto".
• VRAM math: 4B unquantized ≈ 8.8GB. On a 12GB card, 8-bit 4B is comfortable; 8B wants quantization and a bigger card.

Install the pack once via ComfyUI Manager (search "qwen3-vl-comfy-ui") or cd ComfyUI/custom_nodes && git clone https://github.com/Granddyser/qwen3-vl-comfy-ui, then pip install -r requirements.txt (on Linux, add --break-system-packages for qwen-vl-utils). Restart ComfyUI and you'll find all the Qwen3-VL nodes under the Qwen3-VL category.