🍭大炮-Qwen3VL@炮老师的小课堂

You want Alibaba's Qwen3-VL reading your images locally, inside the graph, and you don't want to babysit a second program or pay for an API call. This node is that. Qwen3VL_Advanced is the flagship of the 大炮-Qwen3VL pack (from "炮老师的小课堂", a Chinese community author) - one node that reverse-prompts single or multiple images, understands a video as a frame sequence, and hands you a ready-to-use text prompt. The name is not a lie: it runs the model on your own GPU, needs no key, and makes no network calls after the weights are downloaded.

It's a "VLM as a tool in the graph" - bolted onto your workflow rather than part of a diffusion checkpoint. Reach for it to feed a generated image back into a prompt, build captions for a LoRA dataset, or turn a reference image into a structured txt2img prompt for a Qwen3-class encoder. The community pattern is to chain it - VLM describes, LLM rewrites, sampler generates - because two tightly-scoped steps drift less than one open-ended one.

How it works

Under the hood it's plain transformers: AutoModelForImageTextToText plus the Qwen processor and tokenizer, with bitsandbytes handling the 4-bit/8-bit quantization if you ask for it. On first run the node checks ComfyUI/models/prompt_generator/, and if the model isn't there it downloads it - from Hugging Face for the official Qwen models, from ModelScope for the two community (abliterated) ones, which need modelscope installed. The default Qwen3-VL-4B-Instruct runs in about 6GB of VRAM at full precision, ~3.5GB in 8-bit, ~2GB in 4-bit, so it fits a consumer card.

Two design choices are worth knowing. The "保持模型加载" (keep model loaded) toggle defaults to off - the model unloads and frees VRAM after every run, so a caption node doesn't squat on your card between generations. And a seed system (random / fixed / increment) is wired through IS_CHANGED so random mode reruns every time instead of getting cached. That's deliberate, but it's also one reason a workflow with this node re-executes more than you'd expect.

The inputs that matter

• 🤖 模型选择 - 20 options across Qwen3-VL 2B/4B/8B/32B (Instruct and Thinking variants plus FP8), Qwen2.5-VL for backwards compat, and two Huihui abliterated models. Start at 4B.
• ⚙️ 量化级别 - None (FP16) / 8-bit / 4-bit. The VRAM numbers above tell you which one your card wants.
• 🖼️ 最大长边 (default 768) - images are downscaled so the long edge fits this before hitting the vision encoder. This is your OOM lever: big images cause a VRAM spike, and 768 keeps it tame.
• 💭 预设提示词 - five captioning personalities: tag list, simple, detailed, ultra-detailed, cinematic. These become the model's instruction, so pick by what your downstream base wants. ✏️ 自定义提示词 overrides them.
• 🔢 最大令牌数, 🌡️ 采样温度, 🎯 核采样参数, 🔍 束搜索数量, 🚫 重复惩罚 - generation dials. Leave them unless output gets short, rambly, or repetitive.
• 🎬 视频帧数 - how many evenly-spaced frames get sampled from the video input (1–64).
• 🎲 随机种子 + 🎯 种子控制, 🚀 开启TF32加速 (Ampere+ only, notably faster), 🔄 保持模型加载, 🧪 性能诊断 (prints an environment and inference report to the console).

Optional inputs: 🖼️ 图像1–图像4 (wire up to four images for comparison) and 🎥 视频 (an IMAGE batch of frames), plus a 🎯 Qwen3VL额外选项 socket for the pack's option node.

The single output, 文本输出, is a STRING - wire it into a CLIP text encode, a prompt node, a save-text node, or anything that eats text. That's the whole loop.

Installing

ComfyUI Manager is the easy path - search for ComfyUI-Qwen3VL-DP. Otherwise:

cd ComfyUI/custom_nodes
git clone https://github.com/paolaoshi/ComfyUI-Qwen3VL-DP
cd ComfyUI-Qwen3VL-DP
pip install -r requirements.txt

Then restart ComfyUI. The dependency that bites: transformers >= 4.57.0. Qwen3-VL's architecture wasn't in older transformers; the classic symptom is model type 'qwen3_vl' but Transformers does not recognize this architecture. Upgrade and it goes away. The FP8 model variants also hard-require a GPU at compute capability 8.9 or higher (a 4090-class card); on anything older the node raises an error, so just pick a non-FP8 entry.

First run downloads several GB - the 4B weights are ~8GB on disk - so be patient and watch the console. If you already have a Qwen3-VL snapshot, dropping it into ComfyUI/models/prompt_generator/Qwen3-VL-4B-Instruct/ skips the download.

One honest caveat before you trust any VLM node with your machine: this category was weaponized once (the ComfyUI_LLMVISION malware), and this pack is small and Chinese-language-first. Glance at the code before your first run - it's short and readable. Fine to use; just don't install blind.