🚀 V2 Remote VLM Task Runner

V2 Remote VLM Task Runner is the same workhorse as Remote VLM Business Task, just filed under the pack's newer v2/Business menu. If you've read that article, you already know this node: one image in, one of five task modes, text out. The "V2" is a migration marker, not a feature upgrade - the class is a straight alias of the original (the source literally subclasses RemoteVLMTaskRunner and only changes the category).

So why mention it separately? Because on the registry and in search, you'll hit both names, and the difference will matter to you exactly once: when you paste in a workflow built by someone using the other ID. Know that both exist, both work the same, and pick one.

The task modes (unchanged)

• Caption - detailed caption template.
• OCR - text extraction template.
• Prompt Reverse - image-to-generation-prompt template.
• JSON - structured output with a default schema (subject, action, scene, style, text, tags).
• Custom - uses your custom_prompt instead of a preset.

The inputs that matter

• model_config - required REMOTE_VISION_MODEL from V2 Remote Vision Model Config. No config, no run.
• task - the five-mode dropdown.
• max_tokens - default 1024, up to 8192.
• temperature - default 0.2. Low on purpose: for structured work you want determinism.
• timeout - default 300s, up to 1800. Vision models are slow; keep it high.
• image (optional) - the IMAGE to analyze.
• custom_prompt (optional) - overrides the task template when filled.

Output is a single result STRING, and the node is an output node, so results render in the UI.

How it works

The config's service_available flag gates the call (dead server → clean "❌ 服务不可用" message). Your image becomes a base64 PNG data URL, and the request goes to your LM Studio / Ollama / llama.cpp / OpenAI-compatible endpoint over the OpenAI-compatible protocol. No local model loading - the "remote" is protocol, not necessarily cloud; it's happy pointing at 127.0.0.1.

The honest take

Between the two identical runners, pick the one whose menu you'll remember. The v2 folder is where this pack is heading, so a fresh workflow built today will age better with the V2 ID - but the old one isn't deprecated, and the pack's README quick-start still shows the non-V2 path. There's no functional reason to care, which is exactly the kind of boring truth worth stating plainly: this is a rename-in-progress, not a divergence.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/walke2019/ComfyUI-GGUF-VLM.git
cd ComfyUI-GGUF-VLM
pip install -r requirements.txt

Restart ComfyUI; it's under GGUF-VLM → v2 → Business. The only dependency it needs is requests from requirements.txt, plus a running server on the other end.

Troubleshooting

Same as the non-V2 runner: "service unavailable" means the server isn't reachable - check the port and that the model is loaded (then hit Refresh on the config node so the model list isn't stale). If you chose Custom and forgot to type a prompt, it falls back to "Describe the input accurately," which is a mild trap. And if the output ignores your task mode, suspect the config node's system prompt overriding your instruction.