🔍 V2 Remote Vision Analysis

This is the free-form sibling of the pack's business task runners. Where Remote VLM Business Task gives you a task dropdown (caption, OCR, JSON…), V2 Remote Vision Analysis hands you a blank prompt field and says "ask the model whatever you want." It's the node you reach for when none of the presets fit: "is this chair facing left or right?", "what's the dominant color palette?", "are these two logos the same brand?" - anything that's a question about an image, answered by the remote model you configured.

It takes a REMOTE_VISION_MODEL config from V2 Remote Vision Model Config (or its non-V2 sibling), an optional image, and your prompt. The output is a single description STRING. It's an output node, so the answer renders right in the UI - which makes it an excellent node for just poking at a model interactively while you build a workflow.

The inputs that matter

• model_config - required. The REMOTE_VISION_MODEL from the config node.
• prompt - your question/instruction. Defaults to "Describe this image in detail."
• max_tokens - default 1024, up to 8192. The tooltip notes -1 means unlimited; if your server supports it, that's handy for long outputs.
• temperature - default 0.7 here (note: higher than the task runners' 0.2). For freeform analysis that's reasonable; for anything you'll feed into another step, drop it toward 0.2 for reproducibility.
• timeout - default 300s, range up to 1800. The tooltip is explicit: vision models take a while on images, keep it at 300–600.

How it works

Same plumbing as the task runners: the config's service_available flag gates the call, your image becomes a base64 PNG data URL, and the request goes to the server over the OpenAI-compatible protocol. No local model loading, no llama-cpp-python involved - this node is light, and it's the reason the pack's README recommends remote mode as the default for most people. Your GPU stays free for the diffusion model; the VLM runs in LM Studio/Ollama on the side.

Why you'd pick this over the task runner

Flexibility, mostly. The task runners build your prompt for you and are great when they fit. This node is for the 30% of cases where they don't - including chaining, because the description output is just a STRING you can pipe into a prompt template or a text save node. It's also the better node for interactively tuning a prompt before you commit it to a batch: one image, one question, immediate answer.

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, find it under GGUF-VLM → v2 → Remote.

Troubleshooting

If you get a "service unavailable" message, the server isn't reachable - check the config node's base_url and that the model is loaded (this is the single most common failure). If the answer is truncated, raise max_tokens or check the model's own context window. And if the model ignores your specific question and gives you a generic description, the config node's system_prompt is likely overriding your instruction - a model-persona system prompt can steamroll a weak user prompt.