🖼️ Remote Vision Analysis

Every pack's vision pipeline has two halves: the loader that sets up the model and the node that actually looks at your image. This is the second half for remote mode - the part where you hand a picture to whatever vision model is running in LM Studio, Ollama, or Nexa and get a text description back. It's the output end of the 🌐 Remote Vision Model Config → 🔍 Remote Vision Analysis chain, and it's also where you tune how the model talks.

What it takes in

The required inputs are the model config and the knobs:

• model_config (REMOTE_VISION_MODEL) - the output of the Remote Vision Model Config node. Nothing happens without it.
• prompt - your instruction, defaulting to "Describe this image in detail." This is the input you'll actually edit.
• max_tokens - default 1024, capped at 8192. The tooltip notes -1 means unlimited, though the widget range starts at 1, so you'll mostly just leave it.
• temperature - default 0.7. Lower it toward 0.1 for consistent, factual captions; raise it for creative description.
• timeout - default 300 seconds, and the tooltip's advice is worth taking literally: vision models process images slowly, and the author recommends 300–600 seconds. If your analysis keeps erroring out, this is the first thing to check - a slow VLM will happily eat three minutes.

The optional image input is exactly that - optional. The node supports pure-text chat too, which is handy for testing a server without an image or using the vision model as a plain LLM.

What comes out

One output: description (STRING). It's an output node, so it displays its result in the UI - but the string also flows onward, which is the interesting part. Wire it into a Save Text node, feed it into a prompt-builder, or let it drive a second pass of generation. If you're building an auto-captioning loop for a training set, this is where captions come from.

Setup

Install the pack and you're done - no CUDA llama-cpp build needed, because the model runs in the external server:

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

The gotchas that waste people's time

Three, mostly server-side. If you get nothing back, check that the server actually has a vision model loaded - a text-only model behind the same port will produce garbage or empty output. If it times out, raise timeout rather than assuming a crash; the default exists because VLMs genuinely are this slow. And remember the model's the boss: if you're using a vision model with its own system prompt baked into the server config, the system_prompt set in the config node upstream may conflict or be overridden - when captions come back wrong, first suspect the server's own prompt settings, not this node.