AI_Ollama_image

Every other AI chat node in this pack needs an API key and a Chinese-market account signup. This one doesn't - it talks to Ollama, the tool that runs open LLMs and vision models locally on your own machine, so there's no key, no billing, no rate limit beyond your own hardware. The tradeoff is obvious: it's only as good as whatever model you've pulled and however much VRAM/RAM you can throw at it.

The node's own description is basically an install instruction disguised as node metadata: CMD运行命令: ollama run qwen3-vl:latest - "run this command in your terminal: ollama run qwen3-vl:latest." That's the actual setup step. This node doesn't download or manage models itself; it expects Ollama to already be running with the model you want pulled and available, exactly the way you'd use Ollama from the command line normally.

The model_name dropdown offers qwen3.5:latest, qwen3.5:27b, qwen3-vl:latest, and gemma3:12b as presets, but since this is querying your own local Ollama instance, custom_model (a free-text field in the optional section) is genuinely the more useful option once you know what you have pulled - type any model tag you've already got locally and it'll use that instead of the dropdown. Beyond the model, required fields are preset (defaults to "None" until you save one via AI_PresetSave), custom_system_prompt, and analysis_prompt - your actual question about the image. Optional: image_1/image_2/image_3, temperature (0–1, default 0.7), max_tokens (1–8192, default 2048), seed, and a genuinely nice extra - enable_ocr (boolean, off by default), which nudges the model toward reading and transcribing text in the image rather than describing the scene generally. Outputs are analysis_result and system_prompt.

Installing it: the node itself installs the usual way - ComfyUI Manager, search "ComfyUI-Apt_Preset," or cd ComfyUI/custom_nodes && git clone https://github.com/cardenluo/ComfyUI-Apt_Preset followed by install.bat and a restart. But that only gets you the ComfyUI side. Ollama itself is a separate install (ollama.com, or your platform's package manager) - download it, then pull whichever model you want to use, e.g. ollama pull qwen3-vl:latest, before this node has anything to talk to.

Troubleshooting: if every call fails immediately, Ollama's local service almost certainly isn't running - this pack ships a companion node, Ai_Ollama_RunModel, specifically to check and start that service from inside the graph, and it's worth putting ahead of this node in your workflow so a dead service fails loudly instead of this node timing out mysteriously. If a specific model name errors, confirm you actually pulled it (ollama list from a terminal shows what's available locally) - the dropdown options are just suggestions, not a guarantee those exact tags exist on your machine. And if responses are slow or truncated, that's a hardware ceiling, not a bug: vision models are memory-hungry, and qwen3.5:27b in particular needs real VRAM/RAM to run at a usable speed.