EBU LMStudio Make Request

This is the workhorse node in the pack, and its default settings already tell you exactly what it's for. Look at the system_message it ships with - "You are an assistant designed to craft AI image prompts... Respond with just your new prompt" - and you've got the whole idea: you type a rough concept into prompt, a locally-running LLM fleshes it out into something detailed and specific, and you wire the result straight into your conditioning node. No cloud API, no key, no per-call cost. Just whatever model you've already got sitting in LM Studio.

That framing matters more in 2026 than it would have a couple years back. A growing share of image models - Flux, Z-Image, Anima, Krea 2 - are reading your prompt through an LLM text encoder rather than tokenizing it as a bag of CLIP tags, which means they respond to structured natural-language instructions, not (word:1.3) weights or comma-separated booru tags. Asking an LLM to write that kind of prompt for you, in that style, is a genuinely sensible division of labor. And this isn't a hypothetical pairing - people are already running exactly this pattern in the wild, pointing a local chat-completions-style node at their own hardware to generate or expand prompts before they ever hit the sampler, no cloud round-trip involved.

What it actually does

Under the hood it's a plain HTTP call to an OpenAI-compatible chat completions endpoint - the same API shape LM Studio, Ollama, and half the local-LLM ecosystem all speak. prompt is your user message, system_message is the instruction that shapes how the model responds to it, and url is where that request goes (default http://127.0.0.1:1234/v1/chat/completions, LM Studio's own local server). Because it's just a generic chat call, you're not locked into prompt-writing - swap the system_message for a captioning instruction, a style critic, whatever, and it'll do that instead. The default text is a starting point, not a constraint.

The knobs worth actually touching: context_length needs to be at or below what the model was loaded with in LM Studio, or you'll get truncated context. temp and top_p are the usual creativity dials - push temp up if every generated prompt is coming back suspiciously similar. max_tokens caps the response length; the default of 300 is plenty for a single image prompt but tight if you're asking for something more elaborate. seed does double duty: it's passed through as the request's sampling seed for reproducibility, and just as importantly it forces ComfyUI to actually re-run the node - without something changing between calls, ComfyUI's output caching will happily hand you back yesterday's cached result instead of hitting the LLM again. utf8_safe_replace is a safety net for local models (especially smaller quantized ones) that occasionally emit a broken multi-byte character - flip it on if you ever see mangled text in the output.

The single output is generated_text - a plain string. Wire it into a CLIPTextEncode, or into a Show Text node first if you want to eyeball what the LLM actually wrote before it hits your sampler.

Installing it

Through ComfyUI Manager: search "EBU LMStudio" and install. Or by hand - cd ComfyUI/custom_nodes && git clone https://github.com/burnsbert/ComfyUI-EBU-LMStudio, then restart ComfyUI. That's the node pack side, and it's genuinely lightweight - no extra pip dependencies beyond what ComfyUI already has.

The real install is on the LM Studio side, and it's separate software you need running alongside ComfyUI: install LM Studio itself, turn on Developer Mode so the local server is available, download at least one LLM inside it, and confirm CLI access works (lms ls --detailed should list what you've got installed - this pack talks to LM Studio through that CLI as well as its HTTP server). Then restart ComfyUI so it picks up the new node.

Where people get burned

Connection-refused errors almost always mean LM Studio isn't running, Developer Mode is off, or the url doesn't match your actual host/port. If the request goes through but errors on "no model loaded," LM Studio has nothing loaded yet - either load one by hand in the app, or chain EbuLMStudioLoadModel in front of this node so a model is guaranteed to be ready first. And because this node is talking to whatever's local, its output quality is entirely a function of which model you loaded - a small 7-8B local model writing image prompts is not going to match GPT-4-class output, so if the generated prompts feel flat, that's the model, not this node.