LM Studio Unified (URL + API Key)

So you've got a local LLM running in LM Studio and a ComfyUI graph full of images, and you'd like the two to talk. Maybe you want a caption for a generated image, a critique, a rewritten prompt, or a decision ("pick the best frame") that feeds back into the graph. The pieces all exist, but wiring a custom HTTP call into ComfyUI by hand is a small project of its own.

LM Studio Unified (URL + API Key) is that wiring, done for you. It calls LM Studio's OpenAI-compatible chat endpoint with text, an image, or both, and hands back the model's reply as a generated_text string.

How it works

Everything goes through LM Studio's /v1/chat/completions API. The node builds a chat message from your system_prompt and prompt, and if you connect an image, each frame of the batch is encoded to a JPEG data URL and sent as a separate image_url content part (capped by max_images). So vision-capable models actually see your image; text-only models can still use it as a pure text chat node.

It's intentionally dependency-free - plain urllib, no openai SDK, no pip installs. Request details go to the ComfyUI console when debug is on, and IS_CHANGED hashes all inputs so the node re-runs when anything changes and doesn't rerun when nothing does.

The inputs that matter

• base_url - the endpoint. Defaults to http://127.0.0.1:1234/v1, the standard local LM Studio address. It also accepts https://host, /v1-style paths, or a full /v1/chat/completions URL, so it works against remote/Cloudflare-fronted routes too.
• model - the model id from LM Studio's /v1/models list. Must match what you loaded in the app.
• prompt / system_prompt - the user and system messages.
• image (optional) - the IMAGE batch to send.
• api_key / api_key_env_var - optional. A blank api_key falls back to the env var (default LMSTUDIO_API_KEY), which on Windows also checks User/Machine registry values.
• seed, max_tokens, temperature, timeout_seconds, user_agent, debug - the rest of the knobs. user_agent matters for routes behind Cloudflare that block Python's default urllib signature; there's a browser-like one preloaded.

Output: generated_text, the model's reply as a STRING.

Install

Part of the Enviral Design Node Pack. ComfyUI Manager: search "Enviral Design Node Pack". Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/EnviralDesign/comfyUI-enviral-design-node-pack

then restart. No extra dependencies, and no LM Studio install required by the node itself - you just need an endpoint reachable.

Gotchas worth knowing

The single most common failure is a model id that doesn't match anything loaded in LM Studio - check the app's model list, not your memory of the filename. For local use, 127.0.0.1:1234 with no key is the zero-config path; if you've pointed it at a remote route, set the key (or env var) and make sure the server's CORS/auth accepts browser-like traffic. And remember the model you're asking must be vision-capable to describe images - a text-only model will politely ignore the image parts. Debug mode is your friend here: it prints exactly what's being sent.