🔱 Gemini Subject

Gemini Subject is the entry point on the composition side of the Artha pack - the node that answers "what's the main subject?" and hands the answer downstream in a structured form. Its default text_prompt is "A cat with a hat", which tells you the whole vibe: it's the pack's least serious-looking node and also one of its most flexible. It's from the 🔱 Artha-Gemini family (Cyrostar, category Artha/LLM/GEMINI), and its subject output (type ARTHASUBJECT) feeds Gemini Compose - the pack's scene/video counterpart to Gemini Portrait - where it joins camera, style, light and scene into a single composition prompt.

Two modes, and the manual one is almost embarrassingly simple. With use_image off, the node does nothing but pass your text_prompt through to the subject output. No API call, no processing - it's a glorified pass-through wire, and it exists so your prose subject can flow into Compose alongside the structured dropdown nodes. That's fine; not every node needs to be clever.

The interesting mode is use_image. Flip it, pick a reference image, and Gemini Vision takes over with a subject-analysis agent that describes only the main subject - no background, no secondary elements. It starts with the subject type, then physical features, pose, and distinguishing details, and it returns a paragraph beginning with SUBJECT:. If the main subject is a person, the agent is told to go deeper: face, hair, makeup, body, fitness, pose, clothing. And there's an only_main toggle for the case where there's more than one subject - flip it off and the node switches to a plural-agent that describes every subject and their interactions. That's the whole clever bit: one boolean changes the agent from "describe the one thing" to "describe the people and how they relate."

The inputs that matter

• text_prompt - your subject description for manual mode. The default "A cat with a hat" is fine to keep if you just want to move on.
• use_image - the mode switch. Off: pure pass-through, free. On: Gemini reads your image, billed per run.
• only_main - true by default. Turn it off only when the image has multiple subjects you want described individually.
• image - a file dropdown with an upload widget, not an IMAGE tensor socket.

Then the standard pack cluster: api_key, model (gemini-2.5-flash), max_tokens, temperature. Single output: subject.

Install & gotchas

Pack-wide install, once:

cd ComfyUI/custom_nodes
git clone https://github.com/Cyrostar/Artha-Gemini
cd Artha-Gemini
pip install -r requirements.txt

Restart ComfyUI, or search "Artha-Gemini" in ComfyUI Manager. Light dependencies (google-genai, Pillow, numpy, torch, soundfile), no model downloads. A Gemini API key from Google AI Studio is required only for use_image mode.

The usual family traps apply: image is a file picker, not an image socket - upload a file or select one from your ComfyUI/input folder, don't wire a Load Image node into it. Image mode is a billed call every run. And the plain-text-key warning again: api_key lives in your workflow JSON, so prefer the pack's api.json (gemini_api_key) or a GEMINI_API_KEY environment variable. One thing worth knowing that isn't obvious: the manual mode's pass-through means this node is the cheapest way to test the pack - no key, no tokens, instant - and it only starts costing you when you hand it a photo.