APNext Gemini Text Only

This is the no-image Gemini node: pure text in, prompt out. It's for when you don't have a reference picture, just an idea or a template you want Gemini to expand into a full image prompt. Think of it as the cheap, fast prompt-writer of dagthomas's LLM lineup - Gemini Flash is quick and inexpensive, and pointing it at a well-written template gives you consistent, structured prompts without a chat tab.

Having an LLM write your prompt is the modern default, not a gimmick - on 2026 models whose text encoder is itself an LLM (Flux 2, Z-Image, Qwen), you're translating your idea into an instruction the image model reads natively. GeminiTextOnly is the lightweight way to do that in-graph, and its dynamic-template support makes it especially handy for churning out LoRA-caption-style prompts at volume.

How it works

It fills any dynamic tokens into your prompt and sends it to Gemini, returning the generated text. There's no image input - everything comes from custom_prompt (plus additive_prompt). Turn on dynamic_prompt and ##TAG##, ##SEX##, ##WORDS##, ##PRONOUNS## get substituted from the matching inputs, so one template produces consistent captions across many subjects. seed and randomize_each_run give you fresh variations per run.

The inputs and outputs that matter

• custom_prompt - your instruction or template. The whole node runs on this; pair it with CustomPromptLoader to pull a saved one.
• additive_prompt - extra text appended to the instruction.
• gemini_model - the dropdown (2.5 Pro down to Flash-Lite). Flash is the cheap, fast default.
• dynamic_prompt + tag / sex / words / pronouns - template substitution for repeatable captions.
• seed + randomize_each_run - batch variety.

Two outputs: output (the generated text) and clip_l (a trimmed variant sized for a CLIP-L text slot).

How to install it

ComfyUI Manager: search comfyui_dagthomas, install, restart. Manual:

cd ComfyUI/custom_nodes && git clone https://github.com/dagthomas/comfyui_dagthomas
cd comfyui_dagthomas && pip install -r requirements.txt

then restart. Needs google-generativeai (in requirements.txt) and a GEMINI_API_KEY set in the environment before launching ComfyUI.

Where people get burned

• Key set after launch, or missing. Export GEMINI_API_KEY before you start ComfyUI. Read at startup, nowhere else.
• Hosted ComfyUI blocks it. Outbound API calls aren't available on most serverless instances; this is a local-install node.
• Empty template, empty value. With no custom_prompt there's nothing for Gemini to expand. Start from a CustomPromptLoader template if you're unsure.
• Watch the length. Gemini will write a lot if you let it; the KB's attention-cap point (models drift past ~75–100 effective tokens) applies. Ask for concise output in the template, or you're feeding the encoder an essay.
• Model name drift. The dropdown can list a variant your key can't call; that errors at runtime. Edit data/gemini_models.json to match your access.