Gemini ComfyUI Styler

Here's the first thing to know, because the name buries it: this node does not call Gemini, or any API, and it needs no key. It ships inside the ComfyUI-OllamaGemini pack alongside the actual LLM nodes, but GeminiComfyUIStyler is a plain, offline prompt builder - a giant menu of curated prompt fragments, not anything AI-powered. You give it a base prompt, bolt on style keywords from a stack of dropdowns, and it hands back a beefed-up positive and negative string.

What it's for

The problem it solves is vocabulary. You know you want a Blade Runner look, or a specific photographer's feel, or an f/1.2 shallow-depth portrait, but you don't know the exact phrasing that steers the model there. This node has those phrasings pre-written and organized into categories like camera, photographers, artist (over 7,700 of them), movies, Color_Grading, lighting, depth, composition, mood, filter, and dozens more. Pick from a dropdown, and it appends the matching text to your prompt. That's the whole mechanism: label → snippet → concatenation. Every dropdown defaults to None, which contributes nothing, so an untouched styler is a pass-through.

The catalog is large and real, but calibrate the hype - the "500+ styles, 150+ creators" badges are the author's own marketing, and the pack's online footprint is mostly its own release posts. What the node actually gives beginners is tricks experienced users type by hand, a few genuinely high-leverage. Naming a concrete referent, like a specific camera body or a named color grade, tends to beat a pile of generic "realistic, candid, 8k" adjectives: those adjectives are largely SD 1.5/SDXL-era vocabulary that does little on today's models, while the concrete names point at things the model was actually trained on. The Color_Grading options earn their keep for the same reason.

The inputs and outputs that matter

You only have to touch two fields:

• text_positive - your base prompt. The styler prepends/appends the chosen fragments to it.
• text_negative - your base negative prompt, treated the same way.

Everything else is optional dropdowns. Set the two or three that match the look you're after and leave the rest on None. The log_prompt toggle (on by default) prints the assembled prompt to your ComfyUI console; leave it on while you're learning, since seeing what got built is the fastest way to understand the node.

The two outputs are text_positive and text_negative, plain strings. Wire them into your two CLIP Text Encode nodes (positive to positive, negative to negative). It's a mid-graph text transformer, not an output node.

Installing it

The easy path is ComfyUI Manager: open it, search ComfyUI-OllamaGemini, install, and restart ComfyUI fully. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/al-swaiti/ComfyUI-OllamaGemini

then restart. The styler itself pulls no model files and needs no credentials. But be aware of what you're installing: this is a large multi-provider pack that bundles SDKs for Gemini, OpenAI, Claude, Ollama, and Qwen. Its requirements.txt drags those dependencies into your Python environment even though the styler uses none of them, and that's exactly where things break - if one dependency fails to install, the entire pack can fail to import and the node silently won't appear. Install from the real repo above and nowhere else; installing a node pack means running its author's code, and this exact shape - a requirements file pulling openai and anthropic builds - is how the ComfyUI_LLMVISION infostealer spread in 2024, hiding its payload inside fake wheels of those libraries.

Where people get burned

The biggest trap has nothing to do with the node and everything to do with what it feeds. This styler happily builds a fat negative string, but on most 2026 models that negative does nothing. Anything guidance-distilled and running at CFG 1 (Flux Dev, Z-Image Turbo, and every Lightning/Turbo/Hyper checkpoint) skips the negative pass entirely; it isn't weakened, it isn't computed. The negative half of this node only earns its keep on real-CFG models like SDXL, Illustrious, and Pony. On a distilled model, restate the constraint as a positive instead.

The second trap is prompt soup. It's very easy to tick eight dropdowns and produce a Frankenstein prompt, and past roughly 75-100 effective tokens the model drifts, while contradictory styles ("oil painting" + "photorealistic" + "anime") average out to none of them. Restraint wins: two or three well-chosen categories beat twenty. The same logic bounds that 7,700-name artist list - most names do little on a base model that never learned them, so they land best on SDXL-lineage finetunes. If a result looks muddy, flip log_prompt on, read what the node assembled, and cut. And if the node never appears after install, it's almost always that failed dependency - check the startup console for a red import error, don't just reload the browser.