Prompt Studio (AI Enhancer)

Here's the modern reality of prompting: your prompt is an instruction fed to an LLM-encoded model, not a bag of tags, and the single highest-leverage phrase you can write is the name of a real camera. Most people still write prompts like it's 2023. Prompt Studio (AI Enhancer) is the node that does the 2026-style prompt engineering for you - it sends your rough brief to a Gemini text model that returns a full, structured, photorealistic prompt you can feed straight into Gemini image generation.

And it has a personality: every prompt it writes carries a built-in anti-AI realism mandate. Real camera bodies and lenses (Leica M6, Hasselblad 500C/M), specific film stocks (Kodachrome 64, Portra 400), optical imperfections like chromatic aberration and halation, physical grain, pores and micro-texture - the works. That list reads like it was lifted from the community's tell-detection threads, because it basically is: too-smooth skin and plastic sheen are still the first things that give an AI image away, four generations of models later.

Three modes

• Expand - short brief to a full structured prompt. The LLM detects the creative domain (portrait, architecture, fashion, product, fine art) and builds the appropriate section structure automatically: a brutalist-interior brief gets Spatial Composition, Material & Texture, Light Study sections; a portrait brief gets Camera, Subject Lock, Tonal Refinement.
• Refine - take an existing full prompt and surgically strengthen the weak parts while preserving your intent and structure.
• Edit - the iteration loop. Wire the previous enhanced_prompt output back into previous_prompt, type feedback like "change lens to 24mm, make it golden hour," and only the sections you mention change.

That Edit loop is the workflow: Prompt Studio (Expand) → enhanced_prompt → generate → not happy → wire it back into a Prompt Studio in Edit mode → regenerate. You iterate on the instruction, not by re-rolling the image.

The inputs that matter

• brief - your creative direction. One sentence is fine for Expand; paste your full prompt for Refine; leave it empty in Edit mode.
• mode - Expand / Refine / Edit, per above.
• model - Gemini 3 Flash is the recommendation and it's right: ~$0.001 per call, so this node is effectively free to run.
• style - two dozen master-photographer presets (Alec Soth, Daido Moriyama, Fan Ho, …). When set, the style's DNA overrides the default system prompt - a styled Expand, a style-injected Refine, a style-preserving Edit.
• images (optional) - reference images for visual context analysis, if you want the prompt to describe what's in a photo.
• custom_instructions - extra directives for any mode: "only Hasselblad references," "keep it under 500 words," "target Architectural Digest editorial."

Outputs

enhanced_prompt (STRING) is the thing you wire into a generation node's prompt input - Gemini Image Generate is the obvious sibling in this pack. analysis (STRING) is the model's commentary on what it did, useful for learning why the prompt looks the way it does.

Installing it and gotchas

Same pack as the rest of the Gemini Direct suite - Manager search "ComfyUI-Gemini-Direct" or the clone-and-pip install from the README, then restart. It needs the same API key setup (gemini_api_key.txt, GEMINI_API_KEY, or the api_key input) and a billing-enabled key.

A few honest caveats. First, the realism mandate is baked into every system prompt - if you're going for stylized, painterly, or deliberately clean anime looks, this node will fight you, so that's a real limitation. Second, Edit mode fails cleanly if you forget to wire previous_prompt or leave feedback empty - the error tells you exactly which one is missing. And third, this node makes text; it doesn't touch pixels, so don't expect to see an image until you wire the output somewhere.