APNext Gemini Prompt Enhancer

This is the hybrid of the whole pack, and it's cleverer than it first looks. Give it a base prompt and it does two things: it injects cinematic terminology from a big set of dropdowns (lighting, camera angle, shot size, lens, color tone, and a dozen more), and it can run the result through Gemini for a polished rewrite. Then it hands you both versions plus the combination, so you can A/B them and keep whichever reads best. If you want the concrete-photographic-language boost the KB says actually works, delivered semi-automatically, this is the node.

The best part: the random-injection half needs no API key. Leave the LLM toggle off and it's a pure, offline cinematic-term enhancer - free, instant, works anywhere. Flip the toggle on and Gemini smooths the injected terms into natural language. That optionality is unusual and genuinely useful; most "enhancer" nodes force the API call.

How it works

The node takes base_prompt and, according to enhancement_mode, pulls elements from its cinematic dropdowns - each defaults to random, so it rolls a coherent set of camera/lighting/style terms and appends them. intensity scales how much it adds. If use_llm is on, it then sends that enriched prompt to Gemini for a rewrite. You can pipe in a builder node's output through apnext_chain to combine this with the rest of the pack. Because those cinematic terms are concrete physical referents, they're the kind of prompt content that transfers across encoders - they help on SDXL and on the LLM-encoded 2026 models where quality-word spam does nothing.

The inputs and outputs that matter

• base_prompt - your starting idea.
• enhancement_mode - how it enriches: Random Mix (4-6 elements), focused modes (Cinematic/Lighting/Camera/Motion/Style), Full Enhancement, or LLM Only.
• use_llm - the key switch. Off = free offline random injection; on = Gemini rewrite (needs a key).
• intensity (0.1–2.0) - how heavy the enhancement.
• The optional dropdowns - lighting_type, camera_angle, shot_size, lens_type, color_tone, visual_effects, composition and more - pin a specific choice or leave on random.

Three outputs: enhanced_prompt (the final combined result), random_enhanced (just the term-injected version, no LLM), and llm_enhanced (just Gemini's rewrite). Wire whichever wins into your CLIP Text Encode.

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. The offline path needs nothing but the install. For use_llm, set GEMINI_API_KEY in the environment before launching ComfyUI.

Where people get burned

• use_llm on without a key. Turning on the LLM rewrite with no GEMINI_API_KEY set is the usual error. Either set the key before launch, or leave use_llm off and use the free random_enhanced output.
• Three outputs, pick one. People wire enhanced_prompt and wonder why random_enhanced looks different - they're supposed to. Compare all three, keep the best; don't feed all three into one encoder.
• Cranking intensity. Max intensity buries your subject under camera and lighting jargon. Keep it moderate so the base idea still leads.
• random everywhere = a slot machine. Every dropdown on random gives huge variety and low control. Pin the choices that matter (shot size, lighting) and randomize the rest.
• Overlong output on LLM mode. Gemini can pad; the attention cap still applies. Trim before it hits the encoder.