README.md

Monks Image Generation Prompt Enhancer

A ComfyUI custom node that uses the Gemini API to rewrite a raw prompt into one optimised for a specific image generation model and workflow.

The node selects the right prompting strategy based on two inputs: the target image generation model and the purpose. Each combination loads a dedicated system prompt, so the output always matches the expectations of the downstream model.

Features

• Per-model, per-purpose system prompts — no one-size-fits-all rewrites
• Supports Gemini Nano Banana and FLUX.2 Klein out of the box
• Three workflow modes: text-to-image, image editing, multi-image fusion
• Masked API key input
• Raises errors on failure — no silent fallbacks

Installation

1. Copy the monks_prompt_enhancer/ folder into your ComfyUI/custom_nodes/ directory.
2. Install the dependency:

   pip install google-generativeai

3. Restart ComfyUI.

The node will appear under the Gemini AI/TextGen category.

Inputs

Input	Type	Description
prompt	Text (multiline)	The raw prompt to enhance
api_key	Text (masked)	Your Google Gemini API key
prompt_generation_model	Dropdown	Gemini model used to enhance the prompt
target_image_generation_model	Dropdown	Image generation model the enhanced prompt is intended for
purpose	Dropdown	Workflow mode
reference_image	IMAGE (optional)	Reference image passed to Gemini alongside the prompt

Outputs

Output	Type	Description
enhanced_prompt	STRING	The enhanced prompt returned by Gemini

Target Image Generation Models

Gemini Nano Banana

Prompts are written as rich narrative prose structured around subject, setting, details, lighting, atmosphere, and style — optimised for Gemini's native image generation pipeline.

FLUX.2 Klein

FLUX.2 [klein] does not apply prompt upsampling: what you write is exactly what you get. The prompting strategy therefore varies by purpose:

• Text to Image — flowing, highly descriptive prose (no keyword lists). Leads with the main subject, front-loads the most important elements, and ends with an explicit style and mood statement. Lighting description is treated as the single highest-impact element.
• Image Editing — concise, action-oriented directives that describe only the transformation. Base-image descriptions are stripped out entirely; the pixel data carries the visual context.
• Multi Image Fusion — minimal explicit reference prompts (e.g. "Change image 1 to match the style of image 2"). Over-prompting confuses the model's cross-attention mechanisms, so prompts are kept as simple as possible.

Purposes

Text to Image

Expands a raw idea into a detailed, model-ready prompt for generating a new image from scratch.

Image Editing

Refines a description of what should change in an input image. Returns only the transformational action — not a description of what the image already looks like.

Multi Image Fusion

Produces a prompt that defines how multiple input images should interact or transfer properties to each other.

System Prompts

System prompts are organised by target image generation model under prompts/:

prompts/
├── nano_banana/
│   ├── text_to_image.md
│   ├── image_editing.md
│   └── multi_image_fusion.md
└── flux_klein/
    ├── text_to_image.md
    ├── image_editing.md
    └── multi_image_fusion.md

Edit the relevant file to customise the enhancement behaviour for a specific model/purpose combination without touching the node code. To add a new target model, create a new subfolder with the three purpose files and register it in IMAGE_GENERATION_MODELS in monks_prompt_enhancer.py.

Prompt Generation Models

prompt_generation_model	Model ID
Gemini Flash 2.5	gemini-2.5-flash

Usage Examples

Text to Image — no reference image

Write a short description and let the node expand it into a detailed, structured prompt ready for image generation.

Input: Generate a wide open shot of a forest Output: A photorealistic wide-angle shot of a majestic, ancient forest, showcasing towering pine and oak trees reaching towards the sky. The forest floor is covered in rich green moss and ferns, with shafts of soft, dappled golden hour sunlight filtering down through the dense canopy, creating a serene and ethereal atmosphere. Captured with a wide-angle lens, emphasizing the expansive scale and natural beauty of the untouched wilderness, with a shallow depth of field.

---

Text to Image — with Reference Image

Connect an image to give Gemini a visual reference. The node generates a new prompt inspired by the image's composition, style, or content.

Input: Use the reference image and generate a prompt to create that image Output: A high-resolution, studio-lit product photograph of a pink Pepto Bismol bottle with a yellow cap, standing prominently at the center of a clean, light-colored marble dining table. In the softly blurred background, a woman and a man are seated at the table, engaged in a warm, natural conversation. The entire scene is brightly illuminated by bright, natural daylight streaming through a large window, creating soft reflections on the marble surface and adding warmth and authenticity to the lifestyle setting of the product. Camera angle is a straight-on medium shot, with a shallow depth of field keeping the product in sharp focus while gently blurring the background. Ultra-realistic, sharp focus on the product, vertical portrait format.

---

Image Editing — with Reference Image

Connect the image to edit and describe the change. The node strips out any description of the base image and returns only the transformational action.

Input: Remove the Bottle from the Table Output: Remove the pink Pepto Bismol bottle from the table. Fill the space with the existing marble tabletop texture, maintaining the original lighting and reflections.

---

Model Variance — same prompt, different target models

Switching target_image_generation_model on the same input produces structurally different outputs. Gemini Nano Banana receives a rich narrative paragraph; FLUX.2 Klein receives prose or a concise directive tuned to how that model processes text — no keyword lists, no redundant descriptions.