nightcafe-prompt-generator-v0.1

This model is a fine-tuned T5-Base model designed to generate "convoluted" and "over-engineered" image generation prompts from concise, two-sentence descriptions.

Model Description

This model was fine-tuned on a dataset of NightCafe Studio user-generated prompts. The input to the model is a concise, two-sentence description of an image, generated by a multimodal model (Gemma-3N-e4b in this case). The target output is the original, often highly detailed and stylized, human-written prompt.

Base Model: google/t5-base

Training Data:

Source: NightCafe Studio user-generated image prompts and associated image data.
Filtering: Prompts were filtered to be between 75 and 400 characters in length.
Input Generation: Concise, two-sentence descriptions were generated from the original images using a local Gemma-3N-e4b model.

Intended Use

The primary purpose of this model is to explore and demonstrate the concept of "prompt over-engineering" in text-to-image generation. It aims to bridge the gap between simple, objective image descriptions and the complex, often keyword-stuffed prompts used by human users to achieve specific artistic styles and qualities.

Limitations and Biases

Subject Fidelity: The model may sometimes alter the core subject of the input description, especially for less common subjects, defaulting to more prevalent patterns learned from the training data (e.g., transforming "cat" into "woman" if "woman" is highly correlated with certain artistic styles in the dataset). This highlights a bias in the training data and the model's learning process.
Generalization: Performance on prompts outside the 75-400 character range or on subjects/styles not well-represented in the training data may vary.
Output Quality: While designed to be "convoluted," the generated prompts may not always be optimal for actual image generation and might require further human curation.

How to Use

(Instructions for loading and using the model will be added here after deployment.)