prompt-fungineer-355M

Prompt Fungineer 355M

Follow @treadon on X and treadon on Hugging Face for more AI experiments, evals, and projects.

A GPT-2 based model (355M parameters) fine-tuned to expand short image prompts into detailed, creative prompts for AI image generators like Midjourney, DALL-E, Stable Diffusion, and others.

What It Does

Give it a simple prompt like "a cat on a mountain" and it generates detailed, production-ready prompts with lighting, composition, camera settings, and artistic style suggestions.

Input:

a sunset over the ocean

Output:

A breathtaking sunset over a calm ocean, golden hour lighting casting
warm amber reflections across gentle waves, wispy cirrus clouds painted
in shades of coral and violet, shot with a Canon EOS R5, 24-70mm f/2.8
lens, long exposure capturing silky smooth water, dramatic sky gradient
from deep blue to fiery orange, cinematic composition with rule of thirds

How It Works

The model was fine-tuned on a curated dataset of high-quality image prompts (treadon/image_prompts). It learned the structure and vocabulary that image generation models respond to best.

The input format uses structured tags:

• BRF: — Brief/base prompt (your input)
• POS: — Positive prompt (detailed expansion)
• ENH: — Enhancers (camera settings, lighting, technical details)
• INS: — Inspiration (photographer/artist references)
• NEG: — Negative prompt (what to avoid)

Usage

With the Demo

Try it directly: Prompt Fungineer Space

With Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("treadon/prompt-fungineer-355M")
tokenizer = AutoTokenizer.from_pretrained("gpt2")

prompt = "BRF: a castle in the clouds POS:"
inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_length=256,
    do_sample=True,
    top_k=100,
    top_p=0.95,
    temperature=0.85,
    num_return_sequences=4,
    pad_token_id=tokenizer.eos_token_id,
)

for output in outputs:
    print(tokenizer.decode(output, skip_special_tokens=True))

Model Details

Recommended Settings

Options

• Enhancers: Adds camera/lens/lighting metadata (recommended ON)
• Wild Imagination: Lets the model go beyond the base prompt
• Inspiration: Adds photographer references (experimental)
• Negative Prompt: Generates a negative prompt for Stable Diffusion

Limitations

• 355M parameters — produces good but sometimes generic results
• Trained on English prompts only
• Occasionally generates nonsensical photographer names in "Inspiration" mode
• Best for photographic/realistic styles; less effective for abstract art prompts

Links

• Demo: HuggingFace Space
• Training Data: treadon/image_prompts
• Blog Post: Image Prompt Generator

Author

Built by Ritesh Khanna (@treadon)

More from me

For other projects and writeups, see riteshkhanna.com, follow @treadon on X, or treadon on Hugging Face.