ErniePEUnleashed – Composition & Art‑Style Prompt Enhancer

ErniePEUnleashed is a specialized prompt enhancement model built for AI image generation pipelines. More specifically it is a purpose build fine tune of the baidu/ERNIE-Image prompt enhancer, trained on a custom-built dataset of 3804 unique compositions. It transforms short prompts or basic ideas into richly detailed, spatially structured compositions and seamlessly integrates specific art styles into the description, giving you full creative control over the look of your generated images. Unlike generic prompt enhancers that simply pile on keywords, ErniePEUnleashed understands composition: it writes foreground‑midground‑background hierarchy, precise lighting logic, camera angles, and material textures—exactly what an image generation model needs to produce striking, coherent images.

Model Details

Base model: Ministral-3-3B-Instruct-2512 / more exactly the prompt enhancer that comes with baidu/ERNIE-Image

Fine‑tuning method: Full 16‑bit LoRA (rank 32, alpha 64, 7 target modules)

Dataset size: 3,804 examples, unique, custom built dataset with both manual and AI built compositions with a roster of 24 local models and 8 of the online ones, generated with very specific rules.

Training epochs: 3

Max sequence length: 4096 tokens

Model Description

This is a fine-tune of baidu/ERNIE-Image's prompt enhancing model, made by me. It should work as both a prompt enhancing model, as well as just a composition builder in general. It can do both pure compositions, and art style infused compositions for AI image generation, and whilst it is based on the baidu/ERNIE-Image pe model, realistically it can work with most single positive prompt workflows, especially qwen-image-2512, since that was my main specialization.

model.safetensors – is just the normal finished merge of the model and trained lora

ErniePEUnleashed-comfy.safetensors - is the version that works with just the generic comfyui workflows for Ernie-Image

ErniePEUnleashed-bf16.gguf – added the gguf in case people just want to use it, cause why not..

This is the list of Art Styles that it has been trained on:

• "16-bit Pixel Art"
• "8-bit Pixel Art"
• "90s Anime-OVA"
• "American Superhero Comics - Silver Age"
• "Architectural Fantasy Linework"
• "Art Nouveau"
• "Atmospheric Impasto Oil Painting"
• "Charcoal Drawing"
• "Chinese Landscape Relief Sculpture"
• "Cinematic Fantasy Illustration"
• "Disney Afternoon Animation"
• "Expressionist Digital Illustration"
• "Franco-Belgian Comic - Ligne Claire"
• "French Graphic Novel - Metal Hurlant"
• "Golden Age Book Illustration"
• "Gouache Illustration"
• "Grunge Collage"
• "Hanna-Barbera Flat Animation"
• "High Energy Flat Graphic Anime"
• "Japanese Ink Wash"
• "Layered Anime Scenery Illustration"
• "Low Poly 3D"
• "Manga - Seinen"
• "Mirrors Edge Clean Style"
• "Oil Painting - Classical"
• "Oil Painting - Impressionist"
• "Painterly Fantasy Matte Illustration"
• "Painterly Neon Digital Art"
• "Painterly Strong Graphic Novel"
• "Pastel - Soft"
• "Pen and Ink Sketch"
• "Pencil Sketch"
• "Pop Art Graphic Illustration"
• "Psychedelic Expressionist Horror"
• "Retro Graphic Novel Poster Illustration"
• "Risograph Print"
• "Studio Ghibli Aesthetic"
• "Technical Graphic Illustration - Duotone"
• "Vaporwave Aesthetic"
• "Vintage Pulp Fiction"
• "Watercolor"
• "Woodblock Print - Japanese Ukiyo-e"

If you want to use a specific art style, I would very much suggest that you use the full name of the art style you want from this list in the prompt, otherwise I won’t guarantee how well it will infuse art style.

Otherwise, with no mention of art style it will make a similarly well crafted/detailed composition but without introducing any art style to it.

Example from the Example zip:

1.Art Style: 16-bit Pixel Art

My prompt:

 A neon-lit alley market at midnight during heavy rain in 16-bit Pixel Art style.

The enhanced result:

A 16-bit pixel art composition of a neon-lit alley market at midnight, where every element is constructed from carefully placed colored pixels on a medium-resolution grid. The scene unfolds in a narrow corridor of stalls under a sky swallowed by rain.

Foreground:
The immediate foreground is dominated by a vendor's stall constructed from blocky wooden crates and metal shelving, their surfaces rendered with visible pixel edges and subtle color variation. Steam rises from boiling plates in small, blocky wisps that dissolve into the air, their forms defined by sharp pixel transitions. A customer stands at the counter, their dark coat a collection of large, distinct squares with slight color shifts for shading, while their face remains obscured by hood or shadow. Rainwater drips from the eaves onto the polished countertop in distinct, glowing pixels that catch the light, creating sharp reflections on the metallic surface below.

Midground:
The alley stretches into the midground, its walls lined with more stalls each constructed from blocky structures with glowing signs emitting neon light. The signs are made of sharp-edged pixels that bleed light outward in distinct color gradients. Hanging lanterns cast warm yellow light that mixes with the cool blue of the neon, their illumination creating hard-edged shadows on the wet ground. A vendor leans against a stall, their figure rendered as a collection of distinct pixel blocks with slight color variations for shading. Rain continues to fall, creating a misty atmosphere where distant figures become increasingly pixelated and indistinct.

Background:
In the far background, the alley opens into a larger market square under a bright, overcast sky rendered in muted grays and blues. The scale of the buildings is reduced, their outlines becoming simpler pixel patterns against the sky. The rain continues unabated, with water droplets falling in sharp, defined lines that reflect the neon lights in the alley. The overall lighting creates a contrast between the warm, intimate glow of the market and the cool, distant light of the city beyond.

Lighting & Atmosphere:
The scene is bathed in the warm glow of neon lights from the signs and lanterns, contrasted by the cool blue of the rain and the distant city lights. The rain creates a misty atmosphere that softens the edges of the scene but maintains sharp pixel transitions where light catches the droplets. The overall effect is one of a vibrant, bustling market at night, rendered with the precision and detail of an 16-bit era console game.

Mind you, a lot of the naming in this list is just because they are based on my own developing library of style prompts, that aren’t necessarily the thing you might expect with some of the more niche stuff, however you can check out the examples in the Art style results.zip, where you have both the text file with before and after, as well as all the resulting pngs. Now those should just be taken as a first impression of how the art style looks more than anything, as they are one-time renders, all with the same exact settings and static seed. These weren’t me trying to get the best results. Also, all of the examples are based on one sentence prompts, however the model is trained to improve already existing prompts as well, that just aren’t fully fleshed out, or to infuse already well built prompts with the desired art style.

Do you need any specific system prompts?

No, this should work with a simple or no system prompt even, especially for just generating compositions.

I did the training with very generic, simplistic system prompts in the dataset, depending on task, with 4 different, but similar variations for both composition building and prompt enhancement tasks, specifically so it doesn’t learn x system prompt is what it needs to reply to in a certain way, and hopefully learn more of what a good composition is. The examples themselves are all done with just the generic, default system prompt in the ernie workflow, and not a specific one of mine.

For reference, if someone really wants them, these are the system prompts I used in the dataset:

SYSTEM_PROMPTS_COMPOSITION:    

"You are an expert composition writer. The compositions you write will be used for AI image generation. You write detailed, structured compositions with concrete visuals and clear visual hierarchy.",

   

"You are a specialist in writing image generation prompts. Your task is to produce compositions that are highly detailed, logically structured, and grounded in concrete visual elements. Each composition must include a clear visual hierarchy, guiding the viewer's eye through the scene.",

   

"You write detailed, well-structured composition prompts for AI image generation. Your descriptions focus on concrete visuals and a clear ordering of visual importance. You are an expert at this craft.",

   

"As an expert composition writer for AI image generation, you create prompts that are rich in concrete visual detail and structured with a clear visual hierarchy. Your compositions are thorough, logically arranged, and ready to be interpreted by an image generation model.",

SYSTEM_PROMPTS_PE:    

"You are a prompt enhancing expert for AI image generation. The composition you enhance will be fed directly to an image generation model. The model responds well to clear spatial hierarchy, specific visual details, and purposeful compositional logic. Your compositions should be detailed and well structured and expand on the user's input, be it a shorter prompt or a simple idea or request.",

   

"You are an expert at enhancing prompts for AI image generation. The compositions you produce will go directly into an image model, which performs best with clear spatial hierarchy, precise visual details, and deliberate compositional logic. Your enhanced prompts must be detailed, well structured, and expand on whatever the user provides – whether a short phrase, a simple idea, or a basic request.",

   

"Your role is to improve prompts for AI image generation. The final composition will be fed directly to an image model, which responds strongly to clear visual hierarchy, concrete specifics, and intentional composition. You must produce detailed, well-organized prompts that build upon the user's input – no matter how short or simple the original idea.",

   

"You are a prompt enhancement specialist for AI image generation. The composition you create will be interpreted directly by an image generation model, which works best when given a clear spatial structure, vivid visual details, and a logical compositional flow. Your job is to produce rich, structured prompts that expand on the user's starting point – whether it's a brief phrase, a vague concept, or a straightforward request.",

What the training itself tried to teach, and examples it gave of:

Examples of creating just a random, good detiled composition, both with and without art styles.


Examples of creating a good composition based on genre.


Examples of creating a good composition based around a specific emotion.


Examples of creating a good composition based on a specific Universe.


Examples of Infusing an already existing good composition with art styles, without changing the core structure of the composition. 


Examples of Enhancing a prompt with both no art style and art style.


Examples of creating a more targeted, good composition, ie. examples of good portrait, landscape, cityscape, etc.