Brought to you by the CuteCaption team.

https://discord.gg/jbxqAaRw
[CuteCaption is an ultra-modern, insanely customizable caption editor and AI computer vision dataset suite currently in beta on our Discord]

CuteCaption Ultra Instagram & OnlyFans LoRA

[please use double block inference for best results]

This is not your typical LoRA. This is the culmination of a pioneering, world-class dataset curation process designed to generate the most aesthetically beautiful and geometrically appealing figures possible. The problem with many datasets is the tendency to bake in gaudy jewelry or unwanted accessories. By deeply re-balancing the dataset, we ensure the LoRA does not generate these elements by default. This means if you prompt for a "portrait of a woman," you won't get unwanted distractions. It gives you, the artist, final creative control.

Key Features

Unprecedented Control: The pièce de résistance of this LoRA is our pre-training curation. Over 650 images have been painstakingly edited to remove jewelry, accessories, tattoos, and background clutter. This means the model is not biased towards generating these elements, giving you final control over what your subject wears and their environment. The problem with many datasets is the tendency to bake in these items. This approach allows us to ensure the LoRA does not generate these elements by default. This means if you prompt for a "portrait of a woman," you shouldn't get unwanted distractions.

The Dataset (~1650 Images): After curating a catalogue of nearly 25,000 images, we meticulously selected a core set that resonates with a specific, ultra-aesthetic "vibe." This isn't about volume; it's about quality and essence. We understand beauty, and we've handled the taste-making for you.

Dual-Style Versatility: The model is trained with two distinct styles.

Instagram Style: Triggered by 1NST4GR4M, this evokes the classic, polished, and alluring feel of top-tier social media content. The LoRA is trained on a variety of subjects, ensuring it doesn't default to a single face.

OnlyFans Style: For more intimate and suggestive scenes, this tag unlocks the model's NSFW capabilities. With 150-300 images tagged this way, it can generate provocative content without automatically defaulting to it.

The "Vibe" Generator: Forget needing prompts like "beautiful woman" or "almond-shaped eyes." This LoRA is trained on a foundation of aesthetic excellence. Simply prompting for "a woman" will generate a person who embodies the curated vibe of the dataset.

How to Prompt - Mastering the Vibe

This LoRA was trained with detailed and descriptive captions, giving you incredible flexibility. The primary trigger word is 1NST4GR4M. Combine this with other keywords to guide the model. There is a secret, modular token "0QMvzt" which was trained on around 300 images to act as a general detail modifier.

Based on a deep analysis of the master captions, here are the most common tokens and themes to consider using in your prompts:

Style & Mood

These are powerful keywords to define the overall feel of your generation.

Core Styles: Instagram Style, Instagram Style Selfie, OnlyFans Style

Atmosphere: seductive, confident, alluring, serene, casual, glamorous, intimate, playful, provocative

Shot Type: Instagram Style Selfie, selfie, booty shot, close-up shot of midriff, from behind, viewed from the side, low angle

Body & Physique

The model is highly responsive to physical descriptors.

Body Types: curvy physique, slender physique, toned, fit physique, athletic build, hourglass figure, voluptuous figure

Key Features: slim waist, wide hips, toned abs, toned legs, cleavage, booty, buttocks

Breasts: large breasts, ample breasts, medium-sized breasts, small breasts

Skin: fair skin, tan skin, tanned complexion, sun-kissed skin, light olive skin, dark skin

Clothing & Lingerie

The dataset is rich with specific clothing styles.

Swimwear (Very Common): bikini, string bikini, thong, one-piece swimsuit, monokini, high-cut swimsuit

Lingerie: lingerie, lace lingerie, lace bra, thong, garter belt, bodysuit, corset

Apparel: crop top, mini dress, mini skirt, high-waisted shorts, jeans, sheer dress, leggings

Footwear: high heels, sneakers

Poses & Actions

Describe what the model is doing to guide the composition.

Common Poses: posing confidently, seductive pose, looking over her shoulder, arms raised, hands on her head, leaning against, kneeling, lounging, sitting on, reclining on

Actions: taking a selfie, sunbathing, walking on the beach, smiling, looking at the camera

Locations & Environments

The model excels in a variety of settings.

Beach & Water: sandy beach, ocean, poolside, in a swimming pool, yacht, on a boat, tropical setting, clear turquoise water

Indoor: bedroom, modern living room, balcony, kitchen, minimalist room, dimly lit room

Outdoor: lush greenery, garden, rocky shoreline, snowy forest, cityscape at night

How to Use

Getting started is simple.

Write a prompt for an Instagram or OnlyFans model as you normally would. More descriptive and robust prompts tend to produce the best results.

Add the LoRA, and include 1NST4GR4M to your prompt. You may not need the trigger word to get desirable results.

Recommended Strength: Start with the LoRA weight around 0.40 and experiment. Results may vary as you approach 1.0.

Crucial Tip: It may also prove crucial to use DOUBLE BLOCKS in your prompting for best results.

The Creator's Journey: Forging the vibe.

Our journey began over two years ago. For this project, the magic is the dataset. Like a DJ with a keen "ear," curating a model requires a keen "eye." I have always had a deep connection to aesthetic geometry and an even deeper appreciation for beauty. We understand that extremely subtle differences in facial geometry can dramatically alter perception. This LoRA is the result of that understanding and taste-making.

This isn't just about cramming "hotties" into a dataset. There is a vibe here. We started by carefully selecting a "core" set of 600 images from our full catalogue of nearly 25,000. By immersing ourselves in this core vibe, the dataset began to resonate. Over time, we could merely glance at a photo to "know" if it fit. This feedback loop of immersion, curation, and minor pruning allowed the dataset to grow to ~1650 images, each one a piece of a puzzle that fits perfectly. Every single image matters; if there was the slightest doubt, it didn't make the cut.

The dataset is a work of art in its own right, sourced from Instagram, OnlyFans, Flickr, and even old hard drives to find timeless, perfect shots that may no longer be available publicly. Some lower-quality gems were even sharpened and cleaned up just to be included. We've also taken extreme steps with captioning, using an LLM as a base and then manually reviewing every single entry for ground-truth accuracy.

Disclaimer: Although a tremendous amount of love and care went into this LoRA and especially the dataset, we do not want to make claims that it is perfectly performant or the best Instagram / OnlyFans model available. We have explored many different training parameters, and although we HAVE been able to get some extremely beautiful generations, there is still a lot of testing that needs to get done, so although massive love has gone into the dataset, it may still need some work yet. We are launching this model in the hope that we can get enough feedback to launch version 2, of which we have curated an additional 300 images and will begin to include videos in the dataset.