README.md

---
license: apache-2.0
language:
- en
base_model:
- black-forest-labs/FLUX.1-dev
pipeline_tag: text-to-image
tags:
- art
- realism
- lora
---
<h1 align="center">Babes LoRA</h1>

<p align="center">Low-Rank Adapter for Flux. Dev trained on <b>100</b> to <b>170</b> babes images.</p>

---

## Why ?
I have no idea... but it works.
Not only is it good at generating instagram-like images, it's also surprisingly great at realism - leading to better results than most flux realism based LoRA

## Usage
Because of the little amount of data and the training techniques used, it is recommended to use these LoRAs with weights above `1.0`.
At these weights it does not mix very well with other LoRAs.

Weights are provided in `safetensors` format so they can be used with any generation library or platform.

## Preview

| Preview | Version | Recommended Usage/LoRA Weight |
| --- | --- | --- |
| ![base](./assets/base.png) | - | _Raw output from Flux Dev_ | 
| ![1.0](./assets/1.0.png) | base | min=1.44, max=3.0 |
| ![1.1](./assets/1.1.png) | 1.1 | min=1.44, max=3.0 |
| ![1.1a](./assets/1.1a.png) | 1.1a | min=2.0, max=2.44 |
| ![1.1b](./assets/1.1b.png) | 1.1b | min=2.0, max=2.44 |
| ![1.2a](./assets/1.2a.png) | 1.2a | min=1.44, max=2.0 |
| ![1.2b](./assets/1.2b.png) | 1.2b | min=1.44, max=3.0 |
| ![1.3a](./assets/1.3a.png) | 1.3a | min=1.44, max=2.0 |
| ![1.4a](./assets/1.4a.png) | 1.4a | min=1.44, max=2.0 `Latest dataset` |


## BibTeX
```bibtex
@Misc{babes,
  title =        {Babes: Flux Low-Rank Adapter trained on 100 babes images},
  author =       {Hugo Ventura},
  howpublished = {\url{https://huggingface.co/hugovntr/babes}},
  year =         {2025}
}
```