README.md

---
license: apache-2.0
datasets:
- peteromallet/InScene-Dataset
base_model:
- black-forest-labs/FLUX.1-Kontext-dev
tags:
- image
- editing
- lora
- diffusers
pipeline_tag: image-to-image
---

# InScene: Flux.1-Kontext.dev LoRA

## Model Description

**InScene** is a LoRA for Flux.Kontext.dev that's designed to generate images that maintain scene consistency with a source image. It is trained on top of Flux.1-Kontext.dev.

The primary use case is to generate variations of a shot while keeping the background and overall environment, characters, and styles the same:
![samples.png](samples.png)

## How to Use

To get the best results, start your prompt with the phrase:

`Make a shot in the same scene of `

And describe your new image.

For example:
`Make a shot in the same scene of the car up very close to the camera with the driver smiling manically.`


### Strengths & Weaknesses

The model excels at:
- Generating realistic shots that are consistent with the original scene.
- Handling most common photographic and artistic styles.

The model may struggle with:
- Action-oriented prompts (e.g., "punching", "running").
- Uncommon or highly abstract styles.

## Training Data

The `InScene` LoRA was trained on 394 image pairs. This dataset was created by extracting and enriching frames from the WebVid dataset.

You can find the public dataset used for training here:
[https://huggingface.co/datasets/peteromallet/InScene-Dataset](https://huggingface.co/datasets/peteromallet/InScene-Dataset)