![Portada](https://www.modelscope.ai/models/furaidosu/tadv-bg-qwen-image-2512/resolve/master/images/portada.png)

## Overview

This LoRA generates stylized background images with exaggerated, distorted perspective, designed for illustrations and video games.

I originally developed it to recreate the kind of environments commonly seen in point-and-click adventure games — scenes that feel dynamic, playful, and slightly surreal while remaining readable and functional as interactive backgrounds.

## Background

Several years ago, while illustrating backgrounds for a point-and-click game, I began exploring image generation models to accelerate my workflow.

I experimented with training LoRAs using my own illustrations as well as artwork from other artists whose style I admire. However, I could not achieve the exact result I was looking for — particularly in terms of richer detail, stronger perspective deformation, and textured surfaces.

## Development Process

I started with rudimentary LoRAs that produced interesting results only under specific prompts. Using those models, I generated hundreds of images with wide variation in color palettes, moods, lighting conditions, interiors, and outdoor scenes.

To push the perspective further, I also trained several auxiliary LoRAs for Qwen-Image that introduce optical distortions such as fisheye effects and 360-degree panoramic warping. By mixing these distortion LoRAs with style LoRAs, I was able to produce highly dynamic compositions with curved geometry and spatial exaggeration.

I then manually curated and edited these images. Additional details and textures were added using a custom Qwen-Image-Edit-2511 model I built specifically to enrich flat vector-style artwork. I also refined images further with assistance from ChatGPT and Nano Banana.

## Training

The final dataset used to train this model was entirely synthetic and assembled from images generated through combinations of:

* Style LoRAs
* Perspective-distortion LoRAs (fisheye and 360 panorama)
* Manual editing and enhancement
* Texture augmentation passes

Using this curated dataset, I trained the present model, which successfully produces the distinctive visual style I had been aiming for.

## Style and Inspiration

The style is inspired by background art from classic adventure games such as:
* Sam & Max Hit the Road
* Day of the Tentacle
* Looney Tunes environments

while introducing a modern twist through richer textures, stronger depth cues, and more pronounced spatial distortion.

Key characteristics of the generated images include:

* Distorted, exaggerated perspective
* Fisheye-like or panoramic spatial curvature
* Curved or winding paths
* Slightly twisted or elastic-looking objects
* Playful, stylized geometry
* Textured surfaces with depth
* Compositions suitable for interactive scenes

The result is a distinctive background style that feels lively, handcrafted, and ideal for stylized storytelling, animation, or game worlds.
---
#### Model files may be downloaded with ModelScope SDK or through git clone directly.

Download with ModelScope’s Python SDK
```bash
#Install ModelScope
pip install modelscope
```
```python
#Download with ModelScope’s Python SDK
from modelscope import snapshot_download
model_dir = snapshot_download('furaidosu/tadv-bg')
```
Download with Git clone
```
git clone https://www.modelscope.ai/furaidosu/tadv-bg.git
```

<p style="color: lightgrey;">If you are a contributor to this model, we invite you to promptly update the model card content according to <a href="https://www.modelscope.ai/docs/contribute/model-integration" style="color: lightgrey; text-decoration: underline;">the model contribution documentation</a>.</p>



Files with SHA256: 241a02d73a7844137a33c6fbdc135231192fe666c5fde68f79f639dc063d1fde