--- title: tadv-bg-qwen-image-2512 canonical_url: "https://www.modelscope.ai/models/furaidosu/tadv-bg-qwen-image-2512" md_url: "https://www.modelscope.ai/models/furaidosu/tadv-bg-qwen-image-2512.md" repository: furaidosu/tadv-bg-qwen-image-2512 last_updated: 2026-02-17 license: "Apache License 2.0" pipeline_tag: text-to-image-synthesis tasks: - text-to-image-synthesis base_model: - Qwen/Qwen-Image-2512 base_model_relation: adapter parameters: 118.0M tensor_type: - BF16 library_name: - lora - safetensors supports_inference: txt2img downloads: 18 stars: 0 tags: - LoRA - text-to-image --- # tadv-bg-qwen-image-2512 > tadv-bg-qwen-image-2512 - An open-source model by furaidosu on ModelScope. I trained this model on synthetic data. The images from the dataset where generated using previous models and edited on photoshop. Then I upscaled and added details using a custom… furaidosu/tadv-bg-qwen-image-2512 is a 118.0M-parameter text-to-image-synthesis model on ModelScope. licensed under Apache License 2.0. derived from Qwen/Qwen-Image-2512. and supports online inference (txt2img). - **Repository**: furaidosu/tadv-bg-qwen-image-2512 - **License**: Apache License 2.0 - **Tasks**: text-to-image-synthesis - **Parameters**: 118.0M - **Base model**: Qwen/Qwen-Image-2512 - **Online inference**: txt2img - **Tags**: LoRA, text-to-image - **Downloads**: 18 - **Stars**: 0 - **Last updated**: 2026-02-17 Source: https://www.modelscope.ai/models/furaidosu/tadv-bg-qwen-image-2512 --- ![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 ```

If you are a contributor to this model, we invite you to promptly update the model card content according to the model contribution documentation.