---
title: Omni-Section-V2-Clean
canonical_url: "https://www.modelscope.ai/models/deepeshahlawat/Omni-Section-V2-Clean"
md_url: "https://www.modelscope.ai/models/deepeshahlawat/Omni-Section-V2-Clean.md"
repository: deepeshahlawat/Omni-Section-V2-Clean
last_updated: 2026-03-16
license: "Apache License 2.0"
pipeline_tag: text-to-image-synthesis
tasks:
  - text-to-image-synthesis
base_model:
  - "MusePublic/Qwen-image@v1"
base_model_relation: adapter
parameters: 1.9B
tensor_type:
  - BF16
library_name:
  - lora
  - safetensors
  - pytorch
frameworks:
  - pytorch
supports_inference: txt2img
downloads: 201
stars: 1
tags:
  - LoRA
  - text-to-image
---

# Omni-Section-V2-Clean

> Omni-Section-V2-Clean - An open-source model by deepeshahlawat on ModelScope. Omni-Section: Z-Axis Geometric Reasoning LoRA

deepeshahlawat/Omni-Section-V2-Clean is a 1.9B-parameter text-to-image-synthesis model on ModelScope. licensed under Apache License 2.0. derived from MusePublic/Qwen-image@v1. and supports online inference (txt2img).

- **Repository**: deepeshahlawat/Omni-Section-V2-Clean
- **License**: Apache License 2.0
- **Tasks**: text-to-image-synthesis
- **Parameters**: 1.9B
- **Base model**: MusePublic/Qwen-image@v1
- **Online inference**: txt2img
- **Tags**: LoRA, text-to-image
- **Downloads**: 201
- **Stars**: 1
- **Last updated**: 2026-03-16

Source: https://www.modelscope.ai/models/deepeshahlawat/Omni-Section-V2-Clean

---

# Omni-Section: Z-Axis Geometric Reasoning LoRA

## 🚀 Overview
**Omni-Section** is a commercial-grade Low-Rank Adaptation (LoRA) fine-tuned on the **Qwen-Image 2.0** architecture. It introduces **Z-axis geometric reasoning** to the generative pipeline, enabling the instant generation of precise 3D spatial disaggregation (exploded views) of complex objects.

Unlike standard diffusion fine-tunes that treat "exploded views" as stylistic noise, Omni-Section enforces **structural coherence**, isolating individual components while maintaining photorealistic material fidelity.

## 📈 Impact & Utility
*   **Rapid Prototyping:** Reduces 3D modeling/rendering workflows from hours to seconds.
*   **Industrial Design:** Provides alpha-ready, text-free 8K assets for product manuals, technical brochures, and marketing visualization.
*   **Geometric Intelligence:** Bridges the gap between aesthetic generation and structural logic, proving Qwen-Image 2.0's latent space is capable of complex mechanical understanding.

## 🛠 Training Methodology
*   **Base Model:** Qwen-Image 2.0
*   **Dataset:** 16 high-resolution, text-free, 3D industrial exploded view renders.
*   **Resolution:** 1024x1024 (native training resolution).
*   **Hyperparameters:** Rank 64 / Alpha 64 / 3,200 Steps.
*   **Data Purity:** Dataset was scrubbed of all typographic annotations to force the model to focus 100% of its capacity on spatial geometry and material texture.

## 📊 Performance Verification (A/B Test)
| Prompt | Base Qwen-Image | Omni-Section LoRA |
| :--- | :--- | :--- |
| **Nvidia RTX 6090** | Generic aesthetic render | Structurally coherent 3D diagram |
| **Modular House** | Non-functional house | Functional, assembly-ready modular render |

<div align="center">
 <img src="canva1.png" >
</div>
<div align="center">
 <img src="canva2.png" >
</div>
<div align="center">
 <img src="canva3.png" >
</div>

## 🔗 How to Use
Use the trigger word: `omni-section style`
**Recommended Settings:**
- **LoRA Weight:** 0.85
- **CFG Scale:** 7.0
- **Negative Prompt:** `text, words, labels, diagram lines, pointers, watermark, messy, merged components`

## 📁 Dataset
You can explore the clean, text-free dataset used to train this model

---
*Built for the Qwen-Image LoRA Training Competition (Track 1: AI for Production).*