# UniWorld-V2: Next-Level Precision Image Editing with Reinforcement Learning
A **new model**, even better than Google’s Nano Banana at **fine-detail image editing** and with **stronger Chinese understanding**, has arrived — introducing **UniWorld-V2**.
---
## 📌 Quick Comparison Example
**Task:**
“Change the hand gesture of the girl in the middle wearing a white shirt and mask to an OK sign.”
**Original image:**
**UniWorld-V2 result — perfect execution:**
**Nano Banana result — failed interpretation:**
---
## 🏗 Technical Background
Behind UniWorld-V2 is the **UniWorld team** from **RabbitShow Intelligence & Peking University**, with a **new post-training framework** called **UniWorld-R1**.
**Key innovation:**
- **First-time application of RL policy optimization** to a unified image-editing architecture — making it the **first visual reinforcement learning framework**.
- Built upon this, UniWorld-V2 was released, achieving **state-of-the-art performance**.
Benchmarks like **GEdit-Bench** and **ImgEdit** show UniWorld-V2 outperforming even **OpenAI’s GPT-Image-1**.
---
## 🌏 Real-World Capabilities Beyond Supervised Fine-Tuning (SFT)
### **1. Chinese Font Mastery**
- Precise interpretation of instructions.
- Renders **complex artistic Chinese fonts** such as “月满中秋” and “月圆人圆事事圆” with **clear strokes** and **accurate meaning**.
Examples:
Modify any character via prompt:
---
### **2. Fine-Grained Spatial Control**
- Handles “red-box control” tasks.
- Executes edits **only within specified areas** — even difficult ones like “move the bird out of the red box.”
Example:
---
### **3. Global Lighting Integration**
- Understands commands like “relight the scene.”
- Adjusts **lighting cohesively across the environment**.
Example:
---
## 🧠 UniWorld-R1: Core Framework
### 🚫 Problems with Traditional Methods
- SFT models tend to **overfit training data**, reducing generalization.
- Lack of a **general reward model** for diverse editing tasks.
### 💡 UniWorld-R1 Innovations
1. **RL-based unified architecture**
- First image-editing post-training framework using RL policy optimization.
- Employs **Diffusion Negative-aware Finetuning (DiffusionNFT)** — a likelihood-free method that trains efficiently.
2. **MLLM as training-free reward model**
- Uses multimodal LLMs (e.g., GPT-4V) for **task-agnostic evaluation** without extra training.
- MLLM output logits provide **fine-grained feedback**.
**Pipeline:** Sampling → MLLM scoring → DiffusionNFT
---
## 📚 Dataset Overview
- Compiled **27,572 instruction-based image editing samples**.
- Sources: **LAION**, **LexArt**, **UniWorldV1**.
- Added **text editing** and **red-box control** tasks → **9 task types**.
---
## 🏆 Benchmark Results
### GEdit-Bench
- **UniWorld-V2**: **7.83**
- Outperforms GPT-Image-1 [High]: 7.53
- Outperforms Gemini 2.0: 6.32
### ImgEdit
- **UniWorld-V2**: **4.49** → Best among open-source & closed-source.
---
**Framework generality:**
When applied to **Qwen-Image-Edit** and **FLUX-Kontext**, UniWorld-R1 boosted both models’ scores significantly.
---
### Category Gains
- UniWorld-FLUX.1-Kontext: Big gains in **Adjust**, **Extract**, **Remove**.
- UniWorld-Qwen-Image-Edit: Excelled in **Extract**, **Blend**.
---
### Out-of-Domain GEdit-Bench Performance
- FLUX.1-Kontext ↑ 6.00 → 6.74 (beats Pro version: 6.56)
- Qwen-Image ↑ 7.54 → 7.76
- UniWorld-V2: New state-of-the-art.
---
### 👥 Human Preference Study
- Participants preferred **UniWorld-FLUX.1-Kontext** for **instruction alignment** over FLUX.1-Kontext [Pro].
- Image quality slightly favored the official Pro, but editing capability was better with UniWorld-R1.
---
## 📜 Model Evolution
- UniWorld-V1: Industry’s **first unified understanding-and-generation model** — released **3 months before Nano Banana**.
- UniWorld-R1 **paper, code, and models** are **open-source**:
- GitHub: https://github.com/PKU-YuanGroup/UniWorld
- Paper: https://arxiv.org/abs/2510.16888
---
## 🔗 AiToEarn Integration
**[AiToEarn官网](https://aitoearn.ai/)** — an open-source global platform for:
- AI content generation
- Cross-platform publishing
- Monetization & analytics
**Supported platforms**: Douyin, Kwai, WeChat, Bilibili, Rednote, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X (Twitter).
More resources:
- Blog: [AiToEarn博客](https://blog.aitoearn.ai)
- Open source: [AiToEarn开源地址](https://github.com/yikart/AiToEarn)
- Model ranking: [AI模型排名](https://rank.aitoearn.ai)
---
> **End of Report**
