比NanoBanana更擅长中文和细节控制，兔展&北大Uniworld V2刷新SOTA

Core Insights - The article introduces UniWorld-V2, a new image editing model that excels in detail and understanding of Chinese language prompts, outperforming existing models like Nano Banana [1][6][24] - The model is developed by the UniWorld team from Tujia Intelligent and Peking University, utilizing a novel training framework called UniWorld-R1, which incorporates reinforcement learning for enhanced image editing capabilities [5][14] Model Performance - UniWorld-V2 achieved state-of-the-art (SOTA) results in industry benchmark tests such as GEdit-Bench and ImgEdit, surpassing top closed-source models like OpenAI's GPT-Image-1 [5][19] - In GEdit-Bench, UniWorld-V2 scored 7.83, significantly higher than GPT-Image-1's 7.53 and Gemini 2.0's 6.32 [19][21] - The model also led in ImgEdit with a score of 4.49, outperforming all known open-source and closed-source models [19] Technical Innovations - The core innovation of UniWorld-R1 is its use of reinforcement learning for image editing, addressing issues of overfitting and generalization found in traditional supervised fine-tuning models [14] - UniWorld-R1 employs Diffusion Negative-aware Finetuning (DiffusionNFT) for efficient training and utilizes a multi-modal large language model (MLLM) as a reward model, enhancing alignment with human intent [14][15] User Experience and Functionality - UniWorld-V2 demonstrates superior control in editing tasks, such as accurately rendering complex Chinese characters and executing precise spatial edits based on user-defined areas [9][10][11] - The model's ability to understand and apply lighting adjustments to scenes results in a more cohesive and harmonious image output [12] Generalization and Versatility - The UniWorld-R1 framework shows strong generalization capabilities, improving the performance of other models like FLUX.1-Kontext and Qwen-Image-Edit when applied [20][22] - User preference studies indicate that UniWorld-FLUX.1-Kontext is favored for its instruction alignment and editing capabilities, despite slightly lower image quality compared to stronger official versions [23] Availability and Future Research - The UniWorld-R1 framework, along with its paper, code, and models, has been made publicly available on GitHub and Hugging Face to support further research in the field [24]