[{"data":1,"prerenderedAt":51},["Reactive",2],{"news-story-news-1792426066996-krlfi":3},{"id":4,"title_zh":5,"title_en":6,"summary_zh":7,"summary_en":8,"takeaways_zh":9,"takeaways_en":14,"content_zh":19,"content_en":20,"external_links":21,"source":28,"category":35,"impact_score":36,"views_count":37,"likes_count":38,"related_entities":39,"tags":43,"image_url":50},"news-1792426066996-krlfi","LMSYS Arena 开启 OpenAI GPT-Image-2.5 Sunburst 限时直接测评：图像生成与多轮编辑三榜登顶","LMSYS Arena Enables 14-Day Direct Mode Testing for OpenAI GPT-Image-2.5 Sunburst, Ranked #1 Across All Image Arenas","LMSYS 官方宣布在竞技场（Arena）中开放 OpenAI 旗舰图像模型 GPT-Image-2.5 Sunburst 的专属“直接测试模式”（Direct Mode），限时测试 14 天（至 11 月 2 日）。该模型当前在 LMSYS 文生图（Text-to-Image）、单图编辑（Image Edit）及多图编辑（Multi-Image Edit）三大竞技场全部霸榜第一。测试期结束后将继续保留在匿名双盲对战与 Agent 模式中，用于持续校准真实世界 Elo 积分。","LMSYS officially announced a 14-day exclusive direct testing window for OpenAI flagship vision generative model GPT-Image-2.5 Sunburst in Direct Mode (active until Nov 2, 8 a.m. PT). Currently holding the undisputed #1 ranking across LMSYS Text-to-Image, Image Edit, and Multi-Image Edit arenas, the model can now be directly chosen from the dropdown menu before returning exclusively to anonymous Battle and Agent Mode.",[10,11,12,13],"LMSYS 宣布在竞技场中开放 OpenAI GPT-Image-2.5 Sunburst 的直接测试模式（Direct Mode），限时测试 14 天至 11 月 2 日","该模型当前横扫 LMSYS 文生图（Text-to-Image）、单图编辑（Image Edit）与多图编辑（Multi-Image Edit）三大竞技场全部斩获第一","在复杂空间构图、长文本排版与多轮编辑的主体光影一致性上保持极高保真度，有效遏制了生成式编辑常见的风格漂移","测试期结束后模型将继续保留在双盲对战与 Agent 模式中，用于持续校准全球大模型真实世界 Elo 积分",[15,16,17,18],"LMSYS Arena enabled 14-day exclusive Direct Mode testing for OpenAI GPT-Image-2.5 Sunburst until November 2 at 8 a.m. PT","The model holds the undisputed #1 ranking across Text-to-Image, Image Edit, and Multi-Image Edit Arenas","Demonstrates superior spatial reasoning and ambient lighting consistency across multi-image composition, mitigating visual drift","Following Direct Mode, the model will remain anonymously active in Battle and Agent modes to power Elo ratings","### 核心背景与行业痛点\n在多模态视觉生成与交互式图像编辑领域，大模型评测长期依赖容易受审美偏置影响的人工抽检或静态合成指标（如 FID、CLIP-Score）。然而，在真实多轮图生图、精细局部修改和多图一致性融合等实际生产需求中，静态跑分往往无法反映真实用户体验。LMSYS 盲测竞技场通过匿名双盲投票建立的真实世界 Elo 积分体系，已成为全球大模型生成能力公认的试金石。针对 OpenAI 此前低调发布的最新图像基座，LMSYS 团队今日正式宣布在竞技场提供限时 14 天的 `GPT-Image-2.5 Sunburst` 原生直测支持。\n\n### 架构亮点与底层机制\n1. **三冠王视觉基座（Triple-#1 Visual Powerhouse）**：\n   - GPT-Image-2.5 Sunburst 在 LMSYS **文生图（Text-to-Image）**、**单图编辑（Image Edit）** 以及 **多图联动编辑（Multi-Image Edit）** 三大专业竞技场中全面位列全球第一；\n   - 具备卓越的指令遵循与多轮视觉上下文保持能力，能够在保留输入图像核心光影、主体结构与微观材质的前提下执行复杂空间修改；\n2. **限时直选模式与双盲 Elo 体系分离**：\n   - 开启为期 14 天的 Direct Mode（直接模式），开发者可在下拉菜单中直接选定 GPT-Image-2.5 Sunburst 进行任意复杂提示词测试与编辑流水线调试；\n   - 11 月 2 日上午 8 点（太平洋时间）直选期结束后，该模型将继续常驻在双盲 Battle 模式与 Agent 模式中，其匿名盲测投票数据将继续支撑全球权威 Elo 排行榜的动态校准。\n\n### 权威 Benchmark 与实测跑分对比\n根据 LMSYS 官方公告与竞技场历史胜率统计（Figure 1, Arena Vision Leaderboards）：\n- **文生图竞技场（Text-to-Image Arena）**：\n  - GPT-Image-2.5 Sunburst 位居总榜第一，在复杂空间关系提示（Spatial Composition）、文本嵌入排版（Typography）和微距真实质感（Photorealism）测试中胜率全面领先竞品；\n- **图像编辑与多图编辑竞技场（Image Edit & Multi-Image Edit Arena）**：\n  - 稳居单图编辑与多图协同编辑双料冠军；\n  - 能够精准解析“在保留原图人物姿态与光影角度的前提下将背景置换为赛博朋克雨夜”或“将图 A 中的物体以物理透视合理融合进图 B”等高难度操作，极大缓解了扩散模型常见的局部坍缩与风格漂移（Visual Drift）；\n- **社区反馈与真实胜率**：\n  - 在盲测对抗中展现出显著高于常规扩散模型的长提示理解鲁棒性，在超过 9,000 次曝光与社区盲测交互中保持胜率第一梯队。\n\n### 开发者实战落地与开箱指南\n- **限时直测入口**：开发者即日起可登录 `arena.lmsys.org`，进入 Direct Mode 下拉列表直接选择 `GPT-Image-2.5 Sunburst`，无需进入随机盲测池即可直接测试；\n- **提示词工程建议**：测试表明，GPT-Image-2.5 对自然语言复合意图（如“先保留左侧阴影，再替换右侧材质”）的解析力远胜传统基于特定 Tag 的控制方式。开发端侧创意工作流或电商品宣生成工具的团队，建议抓紧 14 天窗口期建立金标测试集与自动化评估管线。","### 核心背景与行业痛点\nIn multimodal visual generation and conversational image editing, benchmark evaluation has long suffered from aesthetic subjectivity and unrepresentative static metrics (such as FID or CLIP scores). In real-world enterprise scenarios—such as multi-turn inpainting, local semantic refactoring, and multi-asset composition—static scores fail to capture human creative preference. LMSYS Arena's blind pairwise Elo evaluation has become the gold standard for measuring generative models. To give the community transparent access to OpenAI's latest visual engine, LMSYS announced a 14-day dedicated Direct Mode testing period for `GPT-Image-2.5 Sunburst`.\n\n### 架构亮点与底层机制\n1. **Undisputed Triple-#1 Leaderboard Status**:\n   - GPT-Image-2.5 Sunburst currently holds the #1 ranking across three distinct LMSYS leaderboards: Text-to-Image Arena, Image Edit Arena, and Multi-Image Edit Arena;\n   - Demonstrates unmatched instruction following and visual consistency, retaining underlying illumination, fine textures, and perspective geometry across sequential modifications;\n2. **Direct Mode Access Decoupled from Anonymous Elo Scoring**:\n   - For 14 days, developers can select GPT-Image-2.5 Sunburst directly from the dropdown menu to test complex prompts, edge cases, and multi-image transformations;\n   - Following the conclusion of Direct Mode on November 2 at 8 a.m. PT, the model will remain accessible anonymously in Battle and Agent modes to power real-world Elo leaderboards.\n\n### 权威 Benchmark 与实测跑分对比\nAccording to official LMSYS announcements and live arena win-rate telemetry:\n- **Text-to-Image Arena**:\n  - Holds the overall #1 Elo ranking, demonstrating decisive win margins on complex spatial composition, typographic rendering, and photorealistic physics;\n- **Image Edit & Multi-Image Edit Arenas**:\n  - Ranks #1 in both single-turn surgical editing and multi-image asset synthesis;\n  - Faithfully executes difficult semantic interventions—such as swapping backgrounds while preserving subject geometry or seamlessly compositing objects from Image A into Image B under matching ambient lighting—effectively eliminating severe visual drift;\n- **Community Adoption**:\n  - The announcement garnered thousands of blind test votes within hours, cementing its benchmark dominance across professional creative prompts.\n\n### 开发者实战落地与开箱指南\n- **Direct Access URL**: Developers can immediately navigate to `arena.lmsys.org`, switch to Direct Mode, and select `GPT-Image-2.5 Sunburst` from the model menu;\n- **Workflow & Evaluation Recommendation**: Given the 14-day direct testing window, teams developing automated design pipelines, photo restoration tools, or creative agents should run comparative test suites to benchmark prompt sensitivity and compositing fidelity against local open-source diffusion models.",[22,25],{"title":23,"url":23,"type":24},"https://arena.lmsys.org","demo",{"title":26,"url":26,"type":27},"https://openai.com","other",{"platform":29,"author_name":30,"author_handle":31,"author_avatar":32,"tweet_url":33,"published_at":34},"twitter","LMSYS Org","@arena","https://unavatar.io/x/arena","https://x.com/arena/status/2112199487217434855","2026-10-19T16:07:46.996Z","Models",93,40,0,{"models":40,"agents":41,"vendors":42},[],[],[],[44,45,46,47,48,49],"OpenAI","GPT-Image-2.5","LMSYS","Arena","Direct Mode","Image Edit","https://pbs.twimg.com/media/HVAJuOmaYAAQCxR.jpg",1792876591232]