OpenAI Launches ImageGen 2.0 to Drive AI Productivity Over Haul

OpenAI推出ImageGen 2.0推动AI生产力变革 该平台通过支持多图一致性、精准文字渲染以及与Codex代码生成的直接集成，使图像生成从休闲娱乐转变为关键的工作场所工具。 讨论详细阐述了ImageGen 2.0带来的巨大范式转变，重点指出了其在文字渲染、多语言支持、逼真质感以及原生画幅比例灵活性方面的突破。嘉宾们概述了该工具如何从新奇的创意载体转变为强大的生产力资产，并广泛应用于信息图表、教育内容和内部演示中。他们还探讨了ImageGen与Codex结合在游戏和应用设计中产生的强大协同效应、针对长篇叙事增强的角色一致性，以及该技术未来向高度个性化“创意智能体”的演变。 ImageGen 2.0 发布后两周内周使用量激增超 50%，助力 ChatGPT 平台每周图像生成量突破 15 亿张，并在全球引发多样化的视觉创作潮流。 (0:00) 该模型在文本渲染、多语言支持与写实逼真度上实现重大跨越；网格扩展测试显示其能稳定渲染超 100 个独立物体，远超 DALL-E 3 不足 10 个的表现。 (3:16) 原生支持任意宽高比带来了 360 度全景渲染等新功能；团队在后训练中注重人类审美品味与 Token 效率，使模型不仅能精准输出学术级生物图表，也能智能模拟人类特有的“不完美”复古质感。 (7:31) 随着模型从提示词工程迈向“创意智能体（Creative Agent）”，ImageGen 与 Codex 的深度整合已支持零样本（Zero-shot）应用构建与游戏精灵表（Sprite sheet）开发，并大幅改善了跨多图的角色一致性。 (25:19)

OpenAI product lead Adele Li and researcher Kenji Hata argue that ImageGen 2.0 marks a transition for generative vision from an experimental novelty into a structured productivity medium, driven by advances in photorealism, layout comprehension, and cross-system workflows. Scaling Model Capacity and Visual Fidelity (3:16) The rollout of ImageGen 2.0 precipitated a sharp increase in user activity, with weekly usage expanding by more than 50% in the two weeks following launch to reach over 1.5 billion generated images per week on ChatGPT. Technical gains are evident across multilingual text rendering, photorealism, and geometric handling. To quantify underlying intelligence scaling, Kenji highlights an internal grid benchmark measuring how many discrete objects a model can correctly depict: while DALL-E 3 reliably managed 5 to 8 objects and ImageGen 1.5 handled 25 to 36, ImageGen 2.0 exceeded 100 objects. The architecture also unlocked emergent layout flexibility by natively supporting arbitrary aspect ratios without requiring fixed dimensions. This native capability enabled users to generate extreme panoramic compositions and narrow vertical formats, prompting the release of a 360-degree rendering mode. Concurrently, post-training optimization prioritized token efficiency and aesthetic alignment, training the system to render high-fidelity human anatomy as well as intentionally degraded styles like crayon drawings and low-fidelity digital sketches. Creative Direction and Taste Over Prompt Engineering (17:08) Interaction design has shifted away from rigid prompt engineering toward genuine creative direction. Early systems rewarded esoteric prompt phrasing, but modern workflows increasingly mirror professional design critiques, benefiting from domain-specific vocabulary and reference-based guidance. By consulting professional artists, designers, and marketers, the development team integrated reference-upload features that allow users to anchor technical variables—such as depth of field and color palette—directly to source imagery. This architectural flexibility addresses the reality that most consumer prompts are inherently ambiguous, such as requests to make a subject look better. Personal evaluation suites, including Li's benchmark of 100 family and personal photographs, test whether the system can correctly maintain facial identity, individual context, and nuanced human preferences. The guests emphasize that steering modern generative systems depends far more on creative judgment and contextual reference than on mechanical prompt syntax. Productivity Adoption in Education and Enterprise (18:59) Image generation has broadened from casual entertainment into functional enterprise and educational applications. An alpha testing channel for educators demonstrated that the model could accurately illustrate complex scientific concepts, including graduate-level biological diagrams submitted by university faculty. This capacity to parse dense technical concepts and organize visual hierarchies enables automated generation of accurate study materials and infographics. Internally at OpenAI, this structural capability has altered daily operations: over 50% of the slides in internal presentation decks are now generated with ImageGen. Rather than merely synthesizing decorative imagery, the system optimizes composition by determining spatial layouts and legibility. Commercial adoption reflects this functional shift, with real estate agents staging properties, content creators generating video thumbnails, and educators replacing static posters with custom diagrams. Visual Consistency and Multi-Tool Synthesis (25:19) Historic barriers around character drift and stylistic discontinuity have lessened through improved multi-image coherence. Creators can generate character model sheets depicting consistent figures across multiple poses and re-feed those assets back into the model, supporting multi-page narratives such as ten-page comic books and multi-slide corporate decks without visual divergence. Simultaneously, the combination of ImageGen with Codex supports zero-shot application and video game development. Developers can prompt the image model to produce multi-panel contact sheets or sprite sheets, which Codex immediately imports into code to build functional software interfaces and interactive assets. For complex compositional tasks, the guests note that routing requests through pro or thinking models allows the system to deliberate, run web searches, and inspect uploaded files before rendering. OpenAI aims to evolve ImageGen into collaborative 'creative agents' capable of handling end-to-end design, architecture, and event planning tasks, though direct in-canvas image editability and completely seamless visual persistence remain active areas of ongoing development.

OpenAI 研发人员与产品负责人指出，ImageGen 2.0 的推出标志着图像生成技术正在从早期的娱乐探索跃升至融合多学科知识的“文艺复兴”阶段。该模型在网格对象渲染、文本保真度及审美对齐上取得重大突破，正加速融入专业工作流并逐步向协作型创意代理演进。 核心能力跃迁与规模化生产力落地 (3:16) 根据产品负责人 Adele Li 和研究员 Kenji Hata 披露的数据，ImageGen 2.0 发布后两周内，ChatGPT 上的图像生成周使用量增长了 50% 以上，目前每周生成的图像总数已突破 15 亿张。在技术表现上，模型实现了跨越式的代际提升，尤其体现在消除面部与肢体畸变、大幅提升文本渲染准确性，以及原生支持多语言能力，使其在亚洲和欧洲市场迅速引发共鸣。Kenji Hata 指出，图像生成已正式跨出“好玩但无用”的范畴，步入严肃的生产力领域。 在评估模型智能水平的网格缩放测试中，研发团队展示了模型在单一网格内正确渲染随机对象数量的阶梯式进展：早期的 DALL-E 3 仅能生成 5 到 8 个对象，ImageGen 1 约为 16 个，ImageGen 1.5 稳定在 25 到 36 个，而 ImageGen 2.0 在内部测试中已能成功渲染超过 100 个对象。这种信息承载能力的大幅提升，为信息图表、复杂学术图示等高密度视觉内容的生成奠定了确定性的技术基础。 交互范式转变与人类审美对齐 (17:08) 在交互层面，传统的“提示词工程”正被更具专业性的“创意指导”所取代。团队观察到，拥有艺术背景的用户即便缺乏提示词技巧，也能凭借其专业词汇与审美素养获得更惊艳的效果。为此，开发团队直接吸收专业艺术家与设计师的工作流经验，允许用户直接上传参考图像注入上下文，使模型能够精确捕获参考素材的“神韵”并控制景深等技术参数。针对大量用户习惯输入极度模糊提示词的现实，模型被设计为能够自主推导并提炼这些模糊需求。 后训练阶段的核心任务是向模型灌输人类对美感与真实性的主观判断。为了兼顾生成速度与计算开销，团队重点优化了 Token 效率。有趣的是，高度智能化的模型还催生了一种“刻意追求不完美”的文化潮流，例如模仿微软画图、蜡笔涂鸦或复古低保真图像。Adele Li 强调，能够精准生成这种“粗糙感”恰恰证明了模型具备极高的智能深度，反映了用户对真实感、怀旧与个性化表达的需求。 跨学科知识整合与软硬件工作流协同 (18:59) ImageGen 2.0 展现出在不同视觉风格与学科之间自如切换的能力，涵盖从建筑工程图纸到儿童绘本插画的广泛跨度。在面向教育工作者的阿尔法测试中，一位生物学教授上传了研究生水平的高难度教科书概念，模型输出的科学准确性获得了“完全精准”的评价。视觉沟通的效率也在团队内部得到印证，目前开发团队内部演示文稿中超过 50% 的幻灯片均直接使用该模型进行构图与绘制。 此外，该模型与编程模型 Codex 的协同效应正深刻重塑应用与游戏开发流程。用户可利用 ImageGen 一次性生成多方案设计样本（contact sheets），并无缝由 Codex 转化落地为功能性网站或应用代码；独立游戏开发者则广泛利用该功能自动生成一致的精灵图（sprite sheets）。借助对先前视觉输出的稳定引用能力，创作者得以生成涵盖多姿态的角色设定表，解决了长期困扰行业的多图一致性难题，从而支持制作连续 10 页的长篇连环画或幻灯片。 原生比例适配与未来创意代理构建 (22:15) 在底层架构设计阶段，团队即进行了风险对冲训练，使模型原生支持任意长宽比生成，不仅支持了超长全景画和书签等特殊构图，更直接催生了 ChatGPT 网页与移动端上线的 360 度全景渲染功能。对于高阶用户，团队建议通过 Thinking 或 Pro 模式开启深度推理功能，借助系统自主联网搜索、解析上传文档与调用工具的能力，以开放式的提示引导模型自主优化构图。 展望未来，团队的目标是超越静态图像生成，迈入“创意代理（Creative Agent）”时代。该代理旨在成为能够深刻理解个人偏好与流程意图的协作助手，兼具室内设计师、建筑师乃至婚礼策划师的角色。产品演进的后续路线图将集中于进一步增强构图逻辑、提升风格多样性，并最终在产品内实现对已生成图像的直接原生编辑能力。 尽管 ImageGen 2.0 在科学准确性、多图一致性以及跨工具协同方面展现了显著的生产力价值，但开发团队明确指出，系统目前尚未达到完美状态。未来的关键挑战在于如何在保持模型自主推理与美学直觉的同时，提供更精细的局部编辑交互能力，并切实解决图像在复杂长序列创作中可能出现的细微偏差，从而实现向全功能个人创意代理的平稳过渡。

## Background 本期内容由 OpenAI 官方频道于 2026 年 6 月 3 日发布在 YouTube 平台上。视频简介指出，当前 ChatGPT 用户每周生成的图像量已超过 15 亿张，本期播客由主持人 Andrew Mayne 携手产品负责人 Adele Li 与研究员 Kenji Hata，共同探讨 Images 2.0 发布以来的技术跃迁、病毒式传播趋势，以及从 DALL-E 时代演进至具备写实感、文字渲染、多语言支持和角色一致性的下一代创意助理。 ## TL;DR 节目全面梳理了 Images 2.0 模型带来的“文艺复兴”级演进，剖析了其在高保真文本排版、极致写实摄影、多语言支持以及单图容纳百个实体的世界认知等四大维度的核心突破。嘉宾们探讨了模型从娱乐涂鸦向专业生产力的转变，包括 360 度交互全景、严谨学术教材插图以及通过“我我我测试”验证的面部与上下文一致性。对谈还深入展示了图像模型与 Codex 编程智能体融合的革命性体验，用户可实现从 UI 概念图到前端代码、多动作精灵图和连环画角色的无缝生成，并前瞻了具备自主推理能力的未来“创意智能体”。 ## Atmosphere & style 整场对谈气氛热烈、松弛且洋溢着技术突破带来的由衷兴奋，主持人和一线研发人员仿佛围坐在一处轻松复盘一场重大的技术战役。对话在硬核架构演进与生动鲜活的研发趣闻之间游刃有余地切换——从早期自嘲模型拼写如同大猩猩乱涂，到亲眼见证模型首次质变生成海边女性照片时的惊喜共鸣，极富感染力。整期节目兼具极客探索的技术深度与人文审美的温情，生动传递出创作者重获工具赋能、重新审视“不完美之美”的情感共振。 ## Bulletin - [02:27](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=147s) OpenAI 推出 Images 2.0 模型，两周内推动 ChatGPT 图像生成周使用量激增超 50%，突破每周 15 亿张。 - [05:25](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=325s) 通过优化变量绑定与注意力机制，模型在单图独立实体渲染上限上从 DALL-E 3 的 5-8 个跃升至 100 个以上，实现高密度信息图与规整排版。 - [07:35](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=455s) Images 2.0 原生支持任意宽高比与 360 度全景图生成，ChatGPT 同步上线了原生全景交互查看器。 - [25:18](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=1518s) 图像模型正与 Codex 等代码智能体深度协同，支持零样本将 UI 原型直接转化为前端代码，并依托上下文参考能力解决长期困扰的角色一致性难题。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=0s) Images 2.0 被团队视为图像生成的“文艺复兴”时刻，标志着模型从早期的玩具探索阶段正式步入高保真、跨学科知识融合的生产力工具时代。 - [02:27](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=147s) 新版本核心攻关了四大维度短板：彻底解决乱码的高保真文本渲染、无缝多语言支持、消除假面感与畸变的摄影写实度，以及深度的物理世界认知融入。 - [05:25](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=325s) “百物测试”成为量化多实体渲染极限的标尺，Images 2.0 达成了单幅画面准确呈现 100+ 个随机互不重复实体的能力跨越。 - [07:35](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=455s) 任意画幅比例成为原生涌现能力，支持从超长书签到 360 度全景环视图像的沉浸式生成。 - [07:35](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=455s) 社交网络上风靡将高清照片降维生成蜡笔涂鸦或画图板风格，反映出人类对 AI 粗糙感、怀旧情绪与不完美真实感的深层情感渴望，而拟真不完美本身需要极高的模型智能。 - [12:22](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=742s) 后训练阶段重在平衡物理客观规律与人类主观审美品味，团队通过优化 Token 效率确保在不增加用户等待时间的前提下大幅提升画面精度。 - [14:06](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=846s) 团队采用包含身边亲友真实照片的“我我我测试”作为私人基准，既严格检验面部熟悉度与失真，又评估 ChatGPT 对用户个性化上下文的记忆整合能力。 - [16:32](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=992s) 面对大众模糊简短的口语化 Prompt，模型具备强意图推断与审美补偿能力；同时亦能精准理解专业艺术家的焦段、构图术语或参考图风格迁移。 - [18:50](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=1130s) 模型在教育科研领域展现高度生产力价值，能生成精确的研究生级生物教科书插页，在 OpenAI 内部更有超过 50% 的演讲幻灯片借助该模型制作。 - [22:16](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=1336s) 下一阶段愿景是打造“创意智能体”，能够像私人设计师般理解工作流意图，实现首发命中（First-shot）的多平台营销物料或房源置景生成。 - [25:18](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=1518s) Images 与 Codex 形成开发闭环：图像模型设计界面与游戏精灵图（Sprite sheet），代码智能体即时将其转化为可运行的前端软件应用。 - [25:18](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=1518s) 借助将角色设定图作为上下文回传机制，创作者可实现稳定连贯的角色跨场景输出，支持 10 页连环画或复杂分镜设计。 - [28:08](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=1688s) 在配合“思考模式（Thinking mode）”使用时，建议保持 Prompt 的开放性并明确锚定设计风格（如极简主义），让模型充分自主推理以避免信息过载。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=0s) 图像生成的“文艺复兴”时刻与团队背景 主持人 Andrew Mayne 开启了本期播客，邀请到 OpenAI 图像研究员 Kenji Hata 与产品负责人 Adele Li，共同探讨最新发布的 Images 2.0 模型。Andrew 将这次升级比作一场划时代的技术演进：“如果说 DALL-E 代表着图像生成的石器时代，那么 Images 2.0 就是它的文艺复兴。”新模型不仅在艺术审美上带来了飞跃，还将科学、艺术和建筑设计等跨领域知识融汇于单张画面之中。当团队内部第一次看到输出效果时，一致认定它彻底超越了第一代 ImageGen。 在谈及各自加入图像团队的经历时，Adele 分享了她从投资圈跨界到 AI 核心产品研发的历程。加入 OpenAI 之前，她从事了多年的私募股权投资，并在红点创投（Redpoint Ventures）专注于 AI 及软件领域的风险投资达三年之久。两年前加入 OpenAI 时，她起初负责的是数据与算力基础设施的建设，随后逐步转型为产品经理，并在过去六个月中全力牵头 ImageGen。在她看来，产品经理的职责就是“哪里需要解决问题就顶上去”，不仅要同顶尖研究员紧密协作，更要敏锐洞察市场变化。一年前推出 ImageGen 1.0 时行业环境相对单纯，而如今图像生成赛道玩家林立，ChatGPT 本身的产品生态也发生了巨变，重新定义图像生成在 ChatGPT 体系内的战略生态位极具挑战与活力。 研究员 Kenji Hata 同样在大约两年前加入 OpenAI。他最开始参与的是一个音频领域的探索性项目，在 ImageGen 1.0 正式发布前夕，他以协助者的身份参与其中，随后发现自己对图像研究越来越投入，最终全职扎根于该项目，陪伴了模型从初版到 2.0 的完整蜕变。 ### [02:27](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=147s) Images 2.0 的发布反响与四大能力跃升 Images 2.0 推出仅仅两周，用户的使用热度便迎来了爆炸式增长，整体使用量激增超过 50%，目前用户每周在 ChatGPT 中生成的图像已突破 15 亿张。全球范围内随之掀起了各具特色的病毒式传播狂潮：亚洲用户热衷于利用新模型进行“个人色彩诊断”（Color Analysis）以及制作个性化贴纸表情包；而美国社区则风靡用蜡笔涂鸦（Crayon and Scribble）风格重构经典画面。这种全球范围的共鸣展现了模型的极宽动态范围，也让普通用户无需复杂的提示词就能直观感知到画质、细节保真度与美学维度的巨大跃迁。 ```mermaid flowchart TD accTitle: Images 2.0 核心能力提升矩阵 accDescr: 展示 Images 2.0 在文本渲染、多语言支持、真实感以及世界认知四大维度的协同提升 subgraph CorePillars["Images 2.0 四大核心升级维度"] A["高保真文本渲染<br/>文字排版清晰且具真实语义"] B["深度多语言支持<br/>消除跨文化语种渲染壁垒"] C["极致摄影写实度<br/>告别人脸假面感与形体畸变"] D["全局世界认知融入<br/>深度理解复杂视觉构图与物理规律"] end CorePillars --> Output["赋能全场景生产力与前沿视觉表达"] ``` 面对这种如同跨越了范式般的巨大进步，Adele 阐述了团队立项之初确立的核心愿景：让图像生成超越目前的玩具阶段，实现能够将现实中任何视觉内容高度浓缩还原的跃迁。针对旧版模型的短板，团队重点攻克了四大核心维度： - **文本渲染（Text Rendering）**：彻底打破过去字符乱码的瓶颈，生成的页面文字不仅笔画清晰工整，而且在语义逻辑上完全通顺成词。 - **多语言能力（Multilingual）**：全面适配不同语言的输入与文本呈现，使得亚欧等全球各地非英语母语国家的用户获得丝滑体验。 - **摄影写实度（Photorealism）**：针对用户长期诟病的“假面感”或五官肢体畸变进行系统性修正，让生成的肖像和人物形象看起来更像真人本身。 - **世界认知（World Knowledge）**：将丰富的物理世界常识融入模型认知体系，通过深厚的先验知识精准将抽象概念具象化为高美学质量的视觉输出。 Kenji 补充强调，团队在整个迭代周期中高度重视社交媒体上的用户真实反馈，时刻捕捉模型在真实使用中的边界缺陷，并在新版本中针对性地实施定向抑制甚至彻底修复。 ### [05:25](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=325s) 从娱乐到生产力：文本渲染、信息图与“百物测试” 在研究团队看来，新模型带来的最深刻变革在于应用场景从“无目的的娱乐消遣”全面拓展至“严谨的生产力创作”。Kenji 指出，随着文字渲染精度的大幅提升，用户开始普遍使用该模型制作高密度的信息图表（Infographics）和带有精确说明的图文内容，这在以往是不可想象的。 针对文本与排版能力的突飞猛进，Andrew 调侃道，早期 AI 拼写“OpenAI”时表现得就像大猩猩在乱涂乱画，而现在它甚至能排出一整页细密规整的书页文字。Kenji 解释称，这种能力的跃升是底层“变量绑定”（Variable Binding）与注意力机制稳健进步的自然结果，并非毫无预兆的突变。 ```mermaid flowchart LR accTitle: 实体绑定与对象数量渲染演进图 accDescr: 展示从 DALL-E 3 到 Images 2.0 能够稳定渲染的独立对象数量提升轨迹 D3["DALL-E 3<br/>稳定呈现 5 - 8 个对象"] --> I1["Images 1.0<br/>提升至约 16 个对象"] I1 --> I15["Images 1.5<br/>稳定处理 25 - 36 个对象"] I15 --> I2["Images 2.0<br/>突破 100+ 个随机对象"] ``` Kenji 披露了团队内部衡量模型构图与绑定极限的一项基准测试：要求 GPT 生成包含 100 种不重复随机物体的长列表，随后直接输入图像生成模型，检查画面网格中究竟能正确绘制出多少个独立物体。从模型迭代的历史演进来看： - **DALL-E 3 时代**：画面在容纳 5 到 8 个不同物体时就会出现重叠、漏画或混乱。 - **Images 1.0**：稳定呈现数量提升至约 16 个。 - **Images 1.5**：已能稳定且高质地排列 25 到 36 个独立实体。 - **Images 2.0**：如今能够在单张画面中稳定画出 100 个以上彼此分明的物体，且准确率逼近 100%。 这一量化进展令 Andrew 深有感触。他回忆起早期在 Ada、Babbage 和 Curie 等远古模型上做过的类似测试——让模型列出 100 本科幻小说，往往数到第 22 本时模型就会耗尽上下文开始死循环重复；而如今的多模态模型在复杂的实体连续追踪上已经展现出惊人的稳定性。 ### [07:35](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=455s) 任意宽高比、360 度全景与真实感不完美热潮 除了复杂实体的处理能力外，Images 2.0 展现出了一项强悍的“涌现能力”（Emergent Capability）——对任意比例画面的原生支持。用户不再受制于传统的正方形或固定画幅，而是可以自由生成超细长书签甚至超宽幅的壮丽全景画。在这一特性的基础上，团队进一步开发了生成 360 度环形全景图的能力。由于交互效果极为惊艳，ChatGPT 团队火速在网页端和移动端上线了原生的 360 度全景交互式查看器，让用户能够拖动画面、沉浸式置身于生成的虚拟环境之中。Andrew 兴奋地分享了自己动手尝试的第一个案例：他把名画《打牌的狗》（Dogs Playing Poker）做成 360 度视角，自己如同围坐在牌桌旁的一只狗一样环顾四周，沉浸感趣味十足。 伴随着画幅与认知能力的扩展，模型展现出极大的美学适应弹性，无论是面向幼儿的童趣绘本、趣味梗图，还是面向顶级咨询公司的商业 PPT 附图，它都能拿捏得当。然而，在社交网络上引发最剧烈轰动的，却是一种反其道而行之的趋势：网民们纷纷将真人照片或精美图像喂给模型，要求它反向生成粗糙低保真的微软画图板（Microsoft Paint）风、水彩乱涂或蜡笔简笔画风格。 针对这种“大材小用”的滑稽热潮，Adele 点出了背后深刻的认知逻辑：“要创造出真正‘不完美’的东西，实际上需要非常高的智能。”（"It takes a lot of intelligence to actually create something that is imperfect."）人类用户在面对强大而精密的 AI 时，内心真正渴望的往往是真实感（Authenticity）、怀旧情绪（Nostalgia）以及对不完美的接纳。他们借助最先进的 AI 模拟粗糙、展现瑕疵，从而以一种极具亲和力且带有自嘲幽默的方式与智能系统建立情感链接。 ### [10:30](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=630s) AI 赋予的自我表达与突破时刻 在谈及图像生成的乐趣时，Adele 指出，用户不仅希望通过生成的画面展现美好精致的形象，更渴望展现自己幽默搞怪、不落窠臼的一面。通过 AI 实现自我表达是团队极为看重的方向，这也是公司的核心使命之一：既让大众更轻松地获取并分发前沿智能，又赋予每个人前所未有的个性表达通道。 当主持人 Andrew Mayne 询问模型研发过程中是否存在某个让大家确信“大功告成，可以发布”的质变瞬间时，Kenji 讲述了研发团队的日常工作习惯。在模型训练过程中，研究人员会不时保存 Checkpoint（检查点）并从中抽样生成图像来评估模型表现。就在他们从新模型的一个早期检查点抽样出一张画面——似乎是一位女性眺望海边的场景——并将其与 ImageGen 1 生成的同类图像放在一起对比时，所有人瞬间达成共识：质的飞跃已经发生，毫无悬念地超越了前代。 Andrew 由此回忆起早期 DALL-E 模型的迭代过程：最开始的生成结果常常带有飘忽的卷须状纹理（wispy tendril），当时的研究员曾笃定地预测“大概再跑两次训练运行就能彻底消除”，随后模型果然奇迹般地变得清晰锐利。对比更早年间玩 GAN（生成对抗网络）时甚至需要眯起眼睛猜测“这到底是不是一辆皮卡”，如今图像的进化速度令人惊叹。这一代模型最惊艳的飞跃在于真正的真实感（photorealism）——彻底摆脱了过去那种过于油润、宛如精修杂志封面的塑料感，呈现出宛如出自专业摄影师之手的真实质感。 - **从 GAN 到扩散与前沿模型**：从早年模棱两可、依稀可辨轮廓的像素块，到如今纤毫毕现的真实照片级质感，生成图像跨越了“假想与真实”的鸿沟。 - **标志性的抽样对比**：研究团队通过一张海边女性的抽样图，与前代图像并排比对，凭借肉眼可见的细节与真实度差距，确立了新架构的成功。 ### [12:22](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=742s) 效率与审美品味：后训练中的艺术平衡 Andrew 提出了许多用户的困惑：在没有无限追加算力的情况下，团队是如何做到让图像质量大幅提升，同时又无需让用户干等一个小时的？在早期 DALL-E 时代，用户提交需求后，往往要等待一个小时才能在社交平台上看到成果，而如今 ChatGPT 可以在几秒钟内快速出图。 Kenji 解释称，团队从版本 1、1.5 到 2 的每一次发布中都吸取了大量工程经验。针对推理速度与资源开销，团队投入了巨大精力来优化模型的 Token 效率（Token Efficiency），使其能够用更少的 Token 生成细节更丰富、质量更高的画面。 ```mermaid flowchart TD accTitle: Images 2.0 后训练能力架构 accDescr: 展示模型在后训练阶段如何融合科学逻辑能力与主观艺术审美品味 subgraph Cognition ["认知与世界知识"] W1["科学与数学概念理解"] W2["空间几何与物理常识"] W3["上下文语境与文本记忆"] end subgraph Aesthetics ["审美与风格对齐"] A1["自然真实感 (Photorealism)"] A2["多元视觉风格适配"] A3["专业艺术与设计品味"] end Cognition --> O["Images 2.0 综合表现"] Aesthetics --> O ``` 除了速度优化，Adele 强调了 Post-training（后训练）阶段所面临的独特挑战。训练团队不仅要教导模型理解客观世界知识——例如各种科学概念、数学逻辑以及物体在三维空间中的真实形态，还必须深入探索人类主观的“审美品味”（Taste）。团队必须解决一系列深层哲学问题：什么样的构图能与用户产生共鸣？什么才算真正的美丽？如何让画面兼具质感与真实感？研发目标是让该模型成为当前艺术审美能力最强、创作跨度最广的模型，无论用户需要的是严谨严肃的商业专业输出，还是日常随性的个人创意，模型都能游刃有余地给出最佳视觉方案。 ### [14:06](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=846s) 独特评测法：“我我我测试”与拿橙汁的女人 在讨论如何评测模型水平时，Adele 揭秘了自己独创的私人 Benchmark 基准——“我我我测试”（"me, me, me eval"）。这套评估集包含了她本人、家人与好友的大约 100 张真实照片。她会尝试用模型把身边的亲友置于各种滑稽有趣的场景之中，或是为每个人生成定制的生日贺卡。 Adele 认为这项测试绝不仅是自娱自乐，而是极具实用价值的真实检验： 1. **面部熟悉度与失真识别**：人对自己和最亲近之人的面部特征最敏感，极微小的面部结构瑕疵都能被一眼看穿； 2. **ChatGPT 个性化记忆联想**：评测不仅考验模型的生图底模能力，更考验 ChatGPT 对上下文的理解能力——ChatGPT 是否还记得用户有一个弟弟、父母各自的兴趣爱好是什么，以及能否在画面生成中自然融入这些具有情感温度的个性化细节。 Kenji 和 Andrew 也分享了研发团队中口口相传的经典评估题目。例如研究员 Divya 最钟爱的基准提示词是“一个手拿一壶橙汁的女人”（a woman holding a jug of orange juice），整个团队曾生成过海量拿橙汁女性的测试图；此外还有测试空间与左右关系的“左手写字、右手戴手表、背景时钟显示特定时间”。而在图像模型 1.0 或 1.5 时代，一道著名的难题是画出“倒满到杯沿的红葡萄酒”（folded to the rim），在旧模型中往往需要极为冗长严密的描述才能勉强实现。 Andrew 还分享了一次充满极客趣味的对抗测试：曾有观点认为此类扩散模型绝对不可能生成像素精密的像素画（Pixel Art）。听到这个断言后，Andrew 偏要尝试，直接给模型输入了一个 64×64 的网格指令，命令它严格在网格范围内绘制像素艺术。结果模型不可思议地精准填格，画出了规范的像素画，展现出惊人的指令依从性（Promptability）。 ### [16:32](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=992s) 从模糊 Prompt 到艺术家级专业控制 针对模型的提示词响应，Kenji 提到许多日常用户输入的 Prompt 其实非常模糊，往往只有一两句“让它更好看一点”、“把我画得可爱一点”。模型的系统引导机制（Harness）与模型本身的性格，必须具备从模糊描述中准确提炼用户潜在真实意图的能力。这种能够持续给用户带来“超出预期之惊喜”的特质，正是该模型最吸引人的魅力所在。 | 用户交互分层 | 提示词输入特征 | 模型呈现效果与价值 | |:--- |:--- |:--- | | **大众小白用户** | 简短、抽象、口语化（如“让我变可爱”、“好看一点”） | 模型自主推断意图并进行审美润色，带来惊喜感 | | **专业艺术设计从业者** | 具象、专业术语（景深、焦段、特定配色与构图） | 精确可控，模型能充当执行力极强的创意放大器 | | **上下文参考输入** | 直接上传风格参考图、草图或情绪板灵感图 | 捕捉视觉灵魂与神韵，无缝迁移并重构新画面 | Andrew 感叹道，早年在 DALL-E 时代大家都自诩为“Prompt 工程师”，自己曾为拼凑词藻生成一只“太空中的浣熊”而洋洋自得；然而当真正具备专业背景的艺术家入场时，他们使用行业本身的语言（如景深、镜头焦距、艺术流派）所创造出的视觉杰作彻底拉开了差距。 Adele 对此深表赞同。她透露团队在开发过程中与大量先锋艺术家、设计师和营销专家保持着极紧密的合作，将各专业领域的最佳实践与思考方式提炼并固化到了模型的交互逻辑中。此外，她推荐了一个极其实用的技巧：直接在提示词中上传灵感参考图或上下文素材。模型展现出非凡的“神会”能力，能够敏锐捕捉上传图像中的精神内核与视觉风格，并将其精准转化到全新的生成结果之中，从而为专业创作者提供了前所未有的掌控度。 ### [18:50](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=1130s) 创意放大器与学术教育图表的质变 在探讨模型的多功能性时，Adele 兴奋地指出，该模型大幅拓宽了大众的创意边界。以往在不同视觉风格间切换往往需要切换不同工作流，而现在的图像生成模型能够在严肃严密的“建筑设计线框图”与充满童趣的“童书绘本插画”这两种截然相反的审美向量之间自如切换。 在科研与教育领域，这种精准的图表化与视觉表达能力展现出惊人的生产力潜力。Adele 透露，OpenAI 内部设有一个专供模型试用的 Internal Alpha（内部内测）测试频道，其中包含一个专门面向教育工作者的子频道，成员涵盖小学教师到高校教授。 在该测试群组中，一位大学生物学教授曾利用该模型生成极为晦涩复杂的研究生水平生物学教科书配图插页，其精细程度与科学准确性让教授本人惊叹“无可挑剔”。从抽象复杂的分子生物学机制，到中小学课堂上直观生动的概念图解，模型在将庞杂知识浓缩为一目了然的可视化图像方面展现出顶级实力。教师和学生们正广泛运用这一工具制作个性化教案、备考学习指南，彻底颠覆了传统教学材料的制作流程。 ### [20:37](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=1237s) 个性化学习与信息图表：从课堂海报到内部演示 谈到图像生成的未来应用场景，个性化学习被视作一个极具潜力的发展方向。教育工作者可以借助图像模型，根据每个孩子的语言背景和理解偏好，量身定制他们能够看懂的教学素材。团队正积极探索如何将图像生成深度融入 ChatGPT 的整体学习体验中，让抽象概念的传授自然而然地伴随着直观的视觉辅助。这让人回想起多媒体尚未普及的学生时代——教室里那些用以解构复杂知识的教学挂图与信息图表（Infographics）往往拥有极强的吸引力，学生可以驻足细看其中的每一个微小细节。图像模型不仅擅长构思内容，更在页面排版和文字渲染（Text Rendering）上展现出惊人的成熟度。 这种视觉沟通方式的变革在团队内部早已悄然发生。如今在 OpenAI 的内部汇报演讲中，超过 50% 的演示幻灯片都是直接借助图像生成模型完成的。模型不仅理解“应该表达什么”，还精通“如何以最佳视觉形式呈现”，这被团队形容为一种“超能力（superpower）”。未来的探索重心在于进一步优化排版构图的多样性，并探索如何在产品端直接为用户提供可编辑、可调整的交互能力。 - **教学定制化**：利用图像生成将枯燥的抽象知识转化为兼顾学生语言与偏好的图解素材。 - **内部渗透率**：OpenAI 内部演讲与汇报中，已有超过半数的幻灯片视觉内容由图像生成工具直接辅助完成。 - **产品演进方向**：重点突破画面的排版构图与文本融合，并规划在原生产品中加入更细致的局部编辑功能。 --- ### [22:16](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=1336s) 从工具到创意智能体：探索下一代设计生态 当被问及图像生成的下一阶段演进时，团队明确指出了“创意智能体（Creative Agent）”的愿景。未来的图像模型不再仅仅是一个被动接收提示词的单点渲染工具，而是能够常驻在用户身边的创意助理——它理解用户的工作流、审美风格与最终目标，化身为每位普通人的私人室内设计师、专属建筑师乃至婚礼策划师。 这种对用户意图的深度解析在日常场景中已经初露锋芒。主持人 Andrew Mayne 分享了自己的一次真实经历：每当出版新书时，他都需要为 X（前 Twitter）或 Facebook 等社交平台更新横幅头图。他只是给模型发去指令，让它“找到我的新书封面，并生成一张合适尺寸的社交媒体头图”。模型直接在“首发命中（First shot）”的情况下做到了完美匹配，无需反复调试就自动完成了背景检索、风格延伸并适配了恰当的画面宽高比（Aspect Ratio）。研究团队透露，任意比例渲染在研发早期就经过了充分的技术风险规避训练，模型如今完全具备结合上下文自主调研并推断用户需求的能力。 从写书的作者到借助虚拟置景（Virtual Staging）与户型渲染制作房源列表的房地产经纪人，再到设计视频缩略图的 YouTube 创作者以及希望与粉丝建立更深视觉联结的顶级艺术家，图像生成正迅速成为跨行业工作流中的必备利器。在团队看来，这标志着图像生成领域彻底迈入了全新的范式：“如果说 DALL-E 代表着石器时代，那么 Images 2.0 就标志着文艺复兴的到来。”它不仅承载着纯粹的美学艺术，更将科学、构图和现实世界的物理架构知识熔铸为一体，产出高度可信的专业级视觉成果。 ```mermaid flowchart TD accTitle: 创意智能体多领域落地协作流程 accDescr: 展示创意智能体如何理解用户模糊意图并赋能不同专业场景的工作流 subgraph Input["用户意图输入"] A["提出模糊创意需求<br/>(例如：制作新书宣传横幅)"] end subgraph Agent["创意智能体 (Creative Agent)"] B["信息检索与上下文理解<br/>(搜索书籍封面/房源特征)"] C["美学排版与规格自适应<br/>(推断画幅比例/构图/风格)"] B --> C end subgraph Output["跨行业专业落地"] D["出版营销<br/>(社交媒体专属横幅)"] E["房地产行业<br/>(虚拟置景与房源图)"] F["内容创作<br/>(视频封面与社群视觉)"] end A --> B C --> D C --> E C --> F ``` --- ### [25:18](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=1518s) Images 与 Codex 的交汇：从概念草图到零样本应用构建 图像生成与编程智能体（如 Codex）的交叉融合催生了前所未有的生产力跃迁。当下大量开发者与创作者在启动新网站或独立应用时，第一步往往是让图像模型构思视觉设计。当高水平的视觉审美模型与强大的代码生成模型并轨运行时，用户便能以“零样本（Zero-shot）”的方式从零打造出令人惊艳的完整软件项目。 Andrew 分享了自己使用 Codex 搭建网站的奇妙体验：他先让图像生成工具为网站构思不同的视觉概念，并要求以“联络印相样片（Contact sheets）”的形式输出四套并列方案；随后他选中右上角的方案并指令 Codex 进行实现，亲眼见证代码智能体直接依据视觉设计将网页搭建成型，整个过程“宛如魔法”。在游戏与趣味应用开发中，这种协同同样展现出惊人的创造力——例如在开发虚拟宠物时，只需提出想要一只渡鸦（Raven），Codex 就会自动调用图像生成工具，通过多轮迭代生成一整套精准的精灵图（Sprite sheets）。 在游戏资产设计方面，保持跨图像的视觉连贯性（Consistency）曾是困扰行业已久的痛点，用户以往不得不依赖复杂而脆弱的临时工作流来拼凑效果。而在当前架构下，模型表现出了卓越的上下文参照能力，甚至支持用户构思包含一致角色与情节走向的 10 页连环漫画。创作者可以先生成包含多角度动作姿态的角色设定图（Character sheet），随后将其作为图像上下文喂回模型，持续生成姿态各异却角色一致的后续画面。这种参照机制就像大语言模型通过扩展上下文窗口（Context Length）赋能长文本推理一样，彻底拓宽了视觉创作的自由度。 - **开发范式转变**：图像模型负责界面 UI 概念原型设计，编程智能体（Codex）负责将视觉稿直接转化为可运行的前端代码。 - **精灵图技巧**：借助思考模式（Thinking mode）或 Codex，先精准生成第一个独立的基础精灵图（Sprite），随后以此为基准延伸生成成套的连贯动图或 Sprite sheet。 - **角色一致性**：模型能够像 LLM 理解上下文一样稳定参考前期图像，支持长篇连环画创作、分镜绘制及角色多姿态连续输出。 --- ### [28:08](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=1688s) 提示词技巧与思考模式：释放模型的推理与审美潜力 在对话的尾声，团队分享了能够立竿见影提升图像生成质量的提示词与使用技巧。Adele 强烈建议用户体验结合了思考模式（Thinking mode）的图像生成版本。在该模式下，模型不仅具备更强大的底层画质与构图表现力，还拥有自主调用搜索、分析文件及使用内置工具的 Agent 能力。 对于思考模式下的提示词策略，Adele 的核心建议是“保持开放性（Be open-ended）”。用户无需在一开始就将细节定得过死，模型自身会在生成前进行推理与信息探索，挖掘出真正契合场景的核心元素；同时，如果能在提示中赋予其明确的审美取向或锚定某种具体的视觉风格，往往能收获远超预期的精致效果。 Kenji 则从个人审美出发补充了另一项关键偏好：清晰明确地声明你所青睐的设计风格。由于当前的图像模型在知识与细节表达上极其充沛，有时画面的信息密度容易过载。作为一个偏好极简主义风格（Minimalist infographics）的人，他通常会在提示词中明确强调干净整洁的视觉质感（Clean look），从而过滤掉过于杂乱的修饰，让最终呈现的图表更加一目了然。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=0s) 如果说 DALL-E 代表着图像生成的石器时代，那么 Images 2.0 就是它的文艺复兴。 — Andrew Mayne > [00:00](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=0s) 哪里需要解决问题就顶上去 — Adele Li > [07:35](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=455s) 要创造出真正‘不完美’的东西，实际上需要非常高的智能。 — Adele Li > [14:06](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=846s) 我我我测试 — Adele Li > [14:06](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=846s) 一个手拿一壶橙汁的女人 — Divya > [14:06](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=846s) 倒满到杯沿的红葡萄酒 — Kenji Hata > [16:32](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=992s) 让它更好看一点 > [20:37](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=1237s) 超能力 > [22:16](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=1336s) 找到我的新书封面，并生成一张合适尺寸的社交媒体头图 — Andrew Mayne > [25:18](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=1518s) 宛如魔法 — Andrew Mayne > [28:08](https://www.youtube.com/watch?v=bH2nP-aCFjk&t=1688s) 保持开放性 — Adele Li ## Mentioned OpenAI, ChatGPT, Andrew Mayne, Kenji Hata, Adele Li, Images 2.0, Images 1.0, Images 1.5, DALL-E, DALL-E 3, Codex, 红点创投 (Redpoint Ventures), Ada, Babbage, Curie, 《打牌的狗》(Dogs Playing Poker), 微软画图板 (Microsoft Paint), Divya, X (前 Twitter), Facebook, YouTube ## Video description People are generating over 1.5 billion images a week in ChatGPT. In this episode, Product lead Adele Li and researcher Kenji Hata share some of the new use cases and trends since the launch of Images 2.0. Together with host Andrew Mayne, they trace the progress from the early DALL-E days and dive into the latest capabilities, including better text rendering, photorealism, multilingual support, world knowledge, aspect ratios, and character consistency. They also explore what comes next as image generation models evolve into more capable creative assistants. Chapters 00:36 How Adele and Kenji came to work on Images 02:27 Images 2.0 launch reception 05:25 Productivity use cases and and 360 images 09:34: Viral trends, authenticity, and imperfection 10:51 Training breakthroughs and photorealism 14:06 Evals, prompting, and creative control 22:16 Creative agents and what comes next 22:27 Images + Codex 28:08 Prompt tips [Source](https://www.youtube.com/watch?v=bH2nP-aCFjk)

[0:00] Andrew Mayne: 如果说 DALL-E 代表着图像生成的石器时代，那么 Images 2.0 就是它的文艺复兴。 [0:00] Adele Li: 哪里需要解决问题就顶上去 [7:35] Adele Li: 要创造出真正‘不完美’的东西，实际上需要非常高的智能。 [14:06] Adele Li: 我我我测试 [14:06] Divya: 一个手拿一壶橙汁的女人 [14:06] Kenji Hata: 倒满到杯沿的红葡萄酒 [16:32] 让它更好看一点 [20:37] 超能力 [22:16] Andrew Mayne: 找到我的新书封面，并生成一张合适尺寸的社交媒体头图 [25:18] Andrew Mayne: 宛如魔法 [28:08] Adele Li: 保持开放性

[7:31] Adele Li: it takes a lot of intelligence to actually create something that is imperfect. [14:05] Adele Li: me, me, me eval [14:05] Adele Li: make me look better [14:05] Adele Li: make me cuter [22:15] Adele Li: personal interior designer, personal architect, personal wedding planner, et cetera, all in one. [23:55] Kenji Hata: if DALL-E was the stone ages, ImageGen 2.0 is the Renaissance. [25:19] Andrew Mayne: zero shot [28:11] Kenji Hata: ImageGen thinking [28:11] Andrew Mayne: minimalist infographics

[0:00] Hello, I'm Andrew Mayne, and this is the OpenAI Podcast. [0:03] On today's episode, we're talking about Images 2.0 with researcher Kenji [0:07] Hata and product lead Adele Li. [0:10] They'll discuss why the new model represents such a major leap forward, the [0:14] evaluations that mattered most during development, and what people are [0:16] creating with it now that it's widely available. [0:17] If DALL-E was the Stone Ages, ImageGen 2.0 is the Renaissance. [0:23] It's not only great artistically and aesthetically, [0:26] but it also incorporates science, art, architecture, all in one image. [0:30] We looked at it and we're like, all right, this is better than ImageGen 1. [0:36] Adele, tell me a little bit about how you became a product manager here. [0:39] So I joined OpenAI a little over two years ago. [0:42] And before OpenAI, I was an investor my entire career. [0:45] Oh, wow. [0:46] So I was in private equity and spent three years at Redpoint [0:48] Ventures investing in AI and software companies. [0:52] And when I first joined OpenAI, it was for a completely different role. [0:55] I was thinking about how do we build out our data and compute [0:57] infrastructure. And over time, made my way over to the product side. [1:01] And for the last six months, have been working on ImageGen. [1:04] It's interesting how you style yourself going from one role, then finding [1:08] yourself into this space here, which is kind of cool to think about the idea [1:11] that you have this sort of ability to be useful in different ways. [1:14] Absolutely. And I think the role of a product manager is just to do the job [1:18] that needs to be done, no matter what it is. [1:22] And for ImageGen in particular, it's been really awesome to flex a lot of [1:25] different muscles when it comes to building products, working with [1:28] researchers like Kenji, but also thinking about what is the gap in the [1:32] market today that we want to fill and what is the opportunity that we want [1:36] to grasp here. It's not the same market that it was a year ago when we [1:40] first released ImageGen 1.0. [1:42] Now it's a very different landscape. [1:44] There are multiple image generation makers out there. [1:47] And ChatGPT is a very different company and product itself too. [1:51] And so really thinking about the evolution of ImageGen and its role within [1:55] ChatGPT has been really, really exciting to me. [1:58] Kenji, how did you end up working on images? [2:01] Actually, like when I first started at OpenAI, I also started about two [2:05] years ago. I was working on like some random audio project initially. [2:11] It was my first project. And then at the time, I just found my way just [2:16] working on helping them work on ImageGen 1.0 prior to the launch. [2:22] And so gradually I moved more and more onto the project and then I became [2:25] full time on it, basically. [2:26] What has the reception been like right now for [2:29] the model? [2:30] In the last two weeks since we launched the model, usage is up [2:33] more than 50%. More than 1.5 billion images are generated every week on [2:38] ChatGPT. And we've seen viral trends emerge across the world. [2:42] All the way from trends in Asia for color analysis and stickers to US where [2:48] crayon and scribble are going viral. [2:51] but also a lot of people exploring emergent use cases. [2:54] And I think it shows the dynamic range of the model, but also how people are [2:59] able to visually grasp the advancement of the model almost immediately. [3:03] I think the visual communication reaction that we've seen from our users for [3:08] them to say, hey, this is the best, highest fidelity, highest quality in a [3:12] static model that we've seen has been really awesome. [3:16] This felt like a really big shift, almost worthy of me not even being images [3:20] too, but almost like just a new paradigm because just the capabilities are [3:24] through the roof. What made that possible? [3:28] When we started working on this project, I think we sat down and we [3:31] discussed what is the step change of capability and use cases that we wanted [3:36] to build towards. [3:38] And we believe that image generation has the ability to do so much more than [3:43] what it does today. you could distill every single output or visual content [3:48] that you see today into an image. [3:51] And so that was the mandate that we sought out to improve. [3:54] And with this 2.0 model, we've improved on various different dimensions. [4:00] One is text rendering. The ability for text on a page is so much better [4:04] fidelity. The language and words actually make sense in their actual words. [4:10] The second of all is multilingual. [4:12] So we've really focused on making this model work in various different [4:15] languages. And we're already seeing that people across the world in Asia and [4:20] Europe are really resonating with these advancements. [4:24] The third is photorealism. I think we really saw a lot of feedback from our [4:28] previous models that the output wasn't very realistic or altered their face [4:32] or their bodies. And so one of our mandates was how do we actually make the [4:37] image feel like more like yourself? [4:39] And so all the things that you think that the model knows, it does because [4:44] it has imbued the knowledge of the world into its conscience and is able to [4:48] visually communicate that back to you as a user. [4:52] And so putting that all together, I think we really get a state-of-the-art [4:56] image generation model that is the best aesthetic model out there on the [5:00] market right now. That really represents a new paradigm for image [5:05] generation, which is a huge part of, I think, AI progress at large. [5:09] that we have an opportunity to work on here. [5:11] We often listen back, listen to feedback on social media too. [5:15] So we kind of just take all these things and basically are just aware of it [5:19] and try to make sure that they're mitigated or completely fixed in some [5:23] cases in the next iteration. [5:25] What kind of use cases are you seeing? [5:27] What are you seeing people do with this now? [5:29] I think one that's particularly [5:30] close to like the research team as a general is like infographics, text. [5:35] I think text in images is like so much better nowadays. [5:38] So I think it just opens up a lot more productive use cases. [5:43] And from the research side, we kind of think image generation used to [5:48] always be about fun and maybe unproductive things. [5:52] But now we're really seeing steps forward into productivity and image [5:56] generation for any type of use case that you can imagine it for. [6:00] So you mentioned text. I remember the early models, no disrespect to [6:04] chimpanzees, but getting to dispel like OpenAI even looked like a chimp did [6:08] it. And then now I'm looking at pages of text and finely detailed stuff. [6:12] And I know that as models get smarter, variable binding, the ability to put [6:16] things next to each other improves. [6:18] But this was just a big improvement. [6:20] Yeah. But I don't think it's completely unexpected. [6:25] I think you see a lot of growth in between. [6:29] Well, first you see between DALL-E 3 and GPT Image 1. [6:34] If you ask for a grid of random objects, you go from maybe [6:40] like 5 to 8 in DALL-E 3 to maybe around 16 in Images 1. [6:45] And then with 1.5, we went to about 25 to 36 consistently. [6:51] And I think now we could probably do over 100. [6:54] I think this is like a test that we might do internally. [6:57] It's just we just ask GPT, give me a list of 100 random objects. [7:02] And then we just send that to our image generator and see how many are [7:07] correct. And usually, you know, it'll get almost all 100 correct. [7:12] And that's, but you see the constant growth over time. [7:17] So I don't think it's like completely unexpected. It's just a steady pace. [7:20] That was a test I used to use for like the really old models back with [7:23] like Ada, Babbage, and Curie, like list 100 science fiction books. [7:26] And then some of them would get, by the time it got to like 22, we'd [7:28] just start repeating stuff. So we realize the model reached the end of it. [7:31] So we've seen stuff too, like 360, 360 degree panoramas. [7:35] How did that happen? [7:36] Yeah, that really came from the emerging capability of [7:38] the model, which is the ability to render images in any aspect ratio. [7:42] We discovered that people were generating really long, amazing panoramics, [7:47] skinny bookmarks as well. And one of the cool capabilities of the model is [7:52] that not only were you able to generate images in this panoramic aspect [7:56] ratio, but you could also render images in the style of 360. [8:00] And we saw that it was really fun to actually view these images in a 360 [8:05] world itself. And so that was a really fun feature that we ended up adding [8:09] into the product and it's available on ChatGPT on web and mobile right now. [8:13] First thing I did was I made a version of dogs playing poker. [8:15] Put that in there so you could sit there like you're one of the dogs looking [8:18] around in there, which was not something I expected, but it's fun. [8:21] Yeah. I mean, it's really awesome to see how people are exploring new use [8:26] cases and fun things that they're creating with the model, even far beyond [8:30] what we expected users to be using it for. [8:32] I think when we were designing the model, we were really deliberate in [8:36] understanding what people really wanted to see from image generation. [8:40] There was a lot of latent demand in image generation. [8:43] People were mostly using it for personal use cases, but we definitely saw a [8:47] lot of inklings of people wanting to push the model in certain directions [8:51] that the model wasn't good at. So text rendering was definitely one of those [8:55] dimensions that we really wanted to improve on. [8:58] Multilingual is another. And I think world understanding generally is so [9:01] much better in this model. And that typically means that now people online [9:06] are sharing a bunch of examples of them creating image done for all [9:09] different kinds of use cases that we didn't even think existed out there. [9:14] So I think the model's understanding of aesthetic beauty across [9:19] multiple different outputs, whether that it's like a fun meme, an image for [9:23] a five-year-old versus a professional consulting deck. [9:28] The expansion of opportunity and outputs has been amazing to see in this [9:32] latest model. It's funny too how one of the things that was trending [9:36] was taking popular images or photos of people and then having the model make [9:42] like kind of janky looking Microsoft Paint versions of that. [9:45] Yes. [9:46] And did you think that was something you would see was that people are [9:48] going to use this incredibly capable tool to then go make, you know, these [9:51] silly looking things? [9:52] Yeah, it's funny because it takes a lot of [9:55] intelligence to actually create something that is imperfect. [9:58] That's what I tell people all the time. Yeah. [10:00] And it's definitely very interesting in the viral trends that we're seeing [10:03] online right now. One thing that I think people are really striving for is [10:09] authenticity, imperfection, nostalgia. [10:12] We're seeing that in the Microsoft Paint prompt, crayons, all different [10:16] kinds of generations that people are creating. [10:19] And that really feels like the theme of consumers is they want to interact [10:23] with AI in a very authentic, imperfect way. [10:26] They want to show their imperfections and use AI to help make them look [10:30] good, but also show a more fun and goofy side of themselves. [10:34] And I think that's self-expression via AI is something that we're really [10:37] excited about. And, you know, I think it's really part of our mission as a [10:41] company to make it easier for people to learn more and distribute that [10:44] intelligence, but also letting them express a version of themselves that [10:48] maybe wasn't possible before. [10:51] Kenji, was there a moment with this model where you're saying to yourself, [10:55] wow, I think this is ready to go? [10:57] You know, as it's training, we take a checkpoint and then we just sample [11:00] from it, right? And just see, okay, how good is this thing? [11:03] And I think we just sampled a model, an image, and we looked at it and [11:07] we're like, all right, this is better than ImageGen 1. [11:11] We were just like, okay. I remember watching the iteration of one of the [11:15] early versions of DALL-E and how at first it was sort of the wispy sort of [11:19] weird sort of the tendril sort of thing and talking to one of the [11:23] researchers like, is that going to go away? [11:26] It's like, I think two, probably two runs away from that. [11:28] And then just like that, the ability to predict that was amazing to me. [11:31] And all of a sudden everything got crisp and clear. Yeah. [11:33] And then also like looking at, you know, years ago I'd played with like, you [11:36] know, GANS and like doing those things. [11:39] You have to squint and say, I think it's a pickup truck or something like [11:41] that. Yeah. So it's interesting what you see is you say, okay, this just all [11:44] of a sudden got much better. Yeah. [11:46] I mean, it was just very obvious. You just take the early checkpoint. [11:50] You just sample an image from it. [11:51] And then you just sample an image from, you know, ImageGen 1. [11:55] And you just look at the two and you're just, there's just. [11:59] Why do I like this garbage? I forgot what the image was. [12:02] It might have just been like a picture of like a woman on the seaside. [12:06] Like, you know, overlooking a seaside. [12:08] We just looked at it and we're like, all right. [12:09] It was like no question. Yeah. That was the big jump was the photorealism of [12:13] going from something that looked that was more of a glossy, idealized [12:17] magazine cover to something that looked like a really good photograph. [12:22] So help me understand, like besides just more compute, how did this happen? [12:26] How did you get a model that's much better and also that doesn't take an [12:29] hour to generate an image? [12:30] The times are still, I remember in the DALL-E [12:33] days, like we would literally have to, you know, tell us what you want. [12:37] And then an hour later, it'd be on Instagram to now these things are in [12:41] ChatGPT and it's faster. How is it getting both more intelligent and you're [12:45] maintaining the same speeds? [12:46] I think we learned a lot in each release, like between 1 and 1.5, now 2. [12:53] And so we take each of the learnings that we've made and we've, you know, [12:57] like, for example, speed, right? [12:59] You know, one of the things is like, oh, can we make the model more token [13:02] efficient or something like that? [13:04] And we did a lot of work to make it produce very good images with less [13:10] tokens. I think the post-training for this model was very interesting in the [13:16] sense that we really had to think about not only does the model understand [13:20] world knowledge and how things look in science, concepts, math, etc. [13:25] in an image, but also what is the taste that will resonate with users? [13:30] What makes the model or output beautiful? [13:33] How do you make it look realistic? [13:35] These are all questions that we had to grapple with when we were [13:39] post-training this model. Because I think that one of the things that was [13:43] really important for us was that this model was the strongest aesthetic [13:48] model out there right now, which means that it has more creativity in [13:52] various different outputs, no matter what that output is, if it's a [13:55] professional output or a personal output. [13:59] And so that range of training and the range of use case, I think, made [14:02] training this model a very interesting problem. [14:05] Do you have any personal favorite benchmark tests you like to do, things you [14:09] say, I want to see it make an image of this? [14:11] I have an eval that I call the me, me, me eval. [14:14] Okay. It's essentially 100 photos of myself and my friends and my family. [14:21] And I put everyone in goofy positions. [14:24] I have about a card or a birthday for every single person. [14:29] And I think it's a really great eval in the sense that you only know the [14:33] people around your faces the best. [14:37] You also want to create funny things with the model and do things that are [14:40] relevant. And so one thing for me as a product manager that I'm testing is [14:45] not only is the raw capability of the model really great, but also does [14:49] ChatGPT understand what I want in that context? [14:53] You know, ChatGPT remembers, you know, that I have a brother, that I have a [14:58] mom and dad and what they like to do. [15:00] And so does the model accurately know how to insert pieces of [15:04] personalization in the moments that matter in the images? [15:08] These are things that I'm testing for. How about you? [15:10] Besides the grid one I mentioned earlier, that's probably the one I've used [15:14] the most. For a while, I think Divya and I were doing a lot about [15:18] photorealism. We're trying real hard to push on that. [15:22] Just basically, I know Divya's favorite one was like a woman holding a jug [15:28] of orange juice. I don't know if you've seen this. [15:31] There's like so many images of a woman holding a jug of orange juice. [15:35] Well, I actually feel like the researchers had a more standard set of images [15:39] than they like to lead on. Yeah, and you get like the standard. [15:41] Can it do somebody writing with their left hand or watching on their right [15:44] hand and a clock showing this? [15:46] I think the big leap of the images, like [15:48] probably 1 or 1.5, was like a half full glass of wine. [15:52] The wine glass folded the rim. Yeah, folded the rim. [15:54] Yeah, exactly. And there were ways I was able to prompt it to do it, but [15:58] it was really – had to get a really descriptive, like, you know, red liquid [16:02] inside this. This one is so fun to prompt. [16:04] There was a thing people said, oh, can it do, like, you know, can it do, [16:07] like, pixel accurate pixel image style art? [16:09] And somebody was like, no, it can't. [16:11] And when I hear that, I'm like, okay, let's try. [16:13] And I found out if I gave it, like, a 64 by 64 grid and I [16:17] said, go draw the art in there. [16:19] It did. It just was able to put art into there. [16:21] And that was amazing to see those kinds of results. [16:24] And that's the promptability of this is insane. [16:27] How do you plan for that? Does it just happen? [16:29] You're like, oh, wow, this is better understanding this? [16:32] People come to ImageGen with very vague prompts. [16:35] Yeah. Make it better. Make me look better. [16:40] You know, make me cuter. All these things are really vague. [16:43] And I think it's really the job of the model and the harness to distill that [16:48] into actually what users want. [16:50] And I think that's a personality of the model that we've trained over time [16:54] that we've really harnessed the power for. [16:56] And honestly, I think it also yields a lot of really surprising results that [17:00] people may not expect. And that surprise is just part of the fun of using [17:04] ImageGen. I've seen like two kinds of prompting sort of emerge. [17:08] And I remember back with DALL-E, I thought like, oh, I'm a prompt engineer. [17:12] I'll be great at this. Like, I'll be really good at this. [17:14] And I'd make a raccoon in space and be like, feel proud. [17:17] And I'd see an artist, somebody who wasn't a prompt engineer, somebody who [17:20] actually came from that world. [17:21] And I'd watch them use their language. [17:24] And they were doing amazing things. Yeah. [17:25] And that seems like that's still holding true. [17:28] Definitely. I mean, we work with a group of artists very closely when we [17:32] develop this model. And we're very inspired by artists, designers, [17:36] marketers, all these different professions that I think have a different way [17:40] of approaching their profession. [17:42] And one of the things that was very important for us is we wanted to take [17:46] the inspiration as well as the best practices for those professions and [17:50] distill that into the way that people interact with the model. [17:54] And so that's something we've deliberately tried to focus on. [17:57] One hack that I've seen work really well is the ability to upload [18:01] inspiration or context into the model. [18:05] And the model has an incredible ability to take the spirit of that context [18:08] and translate it into the output. [18:11] But it's interesting because I think that a lot of people worry that, oh, I [18:14] just push in a button, I get something beautiful. [18:16] And each model, that gets better. [18:19] It's easier, as you said, to not have to put a lot of effort into it. [18:22] But when people do put effort into it, they are getting even more amazing [18:25] results. And it seems like actually that if you're artistically inclined, [18:29] you're getting even greater control because now, like you said, it [18:32] understands more about what you're talking about when you talk about depth [18:35] of field and these other things or whatever you're trying to do. [18:38] And as you mentioned, it was exciting to see with earlier models, artists [18:42] who said, oh, I gave it my originals and it gave me these variations and I [18:45] know which one works. [18:47] And just seeing that as this real creative amplifier. [18:50] Yeah, definitely. I think having creative direction or taste or judgment and [18:54] bring that to the model is the best way to push it further. [18:59] I think one thing about this model that I'm really excited about is how it [19:03] expands the creative outlet for people. [19:06] I think the ability to create multiple different styles or types or [19:11] variations has never been easier than with this ImageGen model. [19:15] And I think it's also understanding of different contexts, like the way that [19:21] it's able to shift what it's like to be generating an architectural diagram [19:26] all the way to the aesthetics of a children's book. [19:30] The ability for it to move so seamlessly across these vectors has been [19:34] really awesome. The ability to do great infographics and diagrams is very [19:38] powerful. What kind of feedback have you been getting from people in [19:41] research and education? [19:43] We actually have an internal alpha channel where we [19:46] test our models. [19:49] And in that, there's like a sub-channel dedicated specifically towards [19:53] educators of any level, like elementary school students all the way up to [19:57] graduate level. One of the coolest things I saw was there was a biology [20:03] professor and he put like these graduate level textbook rendering pages of [20:08] things I had no clue about. And he said it was perfectly accurate. [20:12] I think the ability for this model to distill very complex topics into [20:17] something that is really easy to understand within an image is one of its [20:22] strongest capabilities. [20:25] And we've seen this with students, with teachers who are using ImageGen to [20:30] learn different concepts, to also help them create study guides, to help [20:35] also create personalized content. [20:37] I think personalized learning is a huge trend that we're very passionate [20:40] about. And I think the ImageGen model helps you as a teacher create [20:44] something that every kid can understand in their own language and their own [20:49] preference. And that is something that we're really excited about. [20:53] We're thinking about this in the context of also how do we bring more of the [20:57] elements of ImageGen into ChatGPT at large so that when people are trying to [21:01] learn concepts, we're teaching them with ImageGen. [21:04] I remember when I was in school and kind of prior to a lot of kind [21:07] of multimedia blowing up, posters were a big thing, classroom posters [21:10] explaining stuff. This really reminded me of how powerful an infographic can [21:16] be because it allows you to bring as much attention as you want to it. [21:21] And you can spend the time looking at it and seeing it, and you can put [21:24] a lot more detail into it. I think one really awesome visual shift that I've [21:29] seen with ImageGen is that now in internal presentations, over 50% of the [21:34] slides are created with ImageGen. [21:36] Wow. And that permeation of communication via images is so powerful when [21:41] you're trying to explain your concepts or illustrate what you mean. [21:46] And I think infographics and the text rendering capability, as well as the [21:50] composition of the text on the page, is incredibly powerful with this model. [21:54] The model's understanding of not only what to say, but how to present it is [21:59] a superpower. And I'm really excited about future explorations of this, [22:03] where we can think about how do we make this even better? [22:07] How do we improve the composition, the different kinds of outputs, and also [22:11] make it editable in the product? [22:13] These are directions that we're really excited about. [22:15] How do you see the progression of this? [22:17] This is great, but typically anytime I talk to somebody to open an eye about [22:20] what they're working on, they're like, yeah, this is good, but. [22:23] I think we're still super early in exploring all the different use cases [22:27] that people are really trying to push the model with. [22:30] And so one of the things that we're really excited about is what is that [22:35] next stage for ImageGen, which is to create the creative agent. [22:39] Ultimately, the agent that can work alongside you, be your creative [22:44] assistant, and really understand how you work, what your preferences are, [22:49] what is the output that you want to get to, and build the product and model [22:54] ecosystem that helps users kind of have a personal interior designer, [22:59] personal architect, personal wedding planner, et cetera, all in one image. [23:05] I'll tell you another thing that was kind of amazing was like, all right, [23:07] books. And so every now and then I have a book come out, I've got to [23:09] change my social media headers. And I just went and I said, oh, find my book [23:13] cover and create an appropriate size social media header that I can put on X [23:18] or Facebook or whatever. I'm like, well, let's see. [23:21] First shot. First shot. Right aspect ratio. [23:23] Everything. We basically did that from the start or trained the models to be [23:27] good at that from the start. I remember I worked on the initial de-risk of [23:32] basically it could do any aspect ratio that you ask. [23:37] Yeah. Yeah, you can now really just easily specify the outcome that you [23:41] want. Yeah. Like in the case of yourself, you're like, I want promotional [23:45] material. I don't have an idea. I didn't specify exactly what I wanted. [23:49] But the model was able to do the research and then give it to you in [23:52] the style and aspect ratio that was relevant to you. [23:55] And that's super powerful. We're already seeing this. [23:59] you know you're an author I've talked to real estate agents who are using [24:03] ImageGen to help them create listings for their apartments or stage their [24:08] listings YouTube creators have talked to me about using ImageGen for their [24:12] thumbnails and promotional content I've talked to top artists who want to [24:17] use ImageGen to connect with their fans and I think the ability for all [24:20] different kinds of professions to start to use ImageGen to help them with [24:25] visual creation is super powerful especially if you're working in a visual [24:29] and a creative industry, ImageGen is such a hack in your professional [24:32] toolkit. I think it has to be a part of everyone's everyday workflow in the [24:36] future. This does feel like the, I think it feels like the first time where [24:40] anything I can reasonably come up with, it does a pretty good job of it. [24:44] We think it's a new paradigm for image generation altogether. [24:47] Like if, you know, we set this in the launch video, if DALL-E was the stone [24:50] ages, ImageGen 2.0 is the Renaissance. [24:53] Yeah. And I think that is so true because the model, It's not only great [24:58] artistically and aesthetically, but it also incorporates science, art, [25:03] architecture, all in one image together. [25:06] And I think that composition and knowledge that the model has just means [25:10] that the outputs are so much more trustworthy, are more powerful, and enable [25:15] so many more use cases. I think that ImageGen and Codex is also an amazing [25:19] intersection of the capabilities that we're setting out to create with both [25:23] ImageGen as well as coding agents. [25:26] So many people are using ImageGen as a first step to designing a new website [25:31] or creating a new app. And I think that intersection of having a really [25:36] strong aesthetic model, which is image generation, in combination with [25:40] strong coding abilities, means that now you're able to zero shot really [25:45] amazing apps from scratch with both of these tools. [25:48] Yeah, I asked it in Codex. I said, I took my website. [25:51] I said, could you make me, I had the ImageGen. [25:53] Could you create me some different concepts for it? [25:56] And I did these contact sheets. I asked for contact sheets to that. [25:58] Give me like four images there. And I said, oh, the one on the upper right. [26:01] Can you go make that? [26:02] And I watched Codex go make that, which was like, this [26:05] feels like magic. And then they've implemented as part of pets. [26:10] And so like if you're using Codex and you say, hey, I want to have like [26:13] I have like I love Raven. So I have like a Raven. [26:15] I said, can you make a Raven? And then I watched it pull up the ImageGen [26:18] tool and iterate and make the sprites for it. [26:20] Yeah. Sprite sheets are going viral. [26:22] Yeah. Same with game design. [26:24] People are loving. using ImageGen to help them create new worlds. [26:28] Any hints on how to do better sprite sheets? [26:31] I mean, I've tried to make, you know, GIFs internally. [26:34] And I think if I just use, like, the thinking mode or Codex, and you [26:38] basically just ask it to generate one initial sprite, it's really good. [26:43] And then you can just say, can you make the rest? [26:46] The consistency across multi-images has been amazing. [26:51] We've seen a lot of people try creating 10-page comic books with consistent [26:56] storylines, multi-page slides. [26:59] I think that consistency of characters and aesthetics [27:03] is completely unique to this model. [27:05] That was an example too, where there were a lot of workflows out there for [27:09] working with image models that were kind of janky, but you had to figure out [27:14] how to do. And it's great now because I can do stuff where I can create [27:20] characters and say, make a character sheet with the different poses and [27:23] stuff and just go feed it back in and say, okay, now doing this, now doing [27:26] that, now doing that. And that's just such a – often, sometimes what we need [27:31] is obviously a smarter model, but context length did so much for ChatGPT, [27:36] did so much for coding. And with an image model, it's able to reliably [27:40] reference these references. [27:42] It's incredibly capable. Yeah, for sure. [27:45] And we're still trying to improve that as well. [27:47] It's not perfect today. We're really trying to develop this visual creation [27:53] layer for people because every single person you have an aesthetic or [27:57] personal style or preference. [28:00] And we're really trying to imbue that into the product that we're building [28:04] so that people can get to the output that they're wanting easier and faster [28:08] with ImageGen. Any parting prompt tips for people? [28:12] Well, one of the things I would suggest people try is ImageGen thinking. [28:17] So if you navigate to the thinking or pro models, we have a more powerful [28:21] version of ImageGen in that experience. [28:24] And in that model, you actually are able to search the web, analyze files, [28:29] leverage tools under the hood, which then yields a better quality and higher [28:35] composition photo. [28:37] And the suggestion that I have for prompting that experience is be [28:41] open-ended. I think the model will go and do the exploration itself to [28:46] understand and try to reason and find information that matters. [28:51] And I also think giving it a sense of an aesthetic is also super helpful. [28:56] Using and grounding that in a style has been really fruitful for a great [29:00] result. Good one. Good one. [29:02] I think just being very particular about the [29:04] style or like what you like in general. [29:07] Like for me, I like minimalist infographics. [29:10] Sometimes I think the model can be a little dense. [29:12] And so I just maybe I'm just a simplistic kind of guy so I just like [29:17] very very clean a very clean look so I like that. [29:20] Adele, Kenji, thank you very much.