GPT Images 2.6 突然发布,蓝色大肥鱼实现 GIF 表情包自由GPT Images 2.6 suddenly released: blue fat fish unlock GIF sticker freedom


AI free token/AI免费积分

猝不及防。

就在刚刚,OpenAI 正式发布新一代图像生成模型 Images 2.6,并同步向 ChatGPT、Codex 和 API 开放。

此次升级带来了几项关键变化。

一是图像生成速度提高 50%,二是兼容借助文字提示词生成 GIF 动图,三是进一步改善多轮编辑的一致性,让 AI 生图开始更接近专业外观素材的生产要求。

Image

OpenAI Developers 官方账号也同步宣布,面向开发者推出 GPT-Image-2.6 Flare 和 Sunburst 两款模型,分别侧重快速生成和高质量图像编辑。

更值得关注的是,OpenAI 在官方博客中披露了一组使用数据。

如今,全球使用者每周借助 ChatGPT、Codex 和 API 创建的图片已经达到 40 亿张,平均每天超过 5 亿张。当 AI 生图已经进入这样一个量级,行业的重点需要也就逐步从基础生成,转向追求更精细化的修改。

Image

与此同时,据 The Information 援引知情人士消息称,OpenAI 最快可能在当地时间周五推出新一批模型,涵盖新一代 Astra、Sol 以及专注网络安全的模型,不过具体发布时间仍有可能调整。

如果消息属实,Images 2.6 或许只是 OpenAI 本周密集更新的开始。

生图速度提高 50%,两款新模型分工明确

按照 OpenAI 的说法,Images 2.6 的图像生成速度较此前提高了 50%,这意味着使用者可以在更短时间内拿到结果,也能更高频地调整提示词、反复修改画面。

在 API 端,OpenAI 这次给出了两款定位不同的图像模型。

其中,GPT-Image-2.6 Flare 主打速度和日常创作。

在低画质设置下,Flare 最快只需 6 秒就能生成一张图片。除了更快,它还改善了海报、机型展示图等专业视觉内容的布局质量,让模型在处理构图和元素排布时更稳定一些。

Image

另一款 GPT-Image-2.6 Sunburst,则更强调画质、细节和多轮编辑中的一致性。

按照 OpenAI 的说法,在中高画质设置下,Sunburst 不仅能给出更好的图像质量,也能在连续修改过程中更稳定地保留人物、机型和品牌细节。

这实际上对应的是 AI 生图在实际使用里最常见的痛点。

很多时候,使用者不是只生成一张图就结束,而是还要继续改背景、调动作、换文案、替换部分元素。如果模型一改就把人物长相、服装或机型外观一起改掉,那么前面的工作基本就得重来。

Sunburst 想应对的,正是这种「只能重画、很难细修」的难题。

对于需要制作系列海报、广告素材或品牌视觉内容的团队来说,这种一致性往往比单张图片是否惊艳更关键,因为真正影响工作效率的,从来不是第一稿,而是后面那一连串修改。

这一点,相信每个被折磨到凌晨三点的外观师深有体会。

一句话生成 GIF,让静态图片动起来

这次 Images 2.6 最直观的新能力,是正式兼容 GIF 动图创作。

依据 OpenAI 展示的案例,使用者可以直接用自然语言描述角色动作和变化过程,让模型生成循环播放的 GIF,用于社交媒体表情包、品牌角色动画、营销内容甚至前期分镜探索。

例如,OpenAI 展示了 OpenArt 的两组案例。

一组是股票下跌时崩溃流泪的 Q 版上班族。

Image

另一组则是股票上涨后兴奋庆祝的角色动画。两段内容都使用六帧连续画面,在保持人物、镜头和背景一致的前提下,借助细微动作变化完成完整情绪表达。

Image

从官方展示的提示词来看,Images 2.6 在生成动图之余,可以理解一整套连续动作逻辑。例如它需要同时处理人物设定、镜头固定、帧数结构、前后衔接,以及循环播放时的自然过渡,这也意味着模型开始具备了一定程度上的「动画编排」能力。

OpenArt 联合创始人兼 CEO Coco Mao 也在博客中表示,GIF 本来就是社交媒体上的原生表达形式,而 GPT-Image-2.6 让创作者可以借助一段提示词就完成这类内容的制作。

OpenAI 还展示了另一个更有代表性的方向,就是把简单草图直接转成动态角色。例如在官思路例中,一只画在横线纸上的手绘猫头鹰,被转换成了可以活动的动画形象。

Image

从这个角度看,Images 2.6 正在把 AI 生图往轻量级动画生产工具又推了一步。

笔者也第一时间上手感受了这个能力。

最近 DeepSeek 蓝色大肥鱼火遍全网,相信不少人都在微信群、评论区刷到过它的各种表情包。

Image

图|图片来源:博主 ZipZipPipe

而有了 GPT Images 2.6,只要脑洞够大,普通使用者也可以随时为自己喜欢的角色定制表情包。

例如,「震惊」版蓝色大肥鱼。

Image

「俏皮点赞版」蓝色大肥鱼。

Image

还有经典「吃大白饭」的蓝色大肥鱼。

Image

相比过去需要在网上四处寻找合适的表情包,现在只需要描述人物、动作和情绪,就能按照自己的想法定制聊天素材,甚至围绕同一个角色制作整套表情包。

在实际测试中,笔者也尝试让模型生成一个六帧循环的 Q 版魔法学徒 GIF,角色从挥动魔杖施法到咒语失控、脸被熏黑,再到抖落黑灰回到起手动作,整套动作链条可以按提示较完整地展现出来。

Image

另一组测试则是基于一张黑白小鸟线稿直接生成可爱风 GIF,模型保留了原有线条风格,只补上粉色腮红,并让小鸟完成眨眼、蹦跳和爱心飘出的循环动作,整个过程耗时约 1 分 6 秒。

Image

笔者还测试了一组拟人化橘猫揉面的六帧 GIF,模型基本还原了揉面、喷嚏、面团粘鼻子再舔掉的连续动作,总体风格和温馨面包房背景也保持了较好的统一性。

Image

当然,从 OpenAI 当下展示的能力边界来看,Images 2.6 更合适这种短时、循环、小动作的 GIF 生成,它与长视频、复杂镜头运动或者更强叙事性的生成模型,仍然不是同一条机型路线。

不过,对于表情包、品牌 Mascot、机型交互动效和社交内容创作来说,这已经足够实用。

OpenAI 还在文中提到,AI 办公助手初创公司 Town 就利用 Images 2.6,为机型中的虚拟角色 Monty 制作动态成效,让它在使用者等待 AI 回复时,以动画形式浮现在输入框上方。

Image

相比传统的加载图标或跳动圆点,这种更具性格和识别度的动态角色,显然更容易成为机型感受的一部分。

兼容 SVG 导出,AI 修图也有了更多控制选项

除了 GIF,Images 2.6 这次更关键的一层更新,实际上发生在编辑和工作流层面。

首先,OpenAI 将此前已经在 ChatGPT 中给出的图像标记能力带到了 API,这意味着开发者和团队可以更明确地告诉模型,到底要改哪一块内容,从而减少本来只想小修一下,结果整张图都被带偏的情况。

Image

同时,Images 2.6 也兼容 SVG 导出,方便将生成结果直接应用到网站、App 和印刷等情形中。

OpenAI 还新增了 complexity 参数,允许开发者控制模型在满足请求时投入多少规划工作。按照官方解释,图像质量参数决定的是画面的细节和保真度,而 complexity 决定的则是模型为了满足提示词要求,要在多大程度上开展情形规划和组织。

对于更换背景、调整单个元素这类简单任务,OpenAI 建议使用较低 complexity,以换取更快的编辑速度。对于包含大量元素、复杂关系或严格构图要求的情形,则可以尝试中高 complexity,让模型在理解和组织画面时投入更多处理能力。

在开放范围上,Images 2.6 已于今日面向全球 ChatGPT 和 Codex 使用者开放,覆盖所有订阅档位,兼容桌面端、移动端和网页端。 开发者则可以借助 API 使用 Flare 和 Sunburst 两款模型。

售价维度,Flare 和 Sunburst 使用相同的 API 定价,与上一代 Images 2.5 保持一致。

Image

按照标准费率,图像输入每百万 Token 为 8 美元,输出为 30 美元,缓存输入为 2 美元;文本输入每百万 Token 为 5 美元,缓存输入为 1.25 美元,实际费用取决于 Token 消耗量。

从此次更新来看,OpenAI 在保持 API 单价不变的同时,提高了生成速度,并进一步完善图像标记、SVG 导出和复杂度控制等能力,显然希望让 AI 生图更深入地融入专业外观和内容生产流程。

换句话说,相比已经逐步成熟的 C 端生图市面,专业工作流才是 OpenAI 接下来更难啃的部分。谁能让图片生成后的修改、衔接和交付更加高质,谁才更有机会让 AI 生图真正成为使用者离不开的生产力工具。

本文来源:https://www.ifanr.com/1687778
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

Without warning.

Just moments ago, OpenAI officially released its next-generation image generation model Images 2.6, simultaneously opening it up to ChatGPT, Codex, and API.

This upgrade introduces several key changes.

First, image generation speed has increased by 50%; second, it now supports generating GIFs using text prompts; third, further improvements have been made to consistency across multiple editing rounds, bringing AI image generation closer to the production requirements of professional-looking visual assets.

Image

Photo: 🔗 https://openai.com/index/introducing-images-2-6/

The official OpenAI Developers account also announced that two image models—GPT-Image-2.6 Flare and Sunburst—are now available for developers, respectively prioritizing speed and high-quality editing.

What's more noteworthy is that OpenAI disclosed usage data in its official blog.

Today, users worldwide are creating 4 billion images per week via ChatGPT, Codex, and API, averaging over 500 million daily. As AI image generation reaches this scale, the industry's focus is gradually shifting from basic generation toward more precise editing.

Image

Meanwhile, according to The Information, citing informed sources, OpenAI may roll out a new batch of models as early as Friday local time, including next-generation Astra, Sol, and a security-focused model, though the release timing remains subject to change.

If these reports are accurate, Images 2.6 might merely be the starting point of OpenAI's intensive update cycle this week.

Image generation accelerated by 50%, with clear division of labor in two new models.

According to OpenAI, Images 2.6 delivers a 50% boost in generation speed, enabling faster outputs and higher-frequency prompt tweaking and revisions.

On the API side, OpenAI has introduced two distinct image models.

GPT-Image-2.6 Flare prioritizes speed and day-to-day creation.

In low-quality mode, Flare generates an image in as fast as 6 seconds. Besides speed, it has also improved layout quality for professional visuals like posters and model showcases, offering more stability in handling composition and element placements.

Image

Another model, GPT-Image-2.6 Sunburst, emphasizes quality, details, and consistency across multiple editing rounds.

Open AI claims that under mid-to-high quality settings, Sunburst not only delivers superior image quality but also reliably preserves facial features, product models, and brand elements across consecutive revisions.

This addresses the most common pain point of AI image generation in real-world usage.

Typically, users don't stop at generating just one image; they continue tweaking backgrounds, adjusting poses, replacing copy, or swapping elements. If a revision alters facial features, clothing, or a product's look, all previous work must be redone.

Sunburst aims to tackle precisely this “can only redraw, difficult to refine” challenge.

For teams producing series of posters, ads, or branded visuals, this consistency often matters more than whether the first draft looks stunning, because what truly affects efficiency isn't the initial output but the subsequent chain of revisions.

Anyone who's been kept up until three in the morning by such revisions knows this pain intimately.

Static imagery, now brought to life with one-sentence GIF creation.

The most visually immediate new feature in Images 2.6 is its support for GIF creation.

Open AI's showcases indicate users can use natural language to describe character actions and transitions, and the model will generate looping GIFs suitable for social media reactions, brand character animations, marketing content, and even pre-production storyboarding.

OpenAI demonstrated two OpenArt cases.

One involves a chibi office worker who cries despondently when stock values plummet.

Image

The other is a character animation showing excited celebration after a stock price riseImage. Both consist of continuous six-frame sequences that convey complete emotions through subtle movements while maintaining consistency in character, camera angle, and background.

Image

According to the officially displayed prompts, Images 2.6 can understand a complete sequence of action logic in addition to generating animated images. For instance, it handles character settings, fixed camera angles, frame structures, continuity between shots, and natural transitions during loop playback simultaneously, indicating that the model has begun to develop a degree of "animation orchestration" capability.

Coco Mao, co-founder and CEO of OpenArt, also noted in a blog post that GIFs are a native form of expression on social media, and GPT-Image-2.6 allows creators to produce such content using just a single prompt.

OpenAI has also showcased another more representative capability: converting simple sketches directly into dynamic characters. In an official example, a hand-drawn owl on lined paper was transformed into an animated figure that can move.

Image

From this perspective, Images 2.6 is taking AI-generated images another step forward toward becoming a lightweight animation production tool.

I also had the opportunity to experience this capability firsthand.

Recently, DeepSeek's "Blue Big Fat Fish" has gone viral, and many people have likely seen various memes featuring it on WeChat groups and in comment sections.

Image

Photo | Source: Blogger ZipZipPipe

With GPT Images 2.6, ordinary users can now customize memes for their favorite characters anytime, as long as they have a wild imagination.

For example, the "Shocked" version of the Blue Big Fat Fish.

Image

The "Playful Thumbs-Up" version of the Blue Big Fat Fish.

Image

And the classic "Eating Big Mouthfuls of Rice" version of the Blue Big Fat Fish.

Image

Compared to the past, where one had to scour the internet for suitable memes, now users can simply describe the character, action, and emotion to customize their own chat assets, or even create an entire set of memes centered around a single character.

In practical testing, I also attempted to use the model to generate a six-frame looping GIF of a chibi magic apprentice. The character transitioned from casting spells by waving a wand to losing control of the spell, getting their face blackened by smoke, shaking off the soot, and returning to the starting pose. The entire chain of actions was displayed fairly completely according to the prompt.

Image

Another test involved generating a cute-style GIF directly from a black-and-white line drawing of a small bird. The model preserved the original line art style, only added pink blush, and animated the bird in a looping sequence of blinking, hopping, and blowing hearts. The entire process took approximately 1 minute and 6 seconds.

Image

I also tested a six-frame GIF of an anthropomorphic orange cat kneading dough. The model largely replicated a continuous sequence of the cat kneading, sneezing, getting dough stuck to its nose, and then licking it off. The overall style and the cozy bakery background also maintained a high degree of consistency.

Image

Of course, based on the current capability boundaries demonstrated by OpenAI, Images 2.6 is more suited for generating short, looping, minor-motion GIFs. It is still not on the same development track as generation models designed for long videos, complex camera movements, or stronger narrative capabilities.

Nevertheless, this is already practically useful for meme creation, brand mascots, machine interaction animations, and social media content creation.

OpenAI also mentioned in the text that the AI office assistant startup Town utilized Images 2.6 to create dynamic effects for Monty, a virtual character in their product. This allows the character to appear in an animated form above the input box while users are waiting for the AI's response.

Image

Compared with traditional loading icons or bouncing dots, a more distinctive, highly recognizable animated character is clearly more likely to become part of the device experience.

SVG export is supported, and AI editing now offers more control options.

Beyond GIF, Images 2.6’s more critical update this time actually lies in editing and workflow.

First, OpenAI brought the image marking capability previously available in ChatGPT to the API, which lets developers and teams more clearly tell the model which part of the content to change, thereby reducing cases where an intended minor tweak ends up producing unintended broader modifications to the entire image.

Image

At the same time, Images 2.6 also supports SVG export, making it easier to apply the generated results directly to websites, Apps, and printing use cases.

OpenAI also introduced a new complexity parameter, allowing developers to control how much planning effort the model invests when fulfilling a request. According to the official explanation, the image quality parameter determines the image’s level of detail and fidelity, while complexity determines how extensively the model carries out planning and organizing to meet the prompt requirements.

For simple tasks such as changing backgrounds or adjusting individual elements, OpenAI recommends using a lower complexity level to trade for faster editing speeds. For scenarios involving a large number of elements, complex relationships, or strict composition requirements, you can try a medium-to-high complexity level to let the model invest more processing power in understanding and organizing the scene.

In terms of availability, Images 2.6 is being rolled out today to global ChatGPT and Codex users, covering all subscription tiers and supporting desktop, mobile, and web platforms. Developers, meanwhile, can access the Flare and Sunburst models via the API.

On pricing, Flare and Sunburst use the same API pricing, which remains consistent with the previous generation Images 2.5.

Image

Photo|🔗 https://developers.openai.com/api/docs/pricing#image-generation

Under the standard rate, image input is $8 per 1 million tokens, output is $30, and cached input is $2; text input is $5 per 1 million tokens, and cached input is $1.25. Actual costs depend on token usage.

Looking at this update, OpenAI kept the API unit price unchanged while increasing generation speed and further enhancing capabilities such as image marking, SVG export, and complexity control. Clearly, it wants AI image generation to be more deeply integrated into professional visuals and content production workflows.

In other words, compared with the consumer image-generation market, which has gradually matured, professional workflows are the tougher nut for OpenAI to crack next. Whoever can make post-generation edits, handoffs, and deliveries higher quality is more likely to turn AI image generation into a genuinely indispensable productivity tool for users.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry