第19集——图像生成的文艺复兴时刻

你好。

Hello.

我是安德鲁·梅恩，欢迎收听OpenAI播客。

I'm Andrew Mayne, this is the OpenAI podcast.

今天这期节目，我们邀请到了研究员秦健和产品负责人李爱莉，聊聊Images 2.0。

On today's episode, we're talking about Images two point o with researcher Kenji Hata and product lead Adele Li.

他们会谈谈为什么这个新模型是巨大的突破，开发过程中哪些评估最关键，以及现在大家用它做出了哪些东西。

They'll discuss why the new model represents such a major leap forward, the evaluations that mattered most during development, and what people are creating with it now that it's widely available.

如果说DALL是石器时代，那Images 2.0就是文艺复兴。

If DALL was the stone ages, Images 2.0 is the Renaissance.

它不仅在艺术和审美上非常出色，还能把科学、艺术、建筑这些元素融合在一张图里。

It's not only great artistically and aesthetically, but it also incorporates, you know, science, art, architecture all in one image.

我们一看就明白了，这比Images 1强太多了。

We looked at it we're like, all right, this is better than Images 1.

爱莉，能跟我们说说你是怎么成为这里的产品经理的吗？

Adele, tell me a little bit about how you became a product manager here.

我两年前加入的OpenAI。

So I joined OpenAI a little over two years ago.

在来OpenAI之前，我整个职业生涯都在做投资。

And before OpenAI, I was an investor my entire career.

哦，真的吗。

Oh, wow.

我之前在私募股权领域，待了三年在红点创投，投资AI和软件公司。

So I was in private equity and spent three years at Redpoint Ventures investing in AI and software companies.

我刚加入OpenAI时，原本是想做完全不同的工作。

And when I first joined OpenAI, it was for a completely different role.

我当时在想，怎么搭建我们的数据和算力基础设施。

I was thinking about how do we build out our data and compute infrastructure.

后来慢慢转到了产品这边，最近六个月一直在做ImageJunk。

And over time, made my way over to the product side and for the last six months have been working on ImageJunk.

你从一个领域跨界到另一个领域，这种转变挺有意思的，能用不同的方式发挥作用，确实挺棒的。

It's interesting how you style yourself going from one role than finding yourself into this space here, which is kind of cool, you know, to think about the idea that you have this sort of, you know, ability to be useful in different ways.

没错。

Absolutely.

我觉得产品经理就是要干任何需要干的事，不管是什么任务。

And I think the role of a product manager is just to do the job that needs to be done no matter what it is.

特别是做ImageJunk，能在这过程中发挥各种能力真的特别棒——既要和像Kenji这样的研究员合作，又要思考当前市场有哪些空白，我们该抓住什么机会。

And for Images in particular, it's been really awesome to flex a lot of different muscles when it comes to building products, working with researchers like Kenji, but also thinking about like what is the gap in the market today that we wanna fill and what is the opportunity that we wanna grasp here.

现在这个市场，跟我们一年前刚推出Image 1.0的时候完全不一样了。

It's not the same market that it was a year ago when we first released Images one point o.

现在的局面已经大不相同了。

Now it's a very different landscape.

现在市面上已经有好几家做图像生成的公司了。

There are multiple image generation makers out there.

而且ChatGPT这家公司和它的产品本身也已经变了。

And ChatGPT is a very different company and product itself too.

所以，去思考ImageJunk的演变，以及它在ChatGPT里的定位，对我来说特别有吸引力。

And so really thinking about the evolution of ImageNet and its role within ChatGPT has been really, really exciting to me.

Kenji，你是怎么开始做图像项目的？

Kenji, how did you end up working on images?

其实我刚加入OpenAI的时候，也是两年前的事了。

Actually, like, when I first started at OpenAI, I also started about two years ago.

我一开始在做某个音频相关的项目。

I was working on like some random audio project initially.

那是我接手的第一个项目。

Just it was my first project.

后来我就慢慢参与到了Images 1.0上线前的协助工作中。

And then at the time, I just found my way just working on helping them work on Images one point o prior to the launch.

之后我逐渐转向这个项目，最后就全职投入到里面了。

And so gradually, I moved more and more onto the project, and then I became full time on it, basically.

现在这个模型的反响怎么样？

What has the reception been like right now for the model?

上线过去两周，使用量已经增长了50%以上。

In the last two weeks since we launched the model, usage is up more than 50%.

每周在 ChatGPT 上生成的图片超过十五亿张。

More than 1,500,000,000 images are generated every week on ChatGPT.

全球各地都出现了热门趋势，比如亚洲流行色彩分析和贴纸，美国则流行蜡笔涂鸦风格，这些都火了。

And we've seen viral trends emerge across the world, all the way from trends in Asia for color analysis and stickers to US for crayon and scribble that are going viral.

还有很多人在探索各种新兴的使用方式。

But also a lot of people exploring emergent use cases.

这不仅展现了模型的强大表现力，也说明用户能立刻直观地感受到它的进步。

And I think it shows the dynamic range of the model, but also how people are able to visually grasp the advancement of the model almost immediately.

用户们用图片反馈说，这是他们见过的静态模型中画质最高、细节最棒的，这种反应真的特别棒。

I think the visual communication reaction that we've seen from our users for them to say, hey, this is the best highest fidelity, highest quality in a static model that we've seen has been really awesome.

这感觉像是一个巨大的转折，甚至都不该叫 Images 2，而应该说是一个全新的范式，因为它的能力实在太强了。

This felt like a really big shift, almost worthy of maybe not even being Images two, but almost like just a new paradigm because just the capabilities are through the roof.

是什么让这一切成为可能？

What made that possible?

我们刚开始做这个项目时，坐下来讨论过：我们想实现什么样的能力跃升？想支持哪些全新的使用场景？

When we started working on this project, I think we sat down and we discussed what is the step change of capability and use cases that we wanted to build towards.

我们相信，图像生成的潜力远不止现在这些功能。

And we believe that image generation has the ability to do so much more than what it does today.

你今天看到的所有输出和视觉内容，其实都可以转化成图像。

You could distill every single output or visual content that you see today into an image.

因此，我们定下的目标就是彻底改进这一点。

And so that was the mandate that we sought out to improve.

在这个2.0版本中，我们在多个方面都做了提升。

And with this two point o model, we've improved on various different dimensions.

首先是文字渲染。

One is text rendering.

页面上的文字质量有了大幅提升。

The ability for text on a page is so much better fidelity.

文字内容不仅通顺，而且都是真实的词语。

The language and words actually make sense and they're actual words.

其次就是多语言支持。

The second of all is multilingual.

所以我们重点优化了这个模型，让它能支持多种语言。

So we've really focused on making this model work in various different languages.

现在我们看到，亚洲和欧洲各地的用户都对这些改进反响很好。

And we're already seeing that people across the world in Asia and Europe are really resonating with these advancements.

第三点是写实效果。

The third is photorealism.

之前版本的反馈很多都说生成的画面不够真实，人脸或身体都被改得不像样。

I think we really saw a lot of feedback from our previous models that the output wasn't very realistic or altered their face or their bodies.

所以我们的一个目标就是，怎么让生成的画面更贴近用户本人。

And so one of our mandates was how do we actually make the image feel like more like yourself?

模型不只是机械地调用数据，它真正理解了世界的视觉逻辑，能把你想要的图像准确地呈现出来。

And so all the things that you think that the model knows it does because it hasn't viewed the knowledge of the world into its conscience and is able to visually communicate that back to you as a user.

把这些都整合起来，我觉得我们现在做出的这个模型，已经是市场上审美表现最出色的图像生成工具了，它代表了一种全新的图像生成范式——而这正是人工智能整体发展中，我们有机会去推动的重要一环。

And so putting that all together, I think we really get a state of the art image generation model that is the best aesthetic model out there on the market right now that really represents a new paradigm for image generation, which is a huge part of, I think, AI progress at large that we have an opportunity to work on here.

我们也会经常回听用户在社交媒体上的反馈。

We often listen back or listen to feedback on social media too.

所以我们把这些都汇总起来，保持关注，尽量在下一轮迭代中解决或彻底修复这些问题。

So we kind of just take all these things and basically are just aware of it and try to make sure that they're mitigated or completely fixed in some cases in the next iteration.

你们现在看到哪些使用场景？

What kind of use cases are you seeing?

现在人们都在用这个做些什么？

What are you seeing people do with this now?

我觉得对研究团队来说，特别贴近的一个用途是信息图和文字处理。

I think one that's particularly close to like the research team as a general is like infographics, text.

现在图像里的文字效果比以前好太多了。

I think text in images is like so much better nowadays.

这大大拓展了实用型的应用场景。

So I think it just opens up a lot more productive use cases.

从研究的角度来看，过去图像生成总被认为是娱乐性质的，甚至有点不务正业，但现在我们明显看到它正朝着生产力方向迈进，几乎任何你能想到的用途都能用上了。

And from like the research side, we kind of think, you know, image generation used to always be about fun and maybe like unproductive things, but now we're really seeing steps forward into productivity and image generation for any type of use case that you can imagine it for.

你刚才提到了文字。

So you mentioned text.

我记得早期的模型，不是说猴子不好，但那时候生成的东西，连OpenAI的都像猴子干的。

I remember the early models, no disrespect to chimpanzees, but getting to the spell like OpenAI even looked like a chimp did it.

现在我看的都是大段文字和精细到极致的内容。

And then now I'm looking at pages of text and finely detailed stuff.

我知道随着模型越来越聪明，变量绑定和元素排布的能力都在提升，但这次的进步真的特别大。

And I know that as models get smarter, variable binding, the ability to put things next to each other improves, but this was just a big improvement.

是的。

Yeah.

但我认为这其实并不算特别意外。

But I don't think it's like completely unexpected.

我觉得你其实能看到明显的进步，比如从DALL·E 3到GPD Images 1之间。

I think you you see a lot of growth in between well, first you see between DALL E three and, you know, GPD Images one.

如果你要求生成一组随机物体，DALL·E 3可能只能做到五到八个，而Images 1就能做到十六个左右。

There's if you ask for a grid of random objects, you you go from maybe, like, five to eight in DALL three to maybe around 16 in Images one.

到了1.5版本，我们直接提升到二十五到三十六个了。

And then with 1.5, we went to about 25 to 36 Mhmm.

一直都很稳定。

Consistently.

我觉得现在我们很可能能生成一百个以上。

And I think now we could probably do over a 100.

我觉得我们内部可能会做个测试，就是直接问ChatGPT。

I I think this is like a test that we might do internally is just we just ask ChatGPT.

给我列一百个随机物体。

Give me a list of a 100 random objects.

对吧？

Right?

是的。

Yeah.

然后我们把列表丢给图像生成器，看看能准确生成多少个。

And then we just send that to our image generator and see how how many are correct.

通常来说，差不多一百个都能对。

And usually, you know, it'll get almost all 100 correct.

但你能看到它随着时间持续增长。

And that's but you see the the constant growth over time.

所以我觉得这并不算特别意外。

So I don't think it's, like, completely unexpected.

只是稳步前进而已。

It's just a steady pace.

我以前用这个测试来考验早期的模型，比如Adele、Babbage和Curry，让它们列出一百本科幻小说。

That was a test I used to use for, like, the really old models back with, Adele, Babbage, and Curry, like, list a 100 science fiction books.

结果有些模型到二十二本左右就开始重复内容了，因为模型的容量到极限了。

And then some of them would get by the time it got to, like, 22, we'd just start repeating stuff as it relates to the model reached the end of it.

我们也见过类似三百六十度全景图这样的进展。

So we've seen stuff too, like three sixty, 360 degree panoramas.

这到底是怎么发生的？

How did that happen?

对。

Yeah.

这其实源于模型新出现的能力——能够生成任意宽高比的图像。

That really came from the emerging capability of the model, which is the ability to render images in any aspect ratio.

我们发现人们都在生成特别长的精美全景图，还有那种细长的书签式图片。

We discovered that people were generating really long, amazing panoramics, you know, skinny bookmarks as well.

这个模型的一个厉害之处在于，不仅能生成全景比例的图像，还能用360度风格来渲染画面。

And one of the cool capabilities of the model is that not only were you able to generate images in this panoramic aspect ratio, but you'd also render images in the style of three sixty.

对。

Mhmm.

我们发现，把这些图片放在360度环境中实际观看，特别有意思。

And we saw that it was really fun to actually view these images in a three sixty world itself.

于是我们就把这个很有趣的功能加进了产品里，现在在网页版和手机端的ChatGPT上都能用。

And so that was a really fun feature that we ended up adding into the product and it's available on ChatGPT on web and mobile right now.

我第一件事就是做了一个狗玩扑克的版本。

First thing I did was I made a version of dogs playing poker.

是的。

Yeah.

我把那个

I put that

放进去

in there

这样你就能像其中一只狗一样坐在那里四处张望，这完全出乎我的意料，不过挺有意思的。

so you could sit there like you're one of the dogs looking around in there, which was not something I expected, but it's it's it's fun.

对。

Yeah.

我是说，看到大家探索各种新用途，用模型创造出各种有趣的东西，真的特别棒，甚至远远超出了我们原本预期用户会用它来做的范围。

I mean, it's really awesome to see how people are exploring new use cases and fun things that they're creating with the model, even far beyond what we expected users to be using it for.

我觉得我们在设计模型的时候，确实花了很多心思去理解大家真正希望从图像生成中得到什么。

I think when we were designing the model, we were really deliberate in understanding what people really wanted to see from image generation.

图像生成领域其实隐藏着很大的需求。

There was a lot of latent demand in image generation.

你知道的，大家之前主要把它用在个人场景里，但我们确实看到很多人想尝试把模型往某些方向推，而这些方向恰恰是模型不太擅长的。

You know, people were mostly using it for personal use cases, but we definitely saw a lot of inklings of people wanting to push the model in certain directions that the model wasn't good at.

所以文本渲染确实是我们要重点改进的方面之一。

So text rendering was definitely one of those dimensions that we really wanted to improve on.

多语言支持是另一个。

Multilingual was another.

我觉得这个模型对世界的整体理解也提升了很多。

And I think world understanding generally is so much better in this model.

这意味着现在网上很多人分享了各种我们根本没想到的创意用例，都是用图像生成做的。

And that typically means that, you know, now people online are sharing a bunch of examples of them creating ImageGen for all different kinds of use cases that we didn't even think, existed out there.

我觉得这个新模型在不同风格的审美表现上特别出色，不管是搞笑表情包、给五岁孩子画的图，还是专业的商业演示稿，输出的多样性真的让人惊喜。

So I think the model's understanding of aesthetic beauty across multiple different outputs, whether that it's like a fun meme, an image for a five year old versus a professional, consulting deck, the expansion of opportunity and outputs is has been amazing to see in this latest model.

还有个挺有趣的现象，最近特别火的是把一些热门图片或人物照片，用模型生成那种看起来很粗糙的微软画图风格。

It's funny too how one of the things that was trending was taking popular images or photos of people and then having the model make like kinda janky looking Microsoft Paint versions

就是那种效果。

of that.

你当初有想到会有人用这么强大的工具，专门去搞这些搞笑又土气的图吗？

And did you think that was something you would see was that people are gonna use this incredibly capable tool to then go make, you know, these silly looking things.

是的。

Yeah.

挺有意思的是，要故意做出不完美的东西，其实需要很高的智慧。

It's funny because it takes a lot of intelligence to actually create something that is imperfect.

我经常跟别人这么说。

That's what I tell people all the time.

是的。

Yeah.

而且我们现在看到的网络热门趋势确实非常有意思。

And it's definitely very interesting in the viral trends that we're seeing online right now.

我觉得人们现在特别追求真实感、不完美和怀旧感。

One thing that I think people are really striving for is authenticity, imperfection, nostalgia.

从用MS Paint提示词、蜡笔风格，到各种其他生成效果，都能看到这种趋势。

We're seeing that in the MS Paint prompt, crayons, all different kinds of generations that people are creating.

这其实反映出消费者的一个核心需求：他们希望以更真实、更不完美的方式与AI互动。

And that really feels like the theme of consumers is they wanna interact with AI in a very authentic, imperfect way.

他们想展示自己的不完美，用AI帮自己看起来更好看，同时也能展现更有趣、更搞怪的一面。

They wanna show their imperfections and use AI to help make them look good, but also show a more fun and goofy side of themselves.

我觉得通过AI来表达自我，这件事真的让我们特别兴奋。

And I think that self expression via AI is something that we're really excited about.

而且说实话，这正是我们公司的使命之一：让每个人更容易学习、分享知识，同时也让他们能表达出以前可能根本无法呈现的自己。

And, you know, I think it's really part of our mission as a company to make it easier for people to learn more and distribute that intelligence, but also letting them express a version of themselves that maybe wasn't possible before.

健二，这个模型有没有哪个时刻让你觉得，哇，这下真的可以了？

Kenji, was there a moment with this model where you're saying to yourself, wow, I think this is ready to go?

训练过程中，我们会定期保存检查点，然后抽样看看效果怎么样。

You know, as it's training, we take a checkpoint and then like we just sample from it, right, and just see, okay, how good is this thing?

有一次我们抽样生成了一张图，一看，嘿，这比第一代图像强多了。

And I think like we just sampled a checkpoint model, an image, and we looked at it and we're like, all right, this is better than Images one.

我们就说，行了，就这样吧。

We were just like, okay.

我记得当初看DALL·E早期版本的时候，一开始出来的图都是那种飘渺、怪异、像触须一样的东西，我还问过研究人员，这种风格以后会消失吗？

I remember watching the iteration of one of the early versions of DALL E and how at first it was sort of the wispy sort of weird sort of the tendril sort of thing and talking to one of the researchers like, is is that gonna go away?

我觉得离那个目标就差两次迭代了。

It's like, yeah, I think two, probably two runs away from that.

然后一下子就做到了。

And then just like that.

能预测到这种变化让我特别惊讶，突然间所有东西都变得清晰锐利了。

The ability to predict that was amazing to me, and all of a sudden, everything got crisp and clear.

对。

Yeah.

还有，回想几年前，我玩过GAN，做过那些东西。

And then also, like, looking at you know, years ago, I played with, like, you know, GANs and, like, doing those things.

你得眯着眼才能勉强看出那是个皮卡之类的。

You you have to squint and say, I think it's a pickup truck or something like that.

是的。

Yeah.

所以你看，当发现这一切突然间进步这么大，确实挺有意思的。

So it's interesting what you see as you say, okay, this just all of sudden got much better.

是的。

Yeah.

我的意思是，这太明显了。

I mean, it was just very obvious.

你直接拿早期的检查点就行。

You just you just take the early checkpoint.

从它里面生成一张图。

You just sample an image from it.

然后再从另一个地方生成一张图，对比一下，你就明白为什么了。

And then you just sample an image from, you know, Images one and you just look at the two and you're just there's just there's just Why

我怎么就喜欢这种垃圾？

do I like this garbage?

我不清楚这图里到底是什么。

This is I don't know what the image was.

可能就是一张海边女人的照片之类的。

It might have just been, like, a picture of, like, a woman at a sea on the seaside, like Yeah.

俯瞰着海边。

You know, overlooking a seaside.

我们一看，就明白了。

We just looked at it and we're like, alright.

这根本不用怀疑。

There's like no no question.

是的。

Yeah.

最大的突破其实是写实感——从那种光鲜、理想化的杂志封面风格，一下子变成了看起来像一张真正高质量的照片。

That was the big the big the big jump was the photorealism of going from something that looked that was more of a a glossy idealized magazine cover to something that looked like a really good photograph.

所以你帮我理一下，除了算力增加，这到底是怎么做到的？

So help me understand, like, besides just more compute, how did this happen?

你们是怎么让模型不仅更强大，还能把出图时间从一小时缩短到这么短的？

How did you get a model that's much better and also that doesn't take an hour to generate an image?

我还记得DALL·E那会儿，我们得老老实实等着，告诉你我们想要什么。

The times are still I I remember in the DALL E days, like we would literally have to, you know, tell us what you want.

然后一小时后，这些图就发到Instagram了，现在它们直接出现在ChatGPT里，速度还更快了。

And then an hour later, it'd be on Instagram to now these things are in ChatGPT and it's faster.

它是怎么在变得更聪明的同时，还保持同样的速度呢？

How is it getting both more intelligent and you're maintaining the same speeds?

我觉得每次更新我们都学到了很多，比如从1.0到1.5，再到现在的2.0。

I think we learned a lot in each release, like between one, one point five, now two.

所以我们把每次积累的经验都用上了，比如速度这方面。

And so we take each of the learnings that we've made and we've, you know, like for example, speed, right?

比如说，我们能不能让模型更节省token？

You know, one of the things is like, oh, can we make the model more token efficient or something like that?

我们花了很多功夫，让模型用更少的token也能生成高质量的图片。

And, you know, we did a lot of work to make it to make it put produce very good images with less tokens.

我觉得这个模型的后训练过程特别有意思。

I think the post training for this model was very interesting

对。

Mhmm.

我们得仔细想，模型不仅要理解世界知识，懂科学、数学这些概念在图像里怎么表现，还得知道什么样的风格能打动用户。

In the sense that we really had to think about not only does the model understand world knowledge and how things look in, you know, science concepts, math, etcetera, in an image, but also what is the taste that will resonate with users?

到底是什么让模型的输出看起来漂亮？

You know, what makes the model or output beautiful?

怎么才能让生成的画面更真实？

How do you make it look realistic?

这些全是我们在做模型后训练时反复琢磨的问题。

These are all questions that we had to grapple with when we were post training this model.

因为对我们来说，最重要的是这个模型必须是目前审美水平最高的，不管生成的是专业作品还是个人创作，它都更有创意。

Because I think that one of the things that was really important for us was that this model was the strongest aesthetic model out there right now, Which means that it has more creativity in various different outputs no matter what that output is, if it's a professional output or a personal output.

这种多样化的训练和应用场景，让这个模型的训练变得特别有意思。

And so that range of training and the range of use case, I think made training this model a very interesting problem.

你平时有没有特别喜欢拿来测试的案例？

Do you have any personal favorite benchmark tests you like to do?

就是那种你一想就说：‘我倒要看看它能不能画出这个来。’

Things you say, I wanna see it make an image of this.

我有个自创的评估方法，叫‘我我我’测试。

I have a eval that I call the me me me eval.

明白了。

Okay.

其实就是一百张我和朋友、家人的照片。

It's essentially a 100 photos of myself and my friends and my family.

我把大家都摆出各种搞怪的姿势。

And I put everyone in goofy positions.

每个人都有我专门给他们做的贺卡或生日图。

I have about a card or birthday for every single person.

我觉得这个测试特别好，因为只有你自己最熟悉身边人的长相。

And I think it's a really great eval in the sense that you only know the people around your, you know, faces the best.

你同时也想让模型生成有趣、贴合生活的内容。

You also want to create funny things with the model and do things that are relevant.

作为产品经理，我在测试的不只是模型的基础能力有多强，更想知道ChatGPT能不能真正理解我在特定情境下的需求。

And so one thing for me as a product manager that I'm testing is not only is the raw capability of the model really great, but also does ChatGPT understand what I want in that context.

你知道吗，ChatGPT记得我有个弟弟，还有我爸妈，他们喜欢做什么。

You know, ChatGPT remembers, you know, that I have a brother, that I have a mom and dad and what they like to do.

所以，模型能不能在关键时刻准确地把个人化的细节融入图片里？

And so does the model accurately know how to insert pieces of personalization in the moments that matter in the images?

这些正是我在测试的内容。

These are things that I'm testing for.

你呢？

How about you?

除了我之前提到的网格图，那个大概是我用得最多的。

Besides the grid one I mentioned earlier, that's probably the one I've used the most.

有一阵子，我和Divya一直在研究写实风格。

For a while, I think Divya and I were doing a lot about photorealism.

是的。

We're Yeah.

我们特别努力地在推这方面。

Trying real hard to push on that.

简单来说，我知道Divya最喜欢的是一个女人拿着一瓶橙汁的画面。

Just basically, I know Divya's favorite one was, like, a woman holding an or a jug of orange juice.

我不知道你有没有注意到。

I don't know if you see Yeah.

网上有太多女人拿着橙汁瓶的照片了。

There's like so many images of a woman holding a jug of orange juice.

其实我觉得研究人员用的图像标准集比他们愿意承认的要更统一。

Well, I actually feel like the researchers had a more standard set of images like than they like to bleed on.

没错。

Yeah.

你还会看到那些标准测试图，比如一个人用左手写字，右手戴着手表，钟表显示特定时间之类的。

And you get like the standard can it do somebody writing with their left hand or watch on their right hand and a clock showing this?

我觉得图像质量最大的突破，可能是一张半满的葡萄酒杯照片。

I think the big leap of the images, probably one or 1.5 was like a half full glass of wine.

这个

The

酒杯都快满到杯沿了？

wine glass full of the rim?

对，满到杯沿。

Yeah, full of the rim.

对，我试过一些方法来引导它生成，但真的得描述得特别详细，比如里面得是红色的液体。

Yeah, And there were ways I was able to prompt it to do it, but it was, it was really had to get a really descriptive, like, you know, red liquid inside this.

这个提示词特别有意思。

This one is so fun to prompt.

有人问，它能生成像素级精确的图像风格吗？

There was a thing people said, oh, can it do like, you know, can it do like pixel accurate pixel image style art?

有人就说，不行，它做不到。

And somebody was like, no, it can't.

我想亲眼看看。

And I wanna hear that.

我就想，好吧，试试看。

I'm like, okay, let's try.

我后来发现，只要给它一个64乘64的网格，然后说‘在这儿画幅画’。

And I found out if I gave it like a 64 by 64 grid and I said, go draw the art in there.

它真就画出来了。

It did.

直接就把艺术作品填进去了。

It just was able to put art into there.

看到这样的效果，简直太惊人了。

And that was amazing to see those kinds of results.

这种模型的可引导性简直不可思议。

And that's the promptability of this is insane.

你该怎么预估这种事呢？

How do you plan for that?

是碰巧发生的吗？

Does it just happen?

你就会说，天啊，真没想到。

You're like, oh, wow.

这样是不是更容易理解了？

This is better understanding this?

人们用 ImageJan 的时候，给出的提示都特别模糊。

People come to ImageJan with very vague prompts.

是的。

Yeah.

把它改得更好一点。

Make it better.

让我看起来更漂亮些。

Make me look better.

就是让我更可爱一点之类的。

You know, make me cuter.

这些要求都太模糊了。

All these things are really vague.

我觉得模型和配套系统真正的任务，就是把这种模糊的需求转化成用户真正想要的效果。

And I think it's really the job of the model and the harness to distill that into actually what users want.

我认为这是模型经过长期训练形成的一种特质，我们真正发挥了它的潜力。

And I think that's a personality of the model that we've trained over time that we've really harnessed the power for.

说实话，这种模型还会产生很多让人意想不到的结果，而这种意外感，恰恰是使用ImageGen的乐趣所在。

And honestly, I think it also yields a lot of really surprising results that people may not expect, and that surprise is just part of the fun of using ImageGen.

我注意到大概出现了两种不同的提示方式。

I've seen, like, two kinds of prompting sort of emerge.

我还记得当初用DALL·E的时候，觉得自己简直是提示工程高手。

And I remember back with DALL E, I thought, like, oh, I'm a prompt engineer.

我觉得自己在这上面肯定特别在行。

I'll be great at this.

我肯定能做得特别出色。

Like, I'll be really good at this.

比如我生成个太空浣熊，就觉得特别有成就感。

And I and I'm, you know, make a raccoon in space, but I feel proud.

但后来我看到一些艺术家，他们根本不是什么提示工程师，而是来自艺术圈的人，我看他们用自己那一套语言来操作。

And then I'd see an artist, somebody who wasn't a prompt engineer, somebody who actually came from that world, and I'd watch them use their language.

他们做出的东西太棒了。

And they were doing amazing things.

是的。

Yeah.

这一点现在看来依然成立。

And that's that seems like it's that's still holding true.

没错。

Definitely.

我们在开发这个模型时，一直和一群艺术家紧密合作。

I mean, we work with a group of artists very closely when we develop this model.

我们深受艺术家、设计师、营销人员这些不同领域专业人士的启发，他们各自都有独特的做事方式。

And we're very inspired by artists, designers, marketers, all these different professions that I think have a different way of approaching their profession.

对我们来说，非常重要的一点是，既要吸收他们的灵感，也要借鉴他们的最佳实践。

And one of the things that was very important for us is we wanted to take the inspiration as well as the best practices

明白。

Mhmm.

针对这些职业，把他们的做法提炼成用户与模型互动的方式。

For those professions and distill that into the way that people interact with the model.

所以我们一直有意识地在这方面下功夫。

And so that's something we've deliberately tried to focus on.

我见过一个特别有效的方法，就是允许用户上传灵感或背景资料。

One hack that I've seen work really well is the ability to upload inspiration or context

嗯。

Mhmm.

把它们输入到模型里。

Into the model.

模型能非常出色地捕捉这些资料的精髓，并转化到输出结果中。

And the model has an incredible ability to take the spirit of that context and translate it into the output.

这其实挺有意思的，因为很多人担心，只要按个按钮，

What it's it's interesting because I think that a lot of people worry that, oh, I just push in a button.

就能得到一个漂亮的结果。

I get something beautiful.

每个新模型都在不断进步。

And and each each model, that gets better.

正如你所说，确实不用花太多力气就能搞定。

It's easier, as you said, to not have to put a lot of effort into it.

但要是人们愿意多花点心思，出来的效果会更惊人。

But when people do put effort into it, they are getting even more amazing results.

而且如果你有艺术天赋，掌控力会更强，因为现在模型更懂你所说的那些概念，比如景深，或者其他你想表达的东西。

And it seems like actually that if you're artistically inclined, you're getting even greater control because now it under like you said, it understands more about what you're talking about when you talk about depth of field and these other things or whatever you're trying to do.

就像你提到的，早期的模型就让不少艺术家兴奋不已——他们把原创作品上传进去，模型生成了各种变体，他们一眼就能看出哪个最对味，那种感觉就像艺术创作被彻底放大了。

And as you mentioned, it was exciting to see with earlier models, artists who said, oh, I gave it my originals and it gave me these variations and I know which one works and just seeing that as this this real creative amplifier.

是的。

Yeah.

没错。

Definitely.

我觉得，只要把你的创意方向、审美和判断力带进去，才能真正把模型的潜力发挥到极致。

I think having creative direction or taste or judgment and bring that to the model is the best way to push it further.

我觉得这个模型最让我兴奋的一点，是它极大地拓展了人们的创作空间。

I think one thing about this model that I'm really excited about is how it expands the creative outlet for people.

用这个Imogen模型来生成多种风格、类型或变体，以前从来都没这么容易过。

I think the ability to create multiple different styles or types or variations has never been easier than with this Imogen model.

它还能理解不同的使用场景，比如能自如地在生成建筑图纸和儿童读物的画风之间切换。

And I think it's also understanding of different contexts, like the way that it's able to shift what it's like to be generating architectural diagram all the way to the aesthetics of a children's book.

它能在这些不同领域间无缝过渡，真的特别出色。

The ability for it to move so seamlessly across these factors has been really awesome.

制作精美的信息图和图表，这个能力非常强大。

The ability to do great infographics and diagrams is very powerful.

你在研究和教育领域的人群中，听到过什么样的反馈？

What kind of feedback have you been getting from people in research and education?

我们内部其实有个测试通道，专门用来试用我们的模型。

We actually have an internal alpha channel where we test our models.

在这个通道里，还有一个子频道，专门给各个层次的教育者使用，从小学生到研究生都有。

And in that, there's like a subchannel dedicated specifically towards educators of any level, like elementary school students all the way up to graduate level.

我看到一个特别酷的案例，有个生物教授贴了一些研究生教材里的图，内容我完全看不懂。

One of the coolest things I saw was there was a biology professor, and he put like these graduate level textbook rendering pages of things I had no clue about.

但他说是完全准确的。

And he said it was perfectly accurate.

我觉得这个模型能把特别复杂的内容，用图像简洁明了地呈现出来，这是它最厉害的地方之一。

I think the ability for this model to distill very complex topics into something that is really easy to understand within an image is one of its strongest capabilities.

我们看到很多学生和老师用 ImageGen 来理解新概念，做学习笔记，甚至定制专属的学习材料。

And we've seen this with students, with teachers who are using ImageGen to learn different concepts, to also help them create study guides, to help also create personalized content.

个性化学习是我们特别关注的一个大趋势。

I think personalized learning is a huge trend that we're very passionate about.

我觉得 Imagen 能帮老师做出每个孩子都能看懂、符合他们个人风格和偏好的内容。

And I think the Imagen model helps you as a teacher create something that every kid can understand in their own language and their own preference.

这一点真的让我们特别兴奋。

And that is something that we're really excited about.

我们也在想，怎么把 Imagen 的这些功能更好地整合进 ChatGPT，让学习者在理解概念时，能直接通过图像来学习。

We're thinking about this in the context of also how do we bring more of the elements of Imagen into ChatGPT at large so that when people are trying to learn concepts, we're teaching them with ImageGen.

我记得上学的时候，在多媒体还没这么普及之前，海报特别流行。

I remember when I was in school and kind of prior to a lot of kind of multimedia blowing up, posters were a big thing.

教室里贴着解释各种知识的海报。

Classroom poster explaining stuff.

这让我想起信息图有多强大，因为它能让你把注意力完全集中在上面。

This really reminded me of how powerful an infographic can be because it allows you to bring as much attention as you want to it.

是的。

Yeah.

你可以花时间慢慢看，细细琢磨，还能塞进更多细节。

And you can spend the time looking at it and seeing it, and you can put a lot more detail into it.

我觉得用Imagen后，一个特别明显的改变是，现在我们内部的演示文稿里，超过一半的幻灯片都是用Imagen做的。

I think one really awesome visual shift that I've seen with Imagen is that now in internal presentations, over 50% of the slides are created with Imagen.

真厉害。

Wow.

当你要解释概念、说明想法时，用图像来沟通的力量真的特别大。

And that permeation of communication via images is so powerful when you're trying to explain your concepts or illustrate what you mean.

我觉得，这种模型在信息图和文字排版、页面布局方面的能力真的非常强大。

And I think infographics and the text rendering capability as well as the composition of the text on the page is incredibly powerful with this model.

这个模型不仅懂该说什么，还知道该怎么呈现，这简直是个超能力。

The model's understanding of not only what to say but how to present it is a superpower.

我对未来的探索特别期待，比如我们怎么才能把它做得更好？

And I'm really excited about future explorations of this where we can think about how do we make this even better?

怎么优化布局？怎么拓展不同的输出形式？还有怎么让产品里能直接编辑？

How do we improve the composition, the different kinds of outputs, and also make it editable in the product?

这些方向我们都特别看好。

These are directions that we're really excited about.

你认为这会怎么发展？

How do you see the progression of this?

这确实很棒，但每次我跟OpenAI的人聊他们正在做的东西，他们总会说，这不错，不过……

This is great, but typically anytime I talk to somebody with OpenAI about what they're working on, they're like, yeah, this is good, But

我觉得我们在探索各种实际应用场景方面，才刚刚起步。

I think we're still super early in exploring all the different use cases that people are really trying to push the model with.

我们特别期待的是，图像2.0的下一步，是打造一个创意助手。

And so one of the things that we're really excited about is what is that next stage for Images 2.0.0, is to create the creative agent.

最终，这个助手能跟你并肩工作，当你的创意伙伴，真正懂你的工作方式、你的偏好，以及你想要什么样的成果，帮你构建一个生态系统，让你拥有一个专属的室内设计师、建筑师，甚至婚礼策划师，全都集成在一张图里。

Ultimately, the agent that can work alongside you, be your creative assistant, and really understand how you work, what your preferences are, what is the output that you wanna get to, and build the product and model ecosystem that helps users kind of have a personal interior designer, personal architect, personal, you know, wedding planner, etcetera, all in one image then.

说实话，那真没什么特别惊艳的。

I tell you, nothing that was kind of amazing.

就是，嗯，还行吧。

It was like, alright.

书。

Books.

所以每隔一阵子，我的书就出一本。

And so, like, every now and I have a book come out.

我就得换一下社交媒体的封面图。

I've gotta change my social media headers.

我就直接说：帮我找一下我那本书的封面，然后写上‘no’。

And I just went and I said, oh, find my book cover and write no.

生成一个尺寸合适的社交媒体封面图，我好用在 X、Facebook 或者别的平台上。

Create a you know, create appropriate sized social media header that I can put on X or Facebook or whatever.

比如说，先出第一版。

Like, let's say first shot.

第一版。

First shot.

比例要对，所有细节都得到位。

Right aspect ratio, everything.

我们从一开始就做到了，或者从一开始就训练模型能很好地处理各种比例。

We basically did that from the start or trained the models to be good at that from the start.

我记得我当初负责初期的开发，确保模型能生成你要求的任何比例。

I remember like I worked on the initial de risk of every basically, it could do any aspect ratio that you ask.

是的。

Yeah.

没错。

Yeah.

现在你可以轻松地直接指定你想要的效果。

You can now really just easily specify the outcome that you want.

对。

Yeah.

比如你自己，你就说：我需要宣传材料。

Like in the case of yourself, you're like, I want promotional material.

但我自己也没啥具体想法。

I don't have an idea.

我根本没说清楚到底想要什么。

I didn't specify exactly what I wanted.

但模型自己去做了调研，然后按适合你的风格和比例把内容给你出来了。

But the model was able to do the research and then give it to you in the style and aspect ratio that was relevant to you.

这太强大了。

And that's super powerful.

我们已经看到这种趋势了。

We're already seeing this.

你知道，你本身就是一个作家。

You know, you're you're an author.

我接触过一些房地产中介，他们用Imagen来为公寓制作房源图片，或者布置展示效果。

I've talked to real estate agents who are using Imagen to help them create listings for their apartments or stage their listings.

一些YouTube创作者也跟我聊过，他们用Imagen制作视频封面和宣传素材。

YouTube creators have talked to me about using Imagen for their thumbnails and promotional content.

我还跟一些顶尖艺术家聊过，他们想用Imagen来跟粉丝互动。

I've talked to top artists who want to use Imagen to connect with their fans.

我觉得，各行各业的人现在都能用Imagen来辅助视觉创作，这特别强大，尤其是当你从事视觉或创意行业的时候。

And I think the ability for all different kinds of professions to start to use Imagen to help them with visual creation is super powerful, especially if you're working in a visual and a creative industry.

Imagen简直就是你职业工具箱里的神器。

Imagen is such a hack in your professional toolkit.

我觉得，未来它一定会成为每个人日常工作流程中不可或缺的一部分。

I think it has to be a part of everyone's everyday workflow in the future.

这感觉就像是，我第一次遇到这么一个东西，只要我能想到的点子，它都能很好地实现。

This does feel like the I think it feels like the first time where anything I can reasonably come up with that does a pretty good job of it.

我们认为，这为每一代人都带来了全新的范式。

We think it's a new paradigm for each generation altogether.

比如说，我们在DALL·E的发布视频里说那是石器时代，而Images 2.0就是文艺复兴。

Like, if you know, we set this in the launch video of DALL E was the stone ages, Images two point o is the Renaissance.

没错。

Yeah.

我觉得这确实没错，因为这个模型不仅在艺术和审美上出色，还把科学、艺术、建筑这些元素全都融合在一张图里。

And I think that is so true because the model, it's not only great artistically and aesthetically, but it also incorporates, you know, science, art, architecture all in one image together.

我觉得模型所具备的这种构图能力和知识储备，让生成的结果更可信、更强大，也打开了更多应用场景。

And I think that composition and knowledge that the model has just means that the outputs are so much more trustworthy, are more powerful, and enable so many more use cases.

我觉得Images和Codex的结合也非常棒，把图像生成和编程代理的能力完美地融合在了一起。

I think that Images and Codex is also amazing intersection of the capabilities that we're sending out to create with both Images as well as coding agents.

很多人用Imagen作为设计新网站或开发新应用的第一步。

So many people are using Imagen as the first step to designing a new website or creating a new app.

而当强大的视觉生成能力配上强大的编程能力，现在你就能零样本直接从零创建出非常惊艳的应用了。

And I think that intersection of having a really strong aesthetic model, which is image generation, in combination with strong coding abilities means that now you're able to zero shot really amazing apps from scratch with both of these tools.