Author: Compiled by Techub News
This is the content of OpenAI's official podcast, Episode 19. Host Andrew Mayne engaged in an in-depth conversation with researcher Kenji Hata and product lead Adele Li about GPT Image 2.0 (i.e., ImageGen 2.0). This conversation took place approximately two weeks after the model's official launch—by which point the weekly number of generated images had already exceeded 1.5 billion, with multiple usage trends rapidly going viral globally. This is not just a product launch retrospective, but also a candid discussion about a paradigm shift in image generation technology.
From Investor to Product Lead: A Story of Role Transition
Before joining OpenAI, Adele Li spent her entire career in investing. She worked at a private equity firm and Redpoint Ventures, focusing on early-stage investments in AI and software. When she joined OpenAI, she initially worked on planning data and computing infrastructure, far removed from image generation. However, over the past six months, she gradually transitioned to the product side, taking full responsibility for the ImageGen product.
She frankly stated that the essence of a product manager's role is "doing what needs to be done," no matter what that is. And the ImageGen project, in particular, allows her to leverage a variety of skills—working closely with researchers like Kenji, while also constantly thinking about where the market gaps are and where the windows of opportunity lie.
"This is no longer the market it was when ImageGen 1.0 was released a year ago." Adele said. Now, there are multiple competitors in the image generation space, and ChatGPT itself is a very different product. In this context, thinking about the evolving role of ImageGen within the ChatGPT ecosystem is one of the things she finds most interesting.
Kenji Hata also joined OpenAI about two years ago. He initially worked on an audio project, later had the chance to participate in pre-launch work for ImageGen 1.0, and gradually shifted to full-time image generation research, working all the way through to version 2.0.
Data First: Two Weeks Post-Launch, 1.5 Billion Images Per Week
Within the two weeks following the official launch of GPT Image 2.0, image generation usage on ChatGPT increased by over 50%, with the weekly number of generated images surpassing 1.5 billion. At the same time, various usage trends spread rapidly around the world—from color analysis and sticker styles popular among Asian users, to crayon drawing and graffiti styles favored by American users, to name a few.
Adele believes this viral spread itself illustrates a point: users perceived the leap in model capability almost instantly. "Visual communication feedback is the most direct." She said users don't need to read technical reports; they open the model, generate an image, and instantly know whether it's good or not.
Host Andrew shared the same feeling—the magnitude of this capability improvement is so significant that he thinks calling it "2.0" is an understatement; it feels more like a completely new paradigm. So, how did this paradigm shift actually happen?
Three Core Breakthroughs: Text, Multilingualism, and Photorealism
Adele and Kenji attribute the capability leap of ImageGen 2.0 to simultaneous breakthroughs in several key dimensions.
First is text rendering capability. Early image generation models were nearly disastrous at rendering text within images—deformed letters, jumbled words, chaotic layouts. Andrew jokingly said that the "OpenAI" text generated by early DALL-E looked like it was written by a chimpanzee. Now, the model can clearly and accurately render large blocks of text, even complex infographics, within an image.
Kenji quantified this progress using an internal test: ask the model to generate a grid image containing 100 random objects, then count the accuracy rate. From 5 to 8 correct objects in the DALL-E 3 era, to about 16 in ImageGen 1.0, to a stable 25 to 36 in version 1.5, and now version 2.0 can achieve nearly 100 correct. "This isn't a sudden leap, but sustained, steady growth." Kenji said.
Second is multilingual support. The team specifically strengthened the model's ability to understand and generate multiple languages during training. After launch, active feedback from users in Asia and Europe confirmed this direction was correct—users in different language environments can obtain high-quality, localized image outputs.
Third is photorealism. This was one of the most concentrated pain points in user feedback previously: characters generated by older models often had an "overly airbrushed magazine cover feel," with distorted facial and body proportions, lacking realism. Version 2.0 did extensive work in this area, aiming to make images "look more like yourself." Kenji recalled his feeling when he first saw the checkpoint output of the new model: placing it side-by-side with ImageGen 1.0 results, no discussion needed, the difference was clear.
The image he described was a scene of a woman looking out over the sea. "We looked at the two images. Didn't say anything. Just... okay, this one wins."
How to Achieve Both Speed and Quality? The Key is the Post-Training Stage
Andrew raised a question many are curious about: the model has gotten smarter, but the generation speed hasn't slowed down. How was this achieved?
Kenji explained that each version accumulates a large amount of engineering learnings. Taking speed as an example, the team did extensive work to improve the model's "token efficiency"—using fewer tokens to generate higher quality images. This is a process of continuous optimization across each version iteration, not something achieved through a single technological breakthrough.
Adele supplemented this by emphasizing the importance of the post-training stage. She said that when training this model, the team not only has to make the model understand world knowledge—how science, concepts, and mathematics are presented in images—but also answer a more subjective question: What is considered "beautiful"? What is considered "tasteful"?
These questions have no standard answers, yet they directly determine the upper limit of the model's output quality. To this end, the team worked closely with a large number of artists, designers, and marketers, attempting to distill the aesthetic judgments and best practices from these professional fields into the way the model interacts with users.
The team also closely monitors user feedback on social media, incorporating real-world usage issues into the iteration cycle. Kenji said these issues are either mitigated or completely fixed in the next version.
Behind the Viral Trends: Expressing an "Imperfect" Self with AI
Among the usage trends that emerged post-launch, one that the team found both surprising and interesting was users taking this extremely powerful model and specifically generating images that look rough and crude in a "Microsoft Paint style"—"downgrading" celebrity photos or popular images into pixelated doodles.
Adele offered an insightful interpretation of this: "Making an AI generate something 'imperfect' actually requires very high intelligence." This is not a failure of the model; on the contrary, it demonstrates the model truly understanding the user's intent.
She believes this reflects a consumer psychological trend: people crave authenticity, imperfection, and nostalgia. Crayon styles, graffiti styles, retro pixel styles—these viral prompts all point to the same theme: users want to use AI to show a more authentic, more fun side of themselves, not just pursue "perfect output."
"Self-expression through AI is the direction we are truly excited about," Adele said. This also aligns highly with OpenAI's mission—enabling more people to express that "self that was previously impossible to express."
From Entertainment to Productivity: Penetration into Education, Design, and Cross-Industries
Another important shift for ImageGen 2.0 is its move from primarily entertainment use cases towards being a genuine productivity tool.
In the field of education, the team has an internal channel specifically for educators, covering teachers from elementary school to graduate level. Kenji shared a memorable case: a biology professor inputted graduate-level textbook content into the model, which generated highly accurate diagram pages, stating that the content was completely correct.
Adele believes that converting complex concepts into easy-to-understand visual content is one of the model's strongest capabilities. She specifically mentioned the direction of "personalized learning"—teachers can use ImageGen to generate customized learning materials for students with different language backgrounds or preferences. This is a direction she and the team are actively exploring: how to more deeply integrate ImageGen into ChatGPT's learning scenarios, so that concept teaching naturally comes with visual presentation.
In workplace scenarios, Adele revealed an interesting internal data point: over 50% of slides in OpenAI's internal presentations already use images generated by ImageGen. "The penetration rate of visual communication is much faster than we anticipated."
Furthermore, she listed various professional groups already using ImageGen: real estate agents use it to generate property showcase images and virtual staging effects, YouTube creators use it for video thumbnails and promotional materials, artists use it to connect with fans, writers use it to quickly generate images for social media posts...
Host Andrew also shared his own firsthand experience: he gave the model the cover of his book and asked it to generate promotional images suitable for different social media platforms. It got the correct proportions and style on the very first try. "It feels like magic."
360-Degree Panoramas, Sprite Sheets, and Synergy with Codex: Surprises from Emergent Capabilities
Beyond the expected capability improvements, version 2.0 also brought some "emergent capabilities" that the team itself didn't fully anticipate.
360-degree panoramas are one such example. The team found that as they supported generation with arbitrary aspect ratios, users began spontaneously generating ultra-wide panoramic images, even 360-degree wrap-around style images. The team then turned this capability into a product feature. Now users can directly generate and immersively browse 360-degree panoramas on ChatGPT's web and mobile platforms. Andrew immediately used it to generate a 360-degree version of "dogs playing poker," sitting from the dog's perspective and looking around.
Sprite sheets also became an unexpectedly popular use case. Game developers and indie creators are using ImageGen to generate multi-pose sprite sheets for game characters. Combined with Codex's code generation capabilities, they can build a small game with custom characters from scratch. Andrew described what he witnessed: saying "I want a crow" in Codex, watching the system automatically call the ImageGen tool, iteratively generate the crow's sprite sheet, and then Codex integrates it into the game code. "That is magic."
Multi-image consistency is another significant improvement in version 2.0. Kenji mentioned that users are already attempting to create 10-page comic books with coherent storylines, where character appearances and visual styles remain highly consistent across multiple images. This capability previously required significant manual intervention and techniques in workflows, but is now more reliable and fluid.
Next Steps: Creative Agents and Personalized Visual Assistants
Discussing future directions, Adele outlined a clear vision: the Creative Agent.
The picture she describes is this: an AI assistant that truly understands your workflow, aesthetic preferences, and intended output goals, capable of acting as your personal interior designer, personal architect, personal wedding planner—all of which can be reflected in a single image.
The core of this direction is truly injecting "personalization" into every aspect of image generation. Adele illustrated this with her "me-me-me eval": she uses 100 photos of herself, friends, and family as an evaluation set to test whether the model can insert the correct personalized elements in the right context—for instance, if ChatGPT remembers she has a brother and what her parents like to do, when generating a birthday card, can the model naturally incorporate this information into the image?
Kenji added from the research side that the team is continuously optimizing multi-image consistency, the overall experience of the visual creation layer, and making it easier and faster for users to get the output they desire. "It's not perfect today, but we know where the direction is."
Regarding prompt engineering tips, both offered their advice. Adele recommends users try the "ImageGen thinking mode"—in the Pro or thinking model, ImageGen can search the web, analyze files, and call tools, resulting in better generation quality and composition. She suggests using open-ended prompts in this mode, letting the model explore and reason on its own, while giving it a clear aesthetic style as an anchor. Kenji prefers a more personal style; he likes minimalist infographics, so he explicitly tells the model to "keep it clean and simple."
If DALL-E was the Stone Age of image generation, then ImageGen 2.0 is its Renaissance—not just artistic progress, but a comprehensive fusion of science, art, architecture, knowledge, and aesthetics. At the end of this conversation, Adele used this statement to conclude, and perhaps this is the best way to understand the model: it is no longer just a "drawing tool," but a visual intelligence agent that is truly beginning to understand the world, people, and beauty.
