Written by: Techub News Compilation

This is the content from the official OpenAI podcast, episode 19. Host Andrew Mayne engages in a deep conversation with researcher Kenji Hata and product lead Adele Li about GPT Image 2.0 (also known as ImageGen 2.0). This conversation took place about two weeks after the model was officially launched—during which time the weekly image generation count had surpassed 1.5 billion, and various usage trends rapidly gained popularity worldwide. This is not just a product release review, but a candid discussion about the 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 investment. She worked at private equity firms and Redpoint Ventures, focusing on early-stage investments in AI and software. When she joined OpenAI, she was initially responsible for planning data and computing infrastructure, which was quite far from image generation. However, over the past six months, she gradually shifted towards the product side, taking full responsibility for the product work of ImageGen.

She candidly stated that the essence of the product manager role is "doing what needs to be done," regardless of what that task may be. The ImageGen project, in particular, allows her to leverage various capabilities—working closely with researchers like Kenji while constantly considering where market gaps and opportunity windows exist.

"This is no longer the market we had a year ago with the release of ImageGen 1.0." Adele said. Now, there are multiple competitors in the image generation space, and ChatGPT itself has become a distinctly 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-oriented project but later, by chance, participated in the pre-release work of ImageGen 1.0 and gradually shifted to full-time work in image generation research, ultimately progressing to 2.0.

Data Speaks First: Two Weeks After Launch, 1.5 Billion Images Per Week

Within the two weeks after the official launch of GPT Image 2.0, image generation usage on ChatGPT increased by over 50%, with the number of images generated each week surpassing 1.5 billion. Meanwhile, various usage trends rapidly spread globally—from color analysis and sticker styles favored by Asian users to crayon and doodle styles celebrated by American users, among others.

Adele believes this viral spread itself indicates one thing: users have almost instantaneously perceived the leap in the model's capabilities. "The feedback on visual communication is the most direct." She said, users don't need to read technical reports; they can generate an image with the model and immediately know if it's good.

Host Andrew shared the same feeling—this significant enhancement in capabilities made him think that rather than calling it "2.0," it could be better described as a completely new paradigm. So, how did this paradigm shift happen?

Three Core Breakthroughs: Text, Multilingualism, and Realism

Adele and Kenji attribute the leap in ImageGen 2.0’s capabilities to several key dimensions of simultaneous breakthroughs.

The first is text rendering capability. Early image generation models faced disaster when processing text in images—letters were distorted, words were jumbled, and layouts were chaotic. Andrew jokingly mentioned that the "OpenAI" text generated by DALL-E in its early years looked like it was written by a chimpanzee. Now, the model can clearly and accurately present large sections of text within images, even complex infographics.

Kenji quantified this progress with an internal test: having the model generate a grid image containing 100 random objects and then tallying the accuracy. From 5 to 8 objects in the DALL-E 3 era, to about 16 in ImageGen 1.0, and stabilizing at 25 to 36 in version 1.5, the current 2.0 version can almost achieve nearly 100% accuracy. "This is not a sudden leap, but rather a steady and continuous growth." Kenji stated.

The second is multilingual support. The team specifically enhanced the model’s understanding and generation capabilities for multiple languages during training. After launch, the active feedback from users in Asia and Europe confirmed the correctness of this direction—users in different language environments can obtain high-quality localized image outputs.

The third is realistic photography feel. This was one of the most concentrated pain points in previous user feedback: images generated by old models often had a "magazine cover-style over-idealization," with distorted facial and body proportions, lacking a sense of realism. Version 2.0 has made significant progress in this area, aiming for images to "look more like yourself." Kenji recalled his feeling upon first seeing output from the new model's checkpoints: comparing it directly with ImageGen 1.0’s results, it was a clear distinction without the need for discussion.

The image he described was a scene of a woman looking out at the sea. "We looked at the two images and said nothing. It was just... okay, this one wins."

How to Achieve Speed and Quality? The Key to Post-Training Phase

Andrew posed a question that many were curious about: the model has become smarter, but the generation speed hasn’t slowed down—how is this achieved?

Kenji explained that substantial engineering learnings have accumulated with each version. For instance, in terms of speed, the team has put in a lot of work to enhance the model’s "token efficiency"—using fewer tokens to generate higher-quality images. This is an ongoing optimization process in every version iteration rather than being achieved through a single technological breakthrough.

Adele added to the importance of the post-training phase. She stated that while training the model, the team must not only enable the model to understand worldly knowledge—how science, concepts, and mathematics are represented in images—but also answer a more subjective question: what does "good-looking" mean? What does "tasteful" mean?

These questions have no standard answers but directly determine the quality ceiling of the model's output. To this end, the team closely collaborated with numerous artists, designers, and marketers, attempting to distill aesthetic judgments and best practices from these professional fields into the model's interaction with users.

The team also closely monitored user feedback on social media, incorporating real-world usage issues into the iterative loop. Kenji mentioned that these feedbacks are either mitigated or completely fixed in the next version.

Behind the Viral Trends: Using AI to Express an "Imperfect" Self

Among the usage trends that emerged after the launch, the team found one both surprising and interesting: users harnessing this highly powerful model specifically to generate rough, poorly crafted "Microsoft Paint-style" images—downgrading celebrity photos or trending images into pixelated doodles.

Adele offered an insightful interpretation: "To have AI generate something 'imperfect' actually requires a high level of intelligence." This is not a failure of the model; on the contrary, it is a reflection of the model’s true understanding of user intent.

She believes this reflects a consumer psychological trend: people yearn for authenticity, imperfection, and nostalgia. Crayon styles, doodle styles, retro pixel styles—these trending prompts all point to the same theme: users want to use AI to showcase their more genuine and playful sides, rather than just pursuing "perfect output."

"Self-expression through AI is the direction we are genuinely excited about." Adele said, which aligns closely with OpenAI's mission—to enable more people to express that "self that was previously impossible to express."

From Entertainment to Productivity: Education, Design, and Cross-Industry Penetration

Another important transformation of ImageGen 2.0 is its shift from entertainment-centered usage scenarios to becoming a true productivity tool.

In the education sector, there is an internal beta channel specifically for educators, covering teachers at all levels from elementary school to graduate school. Kenji shared a striking case: a biology professor inputting graduate-level textbook content into the model, generating highly accurate illustration pages, and indicating that the content was completely correct.

Adele believes that transforming complex concepts into easily digestible visual content is one of the model's strongest capabilities. She particularly mentioned the direction of "personalized learning"—teachers can use ImageGen to create customized learning materials for students with different language backgrounds and preferences. This is a path she and the team are actively exploring: how to more deeply integrate ImageGen into ChatGPT’s learning scenarios, allowing conceptual teaching to naturally come with visual presentations.

In workplace scenarios, Adele revealed an interesting internal metric: over 50% of slides in OpenAI's internal presentations already use images generated by ImageGen. "The penetration speed of visual communication is much faster than we anticipated."

Moreover, she mentioned various professional groups already using ImageGen: real estate agents generating property showcase images and virtual renovation effects, YouTube creators producing video covers and promotional materials, artists connecting with fans, and writers quickly generating social media images...

Host Andrew also shared his personal experience: he threw his book cover into the model and had it generate promotional images suited for different social media platform sizes, getting the correct proportions and styles on the first try. "It felt like magic."

360-Degree Panoramas, Sprite Sheets, and the Synergy of Codex: Surprising Emergent Abilities

In addition to the expected capability improvements, version 2.0 also brought some "emergent abilities" that the team didn't fully anticipate.

360-degree panoramas are one of these. The team discovered that while the model supports image generation with any aspect ratio, users began spontaneously creating ultra-wide panoramas, even 360-degree surround-style images. The team capitalized on this ability by developing it into a product feature, allowing users to directly generate and immerse themselves in 360-degree panoramas on ChatGPT's web and mobile platforms. Andrew was the first to use 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 independent creators are using ImageGen to generate multi-posed sprite sheets for game characters, which, combined with Codex's code generation abilities, allows them to create a mini-game with custom characters from scratch. Andrew described a process he witnessed firsthand: in Codex, he said, "I want a crow," and then watched as the system automatically called the ImageGen tool to iteratively generate the crow's sprite sheet, later integrated into the game code by Codex. "That's magic."

Multi-image consistency is also a significant improvement in version 2.0. Kenji mentioned that users are already trying to create comic books with 10 pages of coherent storylines, maintaining character appearance and visual style consistency across multiple images. This capability, which previously required extensive manual intervention and skills, has now become more reliable and seamless.

Next Steps: Creative Agents and Personalized Visual Assistants

When discussing future directions, Adele presented a clear vision: Creative Agents.

The vision she described is of an AI assistant that can truly understand your working style, aesthetic preferences, and target outputs, serving as your personal interior designer, personal architect, and personal wedding planner—all of which can be reflected in a single image.

The core of this direction is to truly inject "personalization" into every aspect of image generation. Adele used her "me-me-me eval" as an example: she compiled a test set of 100 photos of herself, friends, and family to test whether the model could naturally insert the correct personalized elements in the right scenarios—for instance, if ChatGPT remembers she has a brother and what her parents enjoy doing, can the model naturally incorporate this information into the image when generating a birthday card?

Kenji added from a research perspective that the team continues to optimize multi-image consistency, the overall experience of visual creation, and making it easier and faster for users to obtain the outputs they desire. "It's not perfect yet, but we know where we're headed."

On prompt techniques, both provided their suggestions. Adele recommended users try the "ImageGen Thinking Mode"—in Pro or Thinking modes, ImageGen can connect to the web to search, analyze files, and call tools, resulting in higher quality and composition. She suggested using open-ended prompts in this mode, allowing the model to explore and reason, while providing it with a clear aesthetic style as an anchor. Kenji leans towards personal style, preferring minimalist infographics, so he often tells the model to "stay clean and simple."

If DALL-E marked the Stone Age of image generation, then ImageGen 2.0 signifies its Renaissance—not just an artistic advancement, but a comprehensive fusion of science, art, architecture, knowledge, and aesthetics. Concluding this conversation, Adele encapsulated it with this statement, which perhaps best represents an understanding of this model: it is no longer just a "drawing tool" but a visual intelligence that is truly beginning to understand the world, people, and beauty.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。