Open AI's New Image Tool Holds Character Faces Across 8 Panels

What This Is

Last week, Sam Altman posted a six -panel comic on X. The protagonist was himself , hunting for GP Us across the globe . The character 's face stayed consistent across every panel, the English dialogue in the speech bubbles was clean and leg ible, and even the paper texture came through. The post hit 450 ,000 views in a single day. The tool behind it: Chat GPT Images 2.0 , released by OpenAI on April 21, running on a model called g pt-image-2.

This update has two core changes. First, a single prompt can now generate up to 8 images while keeping characters , style , and objects consistent. Previously , producing a comic st oryboard or a branded poster series meant generating images one at a time, re -describing the main character each time — and across 8 images , that character might end up with 3 different faces. Second, a "thinking mode" ( requires Plus or Pro subscription) lets the model search the web for real-time information, analyze uploaded files, and plan the image layout before rendering a single pixel. In effect , the " understand the task" step is now b aked into the generation pipeline .

Text rendering is another headline feature . Chinese, Japanese, and Korean characters have historically produced gar bled output in AI image generation. In W ired' s testing , an Images 2.0 poster containing over 20 instances of Chinese text came back correctly spelled with natural typ esetting. API resolution tops out at 2K, with 4K still in beta.

Industry View

W har ton professor Ethan Mollick, after several weeks of continuous testing, offered an interesting verdict : "There's a quality threshold I didn 't anticipate — once you cross it , you can generate high-quality written content, slides, and academic po sters." CNET's fr aming was equally bl unt: Images 2.0 is aimed at "people who need to produce good - looking content fast ." It 's not ch asing Midjourney's artistic and fant astical aesthetic , nor is it trying to be an Adobe-level professional editing tool.

The critic isms are worth taking seriously too . Mollick himself flag ged that the model becomes "very stub born" during iter ative editing — the first two rounds of adjustments work well , then it starts dra gging its feet. OpenAI has also acknowledged several limitations: the model has limited understanding of complex physical structures like orig ami and Rubik's cubes, and struggles with high-density repe ating details such as sand grains. Reddit users found that keyword -stuff ed prompts produce strange grid artifacts in the output — switching to natural language descriptions makes them disappear. All of this points to a tool better suited for "generating first drafts" than "pol ishing final deliv erables."

In direct comparison , Images 2.0 currently leads Midjourney V8 and DALL-E 3 on text rendering and single -prompt batch generation. Midjourney still holds an edge in artistic style exploration. The two tools aren 't really targeting the same user base to begin with.

Impact on Regular People

For enterprise IT: the cost of producing first drafts for brand materials, internal training illustrations , and product promotional images will compress further . That said, the review and revision loop between " first draft" and "final deliv erable" will still require human involvement in the near term.

For individual professionals : the barrier to handling " good enough " visual work — Power Point dec ks, proposals , event posters — is dropping . But if your core value is visual judgment rather than execution, the short -term impact is limited.

For the consumer market: use cases that previously required outs ourcing to a designer — children's picture books, person alized greeting cards, small - run print products — are increasingly viable for individuals to handle themselves. The pricing leverage of small design service providers in these ni ches will narrow .