Instruction Following with ChatGPT Images 2.0

YouTube AI Channels Models

Summary

ChatGPT Images 2.0 adds a layout-planning stage that enables pixel-perfect placement of objects, readable text in hands, and accurate non-standard clock times.

OpenAI researcher Jianfeng Wang explains how ChatGPT Images 2.0 can now follow highly detailed instruction in prompts, especially when it comes to spatial layouts and text rendering.
Original Article
View Cached Full Text

Cached at: 04/22/26, 02:26 AM

TL;DR: ChatGPT Images 2.0 now imagines spatial layouts first, then places every object and letter exactly where you ask, turning even shoulder-level magazine shots and off-beat clock times into pixel-perfect reality. ## From Imagination to Pixel-Perfect Layouts OpenAI researcher Jianfeng Wang introduces the core idea behind the upgrade: the model must first “imagine” the overall composition, then position each element correctly. This two-step mental sketch is what lets ChatGPT Images 2.0 follow long, nit-picky prompts without drifting. ## Magazine-Style Text Rendering ### Shoulder-Shot Challenge Prompt: *“A magazine-style text-art photo shot from a woman’s shoulder level. She holds the word ‘the’ in her right hand and the word ‘view’ in her left hand.”* Previous models struggled to attach readable text to specific hands while keeping the shoulder-level framing. The new release loads, then delivers crisp lettering on each raised palm, preserving the glossy editorial look. ## Clock Faces That Show the Right Time ### Breaking the 10:10 Habit Old behavior: almost every generated clock showed 10:10 because advertising datasets over-represent that “smiley” arrangement. New behavior: user requests for 2:25, 2:30, 9:10, and 7:45 all render the exact hour and minute hand angles. The improvement is visible across analog, digital, and vintage clock styles. ## Object Placement in 2-D Space ### Prompt Breakdown *“Place the apple dead-center, the mug immediately to its right, the book above the mug, the camera on the left, and the basketball below the apple.”* The model silently “thinks,” then outputs an image that keeps every item in the specified relative position. Wang highlights this as a “huge leap” in instruction fidelity. ## Take-away By adding an internal layout-planning stage, ChatGPT Images 2.0 narrows the gap between user intent and generated pixels—whether you need words in hands, non-standard clock times, or a precisely arranged still life. Source: [OpenAI on YouTube](https://www.youtube.com/watch?v=EcP7bzNAEn0)

Similar Articles

Slides & Infographics with ChatGPT Images 2.0

YouTube AI Channels

ChatGPT Images 2.0 in "Thinking" mode can turn 1,000-word prompts or 70-page PDFs into ready-to-use infographics, slide decks, and academic posters without manual editing.

Creating images with ChatGPT

OpenAI Blog

OpenAI's guide on using ChatGPT to generate and edit images from text prompts, covering best practices for writing effective prompts and iterative refinement techniques.