ChatGPT Image 2.0 Gives You Superpowers
TL;DR
Riley calls GPT Image 2 'best in the world by far' because it wins across UI, text, edits, branding, portraits, and 3D mockups — his examples include believable Rust screenshots, fake YouTube and ChatGPT interfaces, product-style branding, and highly accurate app screen renders.
The most jaw-dropping demo was a generated book cover with a barcode that actually scanned to the real title — Riley tested both Good to Great and The Intelligent Investor, then blacked out the ISBN in Canva to confirm the barcode itself still worked.
Multi-edit image manipulation is suddenly absurdly precise — in one prompt, GPT Image 2 correctly handled 11 separate changes to a photo of Riley, including replacing Red Bulls, updating on-screen text to 'vibecode.dev,' adding a pink diamond earring to his left ear, and changing his shirt into a brown turtleneck while preserving composition 'down to the pixel.'
The model’s underrated superpower is image reasoning plus annotation overlays — Riley fed it an '80s Europe' comic and asked for red handwritten-style notes explaining cultural references like the Berlin Wall and Chernobyl, turning image generation into a visual explainer tool.
For builders, the real shift is that image generation is becoming an agent tool, not just a chat prompt — inside OpenAI’s Codex, Riley had an agent pull bookmarked tweets from Readwise, research them, generate annotated GPT Image 2 slides, and export the whole presentation to Canva.
It’s not magic everywhere: counting is still a clear failure mode — when he asked for a 4K image with 175 people and one purple dinosaur, the model’s follow-up numbering overlay mislabeled people, skipped numbers, and ended at 263.
The Breakdown
GPT Image 2 arrives, and Riley immediately declares a blowout
Riley opens with pure conviction: OpenAI’s GPT Image 2 is "best in the world by far, and it’s not even close." He runs through why he’s so hyped — game screenshots, UI generation, text rendering, branding, portraits, photorealism, fantasy art, even 3D modeling — and frames the whole thing as a business and content-creation unlock.
The barcode demo that made him say, "what?"
His first personal test is normal enough: upload four photos and get a magazine cover with realistic likeness and clean text. Then he gets weird in the best way, asking for a Good to Great book image whose barcode scans to the real book — and it works. He repeats it with The Intelligent Investor, then blacks out the ISBN in Canva just to prove the scanner is reading the generated barcode, not just text nearby, and basically loses it on camera.
Eleven edits in one prompt, and it mostly nails every one
Next he pushes image editing hard, asking for 11 distinct changes to a single photo of himself. GPT Image 2 updates the coffee cup text to "Riley Brown," removes Red Bulls, changes a monitor to "vibecode.dev," swaps signage to "GPT image 2," turns a bobblehead into a monkey, adds a pink diamond earring to his left ear, gives him a skin fade, and rewrites the sticky note to "Keep winning." Riley’s main takeaway: it preserved layout almost perfectly while executing a long list of instructions in one shot.
Cartooning Riley, then turning images into explainers
He also tests caricature generation, and the model notices details from the source photos — his ears, coffee references, even background objects. Then he asks for a 2D comic version of himself in the 1980s and follows that with a Europe-themed variant. The most interesting move is his "overlay explanation" prompt: GPT Image 2 adds red handwritten-style annotations explaining references like anti-nuclear imagery, the Berlin Wall, and Chernobyl without changing the underlying image.
Getting started in ChatGPT, plus the select tool and app mockups
Riley shows the simple entry point in ChatGPT: click "create an image," upload references, and wait about 20 to 30 seconds. He likes the new UI and highlights the select tool, which lets you loosely brush over part of an image — like hair — and issue a targeted edit such as recoloring it white. Then he stress-tests one of the biggest commercial use cases: iPhone app mockups, uploading five Vibe Code app screenshots plus a logo and getting surprisingly pixel-faithful phone renders, especially after refining composition with a stronger reference image and a quick annotated screenshot from CleanShot Pro.
Playground control, 4K output, and the first obvious limitation
He points out that OpenAI Playground exposes GPT Image 2 with more granular controls, including 2K and 4K output, though it requires separate API billing. There he tries a crowd-generation prompt with 175 people and one purple dinosaur, then asks the model to label every person to verify the count. That’s where things break: duplicate labels, skipped numbers, overlapping detections, and a final total of 263 — his first clean example of a limitation.
The real story: Codex can now use image generation as a tool
Riley saves his favorite part for last: GPT Image 2 is built into Codex, OpenAI’s agent-style app he compares to Cursor, Claude Code, Lovable, and even document tools like PowerPoint creation. Instead of manually prompting each image, he tells Codex to check Readwise, gather his saved tweets, create a presentation, and generate annotated GPT Image 2 visuals for every slide. The result is a 10-slide deck exported to Canva, and Riley’s bigger claim is the punchline: agents will likely end up prompting GPT Image 2 more than humans do.