OpenAI's gpt-image-2 Just Launched. Here's What It Changes for Photo Editing

OpenAI shipped ChatGPT Images 2.0 on April 21, powered by a new model called gpt-image-2. It's a meaningful upgrade. The improvements in text rendering alone would have made it worth covering, but the addition of thinking mode and 2K output make this a genuinely different product from gpt-image-1.

Here's what actually changed, what it's best for, and where it fits in your workflow if you're already using OutfitGen for photo editing.

What's New in ChatGPT Images 2.0 / gpt-image-2

2K resolution output. gpt-image-1 topped out at 1K. gpt-image-2 doubles that to 2048x2048 natively, which matters a lot if you're doing anything that ends up in print or on a large screen. For in-browser editing it's less critical, but it does mean you have more resolution to work with.

Thinking mode. Before generating, gpt-image-2 reasons through the request. OpenAI says this is the same general approach as their reasoning models — the model works through the prompt, considers ambiguities, and plans the generation before touching a pixel. In practice this makes it noticeably better at multi-step instructions and edge cases. It also means generation can take 60 seconds or more when thinking is on.

Web search grounding. The model can pull real-world context while generating. This is a big deal for anything involving brand logos, current fashion references, real products, or anything that's changed recently. Ask it to generate a product photo with a recent release, and it knows what that product looks like.

8 images from one prompt. Batch generation is now first-class. You submit one prompt and get 8 distinct variations back. Useful for creative exploration or A/B testing visual concepts.

Dramatically better text rendering. Text inside images has been a persistent weakness of diffusion-based models for years. gpt-image-2 is genuinely excellent here — legible small text, proper multilingual rendering (Chinese, Hindi, Arabic), iconography, UI elements, dense compositions. If you've been frustrated by garbled words in AI image output, this is the fix.

~2x speed of gpt-image-1. When thinking mode is off, gpt-image-2 is roughly twice as fast as its predecessor. That's real, useful progress.

gpt-image-2 debuted at #1 on the Image Arena leaderboard at launch. It's available in ChatGPT, Codex, and the API now.

What gpt-image-2 Is Great For

From-scratch creative generation. If you need an illustration, a concept mockup, or a creative visual that doesn't start from an existing photo, gpt-image-2 is excellent. The thinking mode + world knowledge combination means you can give it a complex brief and get something genuinely well-considered back.

Text-heavy images. Infographics, blog header images with words, product packaging mockups, UI screenshots — gpt-image-2's text rendering is the best available right now from any model. If words inside the image matter, this is your model.

Brand and product references. Because of web search grounding, gpt-image-2 knows what things look like. "A photo of a product on a shelf next to a Hydro Flask" will actually look like a Hydro Flask.

Creative riffs and exploration. 8 images per prompt means you can generate a batch of variations and pick the best direction. Good for early creative exploration where you don't know exactly what you want.

What gpt-image-2 Isn't Optimized For

Identity-preserving edits on your own photo. This is the core use case for most photo editing: you have a picture of yourself, and you want to see it with a different outfit, background, or style — without losing what you look like. gpt-image-2 is a generation model at heart, not an edit-first model. It's capable of inpainting and editing, but identity drift (the model subtly changing your face, proportions, or features during the edit) is a known limitation. Thinking mode helps, but it doesn't solve the fundamental architecture difference.

Speed for iterative editing. If you're trying 15 outfit variations on the same photo, 60-second thinking-mode generations add up fast. Even in non-thinking mode, the ChatGPT interface is built for conversation, not rapid iteration on a single image.

Cost for volume. Via ChatGPT Plus ($20/mo) you get image generation as part of the subscription, but the interface isn't optimized for edit-heavy workflows. Via API it's pay-per-token — $8 per million input tokens, $30 per million output tokens. Running 100 photo edits gets expensive quickly.

Where OutfitGen Fits

We're not a general-purpose image generator. We're a photo editor. The specific job we're built for: you have a photo of yourself (or a product), and you want to change something about it while keeping everything else intact.

Our Standard mode runs on Flux 2 Edit, which is fast (~8 seconds) and clean. Our Pro mode routes to Nano Banana 2 (Google's Gemini 3.1 Flash image model), which is specifically trained for subject consistency — the technical term for "keep this person's face intact while changing their clothes." It's widely regarded as the best available model for identity-preserving edits.

That's a different job than what gpt-image-2 excels at. gpt-image-2 is built to generate and reason broadly. Our Pro mode is built to edit faithfully. Both are genuinely good at what they do. They're not really competing for the same use case.

A practical way to think about it: if you want to make an AI illustration with legible text, or explore a creative brief from scratch, use ChatGPT Images 2.0. If you want to see how you'd look in a different outfit, change the background on a product shot, or restyle an existing photo — use OutfitGen. Faster, cheaper, and built specifically for that job.

For a deeper look at how our models stack up against gpt-image-2 across specific editing tasks, see the head-to-head comparison.

Try It

5 free edits, no signup needed. Standard mode is ~8 seconds, Pro is ~20 seconds and does better identity preservation for photos that matter.

Start editing on OutfitGen →

Or go directly to the Clothes Changer if you want to jump straight into outfit edits.