OpenAI's ChatGPT Images 2.5 adds sketch-to-image, better identity matching, and multi-step edits. Here's what changed, tested hands-on.
What is ChatGPT Images 2.5? #
ChatGPT Images 2.5 is OpenAI’s updated image generation and editing model built into ChatGPT, ChatGPT Work, and Codex. It focuses on three things: keeping a person or object looking consistent across edits, handling multi-step edit chains without losing earlier changes, and a new sketch-to-image feature that turns rough drawings into finished images. It’s available across desktop, mobile, and web on all tiers, and OpenAI also shipped two API variants, one more detailed and pricier, one faster and cheaper.
TL;DR #
- Identity consistency is the headline improvement: reference photos of real people now produce results that look noticeably closer to the original subject than the prior model managed.
- Multi-step editing finally holds context between edits, so changing a mug’s color and then adjusting lighting no longer resets earlier changes back to a blank slate.
- The new sketch feature lets you draw a rough shape directly in ChatGPT (via the plus menu, on desktop or mobile) and convert it into a realistic or stylized image.
- Testers reported the model feels roughly 50% faster than the previous version, though nobody independently timed it against a benchmark.
- Combined with GPT-5 Astra’s computer-use abilities, the image model becomes one link in a longer automated chain: generate an image, edit it, hand it to Canva, export a print-ready file, and upload it to a printing service, largely unattended.
- Guardrails around copyrighted or trademarked characters remain strict , sometimes stricter than expected, blocking sketch-to-image outputs that look too close to existing IP.
- A competing model reportedly code-named Spicy Mayo is expected to challenge Images 2.5 on detail accuracy, based on early side-by-side tests.
Other agents ship a demo. Remy ships an app. #
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
What actually changed from the previous model? #
OpenAI’s stated improvements center on four areas: sharper details, more natural lighting and textures, better handling of reference images, and more precise editing that persists across multiple prompts. That last point addresses the most common complaint about the earlier image model: editing one detail (say, changing a mug’s color) used to work fine, but asking for a second edit right after often reset unrelated parts of the image, undoing work that should have stayed untouched.
Testers who tried sequential edits, changing a mug’s color, then asking to keep everything else the same, saw the object hold its position and structure across each step, even as small things like lighting shifted slightly. It’s not flawless continuity, but it’s a real step up from starting over each time.
Reference-image handling also improved. Feeding the model a photo of a real person and asking for a stylized or repositioned version now produces results that read as recognizably the same person, something earlier versions struggled with, frequently drifting into a generic approximation instead. One tester generated YouTube thumbnail variations from a source photo and found the likeness held up convincingly across different poses and styles.
How does the sketch feature work? #
The sketch feature adds a new option under the plus (+) menu in ChatGPT, available on both desktop and mobile. Tapping it opens a small canvas where you draw with a mouse or finger. Once you confirm the sketch, it’s used as a reference image, and you can pair it with a text prompt describing the desired output, for example, “make this look super realistic” or “reimagine this as an oil painting.”
Results vary with how much detail you put into the sketch and how ambitious the prompt is. Simple sketches (a house, mountains, a river) turned into coherent landscape paintings without much fuss. More specific attempts, like sketching a recognizable cartoon character, sometimes triggered content guardrails that blocked generation if the output looked too similar to existing copyrighted characters, even from a rough drawing.
The feature also supports iterative editing on top of a sketch-derived image. You can mark up the generated image directly, asking for an added hat or text overlay, and the model will apply those changes while keeping the earlier edits and the subject’s identity intact, a workflow that would have broken down under the older model.
Is ChatGPT Images 2.5 actually better at consistency? #
Based on multiple independent tests, yes, though with caveats. Side-by-side comparisons of stop-motion and claymation style sequences showed the new model producing noticeably smoother, more coherent frame-to-frame results compared to the prior version, which tended to have subjects “jump” or shift unpredictably between generated frames. One tester noted the very first frame in a sequence still tends to be the highest quality, with small drifts appearing as the model extrapolates later frames.
Remy is new. The platform isn't. #
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Editing consistency held up well in controlled tests too: changing one element of an image (a mug’s color, a croissant swapped for a coffee cup on a poster) left the rest of the composition, layout, and framing intact. That said, testers were careful to note this isn’t pixel-perfect stability. Lighting and background details can still shift subtly even when the main subject stays put.
Where the model still runs into trouble is complex, multi-branch content like detailed flowcharts. Testers found that intricate diagrams with several decision branches sometimes got logic reversed (yes/no paths swapped) or created loops that didn’t make sense, suggesting the model’s gains in visual consistency haven’t fully carried over to complex structured content like text-heavy diagrams.
Can it replace a graphic designer’s workflow? #
Not on its own, but it can do more of the pipeline than before when paired with an agent that can operate other apps. One demonstrated workflow used ChatGPT Images 2.5 to generate and edit a poster image, then handed control to GPT-5 Astra (OpenAI’s computer-use agent) to move the image into Canva, lay out full poster text and design elements, export a print-ready PDF, and upload it directly to an online printing service, all from a single prompt and one permission click. The full run reportedly took about half an hour.
The output stayed editable at every step: the Canva file could still be adjusted, and the chat history could regenerate earlier stages if something needed changing. That non-destructive quality matters more than the speed, since it means a single bad step doesn’t force starting the whole job over.
This kind of chained workflow (image generation, then editing, then file conversion, then upload) is still early and demonstrated by individual testers rather than an official, packaged OpenAI feature. It depends on Astra’s browser and desktop automation capabilities working reliably alongside the image model, not on the image model alone.
How does it compare to other image models right now? #
Testers who tried side-by-side prompts against Google’s Nano Banana line found the results close in some categories and behind in others. In one detailed-scene test recreating a Minecraft-style screenshot, a rumored upcoming model (referred to informally as “Spicy Mayo”) edged out both Images 2.5 and the earlier Nano Banana 2 on small details like inventory icons, though Images 2.5 still performed better than expected, getting most interface elements recognizable.
For photorealistic edits of real people and objects, and for multi-step editing specifically, testers generally rated Images 2.5 as the strongest option available at the time, with the caveat that stylized, painterly outputs (closer to Midjourney’s aesthetic) still aren’t its strength. It’s built for accuracy and consistency, not artistic flourish.
Frequently Asked Questions #
What is new in ChatGPT Images 2.5 compared to the previous version?
The main upgrades are stronger identity preservation from reference photos, edits that persist correctly across multiple sequential prompts instead of resetting, sharper detail and lighting, and a new sketch-to-image feature. Testers also reported it feels faster, with OpenAI citing roughly 50% speed improvements.
How do I use the sketch feature in ChatGPT?
Open the plus (+) menu inside a ChatGPT chat on desktop or mobile, select the sketch option, draw on the canvas that appears, and confirm it. Add a text prompt describing how you want the sketch transformed, and the model generates an image using your drawing as a reference.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
Is ChatGPT Images 2.5 available in the API?
Yes. OpenAI released two API variants: one with more detail at a higher cost, and one that’s faster and cheaper with slightly less detail. Both are usable for building the model into apps or automated pipelines.
Does ChatGPT Images 2.5 still have content restrictions?
Yes, and in some cases the guardrails appeared stricter than expected at launch, blocking sketch-to-image generations that resembled existing copyrighted or trademarked characters even when drawn casually. These restrictions may loosen over time, as has happened with prior OpenAI image releases.
Can ChatGPT Images 2.5 generate accurate images of real people?
It can produce noticeably closer likenesses from reference photos than the previous model, including grabbing publicly available images of a named person to inform the generation. Without an explicit reference and without web search, results are closer to a loose, generic approximation rather than an accurate likeness.