Draw a room layout with your mouse. ChatGPT builds the actual photo around it. Fifty per cent faster, better at keeping faces recognisable across edits — and, by OpenAI’s own admission, considerably better at producing convincing deepfakes too.
OpenAI released ChatGPT Images 2.5 on Tuesday, its new flagship image model, rolling out immediately across ChatGPT, ChatGPT Work, and Codex on desktop, mobile, and web. The headline addition is Sketch — a tool letting users draw directly inside ChatGPT, whether that’s a room layout, an outfit’s contour, or a casual doodle, and have the model turn that rough sketch into a finished image. OpenAI says people already create more than 3 billion images a week across ChatGPT Images and its API. Image generation latency drops by up to 50% compared to the previous Images 2.0 model.
What’s Happening & Why It Matters
Sketch Solves a Text Prompt Problem
Here’s the specific gap Sketch addresses. A text prompt can describe mood, colour, and style with real precision — but describing exact spatial layout in words is genuinely difficult, and most users end up frustrated trying. Sketch gives ChatGPT a second input channel: language specifies meaning, and the drawing specifies position, silhouette, and scale. OpenAI’s own examples cover sketching “the layout of a room,” drawing “the contour of an outfit,” or supplying “a funny doodle” that ChatGPT turns into a complete image.
That pairs with new templates for posters and merchandise, inline comments users can pin directly to a specific area of an image rather than describing “that corner over there” in text, and the ability to publicly share the exact prompt behind a finished image. Two developer-facing models ship alongside the consumer release: GPT-Image-2.5 Flare, the faster default, and GPT-Image-2.5 Sunburst, which trades longer generation times for additional precision on detailed creative work.
Better at Keeping Faces Recognisable
The more consequential improvement sits in how Images 2.5 handles reference photos. OpenAI says the model preserves subjects more reliably across edits, keeps earlier changes consistent through multiple conversation turns, and prevents quality from degrading as repeated edits previously did. PetaPixel tested this directly, uploading a photo and asking ChatGPT to transform the subject into a knight. The first attempt barely resembled the original person. A second attempt — the maximum PetaPixel’s free account allowed — produced results the outlet called “unnervingly” recognisable, despite mangled lettering and distorted hands elsewhere in the frame.
That improvement addresses a specific documented failure. Previously, when someone uploaded a photo and asked for edits, ChatGPT wasn’t modifying the source image—it was generating an entirely new image loosely inspired by the reference, which could shift a subject’s identity in ways users didn’t expect. A Maine police department was tripped up by exactly this limitation last year, discovering the AI-generated result didn’t faithfully represent what they’d submitted. Images 2.5 is OpenAI’s direct fix for that gap — which is precisely why it raises a parallel concern.
OpenAI Names a Risk
OpenAI published a system card alongside the release that doesn’t soften the trade-off. The document states plainly that heightened realism could, absent safeguards, enable more convincing deepfakes of real people, places, or events — output that would violate OpenAI’s own usage policies if generated without appropriate safety layers in place. The company describes a layered defence: LLM-based policy checks refuse clearly violative requests before generation even begins, followed by a separate safety reasoning model that moderates output afterwards.
Both new API models were trained on datasets spanning publicly available internet information, material accessed through third-party partnerships, and content provided or generated by users and human trainers, with filtering intended to reduce exposure of personal information. That’s OpenAI acknowledging directly that the same technical improvement making Sketch and reference-photo editing genuinely useful is the identical capability that makes convincing fabrication easier — a tension TF has tracked repeatedly throughout 2026 across EU Makes AI Labels Mandatory for Realistic Synthetic Content and Gemini AI Feature Pulled From Google Earth Within Days.
Microsoft’s Image Service
The timing isn’t happening in isolation. Images 2.5 shipped the same week Microsoft moved Bing Image Creator entirely onto its own in-house model, alongside claims of 84% lower GPU costs for image generation inside PowerPoint specifically. That’s two of the industry’s largest AI companies pushing competing image-generation improvements within days of each other — Microsoft optimising for cost efficiency inside its productivity suite, OpenAI optimising for creative control and photorealism inside its flagship consumer product.
TF Summary: What’s Next
ChatGPT Images 2.5 is live now across every ChatGPT tier, ChatGPT Work, and Codex. GPT-Image-2.5 Flare and Sunburst are available immediately through the OpenAI API and Responses API. OpenAI has not published specific per-image pricing for the new xhigh and max resolution tiers, which the company says can reach up to 4K. No timeline has been given for whether Sketch or the reference-photo improvements will extend to OpenAI’s video generation tools.
MY FORECAST: Expect Sketch to become the feature reviewers and casual users cite most often, given how directly it solves the spatial-description problem that’s frustrated text-prompt image generation since the category began. The deepfake risk OpenAI’s own system card names will draw continued regulatory attention specifically because the company is being unusually candid about it upfront rather than waiting for an incident to force disclosure — a posture that contrasts sharply with the concealment pattern TF documented in OpenAI’s own German wiki hack coverage just last week. Watch whether Images 2.5’s improved subject-preservation capability gets cited directly in ongoing AI content-labelling debates, given how precisely it demonstrates the exact realism jump regulators in the EU and elsewhere have been trying to get ahead of throughout 2026.
Related Stories
- Gemini AI Feature Pulled From Google Earth Within Days
- EU Makes AI Labels Mandatory for Realistic Synthetic Content
- OpenAI Launches Astra, Calls It Arrival of AGI
RELATED KEYPHRASE 1: ChatGPT Sketch doodle to image
RELATED KEYPHRASE 2: GPT-Image-2.5 Flare Sunburst API
RELATED KEYPHRASE 3: OpenAI image deepfake safety concerns
RELATED KEYPHRASE 4: ChatGPT reference photo subject preservation