Seventeen months after its previous release, OpenAI has refined its image-generation system not by leaping toward new frontiers but by deepening the ground already broken — making the creative act faster, more coherent, and more forgiving of the iterative human mind. ChatGPT Images 2.5 arrives as a quiet acknowledgment that raw capability alone does not serve the artist or designer; what matters is whether a tool can hold a conversation, remember what was asked, and preserve what was built. In this release, the technology bends closer to the rhythms of actual creative work.
OpenAI Launches ChatGPT Images 2.5 With 50% Speed Boost and Enhanced Detail Preservation
Each successive edit builds without quality loss
Mark
So the fifty percent speed boost—is that the main story here, or is there something deeper?
Mimi
The speed matters, but it's really about what you can do with that speed. The real change is that you can now edit an image multiple times in one conversation without everything falling apart. You ask for one change, then another, and the model remembers both.
Luke
How do we know that's actually true? Is that OpenAI's own testing, or has anyone independent verified it?
Mimi
OpenAI published a blog post making the claim. We don't have independent benchmarks yet, but the detail preservation across edits is the technical claim they're making.
Mark
And the Sketch feature—who is that really for?
Mimi
Anyone with a stylus, primarily. It lets you draw a rough layout and the model builds on it. For someone with an Apple Pencil, it's a more natural way to start.
Luke
But most people don't have a stylus. How useful is this feature if it's optimized for a device most users don't own?
Mimi
Fair point. It's a feature for a subset of users, not a universal upgrade. The real value is for designers and artists who already work with styluses.
Mark
What about the two models—Flare and Sunburst? How do you choose?
Mimi
Flare is fast, good for quick turnarounds on brand assets and presentations. Sunburst is slower but more precise, aimed at product photography and marketing where consistency matters.
Luke
And OpenAI is saying Sunburst is better for "pixel-perfect consistency." But what does that actually mean in practice? How much slower is it?
Mimi
They don't specify the exact time difference for Sunburst, only that Flare is two to four times faster than the old model. The trade-off exists, but the specifics aren't public yet.
Mark
What about the security features—watermarking and metadata?
Mimi
They're using C2PA metadata and invisible watermarking to mark images as AI-generated, which is meant to prevent deepfakes and unauthorized reuse.
Luke
Invisible watermarking can be stripped. Has anyone tested whether these protections actually hold up against someone determined to remove them?
Mimi
Not that we know of yet. It's a stated capability, but real-world durability is still an open question.
Le Pouls
- The persistent frustration of generative image tools — losing detail with every successive edit — has pushed designers into costly workarounds, and OpenAI is directly targeting that broken loop.
- Version 2.5 renders images up to 50% faster and introduces a Sketch tool that lets users draw compositional ideas by hand, lowering the barrier between intention and output.
- A new detail-preservation architecture allows users to edit backgrounds, clothing, and faces in sequence without degrading earlier elements — a technical trade-off that previous versions could not escape.
- Two specialized variants, Flare for speed and Sunburst for precision, let users choose their priority rather than accept a single compromise, while C2PA watermarking and invisible metadata guard against deepfake misuse.
- The release is live across ChatGPT's desktop, mobile, and web platforms, with API access for developers — signaling that the competitive pressure in generative image tools is now being fought on the terrain of workflow, not just raw output quality.
Seventeen months after its previous release, OpenAI has refined its image-generation system not by leaping toward new frontiers but by deepening the ground already broken — making the creative act faster, more coherent, and more forgiving of the iterative human mind. ChatGPT Images 2.5 arrives as a quiet acknowledgment that raw capability alone does not serve the artist or designer; what matters is whether a tool can hold a conversation, remember what was asked, and preserve what was built. In this release, the technology bends closer to the rhythms of actual creative work.
OpenAI released ChatGPT Images 2.5 on Tuesday, roughly seventeen months after the previous major version, addressing one of the most persistent frustrations in generative image tools: the erosion of visual coherence across multiple edits within a single conversation.
The headline improvement is speed — images generate up to fifty percent faster than before — but the more consequential change is how the model handles successive instructions. Where earlier versions would degrade unrelated details when a user edited one element of a complex scene, version 2.5 preserves facial features, identifying marks, and background elements across multiple rounds of editing. A user can refine a background, then adjust clothing, then revisit the face, and each change builds on the last without quality loss — even in busy scenes with multiple subjects.
The most visible new feature is Sketch, which lets users draw directly into ChatGPT with a finger or stylus. A rough compositional sketch becomes the visual foundation the model elaborates into a finished image, offering a more intuitive entry point for artists and designers. The update also adds templates for common formats, inline commenting for targeted edits, and prompt-sharing for collaborative iteration.
OpenAI is offering two variants: Flare, the default, optimized for speed and suited to brand assets and web graphics; and Sunburst, which trades speed for greater fidelity, aimed at product photography and precision marketing work. Both carry C2PA metadata and invisible watermarking to identify synthetic images and reduce deepfake risk — protections built into the model rather than applied after the fact.
The release is available immediately across ChatGPT's desktop, mobile, and web platforms, with API access for developers. It reflects a broader maturation in generative AI: as underlying models stabilize, the competitive frontier shifts from what the technology can do to whether a person can actually use it in the way they already work.
OpenAI rolled out ChatGPT Images 2.5 on Tuesday, a refinement of its image-generation system that trades incremental improvements for a fundamentally different creative experience. The update arrives roughly seventeen months after the company released ChatGPT Images 2.0 in April, and it addresses a problem that has dogged generative image tools since their inception: the difficulty of maintaining visual coherence across multiple edits and instructions within a single conversation.
The core upgrade is speed. ChatGPT Images 2.5 can generate images up to fifty percent faster than its predecessor, a meaningful acceleration for anyone who has watched a progress bar crawl across their screen while waiting for a render. But the speed bump is only part of the story. The new model produces images with what OpenAI describes as more natural lighting and richer surface textures—the kind of incremental visual polish that separates a tool from a craft. More importantly, it follows user instructions more reliably across multiple turns of conversation, a capability that matters less for one-off image requests and far more for the iterative workflows that designers and artists actually use.
The technical achievement lies in detail preservation. When a user provides a reference photo and asks the model to generate variations or make targeted edits, ChatGPT Images 2.5 now maintains distinctive features—facial characteristics, identifying marks, the specific qualities that make a subject recognizable—across successive generations. With the previous version, attempting to edit one element of a complex scene would often degrade other details in the image, forcing users to choose between precision and stability. Version 2.5 breaks that trade-off. A user can request an edit to a background while the model keeps facial features intact, then request another edit to clothing while preserving both the background and the face. Each successive instruction builds on the previous ones without quality loss, a capability OpenAI says works even in busy scenes with multiple subjects and complex backgrounds.
The most visible new feature is Sketch, a tool that lets users draw directly into ChatGPT using their finger or stylus on a touchscreen. A rough pencil sketch becomes the visual foundation for the model's generation—the user provides compositional direction or layout, and the model elaborates it into a polished image. For users with a stylus like an Apple Pencil, this opens a more natural entry point into the creative process. The tool also introduces templates for common formats—flyers, product photos—and inline commenting for targeted edits. Users can now share both their prompts and generated images with colleagues, enabling collaborative iteration on a single idea.
OpenAI is offering two distinct variants of ChatGPT Images 2.5, each optimized for different priorities. Flare is the default model, designed for speed. It generates images two to four times faster than GPT-Image-2, with improved support for transparent backgrounds, making it suited for brand assets, presentations, and web graphics where turnaround time matters. Sunburst sacrifices speed for precision, taking longer to render but producing images with greater fidelity and detail consistency—the choice for product photography and marketing materials where every pixel must align with brand standards.
On the security front, OpenAI has implemented guardrails on both prompts and outputs to prevent harmful content generation. The system uses C2PA metadata and invisible watermarking to mark generated images as synthetic, a technical measure intended to prevent unauthorized reuse and reduce the risk of deepfakes. The company says these protections are built into the model itself, not bolted on afterward.
ChatGPT Images 2.5 is available immediately across ChatGPT, ChatGPT Work, and Codex on desktop, mobile, and web. Developers can access both Flare and Sunburst variants through separate application programming interfaces. The release reflects a broader pattern in generative AI development: as the underlying models mature, the focus shifts from raw capability to usability, from "can it do this" to "can a person actually use this in their workflow." For image generation, that shift means tools that remember what you asked for, that let you sketch and refine rather than start over, and that preserve the details you care about while you iterate toward what you want.
Citations marquantes
The model generates images with more natural lighting and richer textures, and follows user instructions more reliably across multiple turns of conversation.— OpenAI
Flare creates high-quality images at two to four times the speed of the previous model, while Sunburst sacrifices speed for greater precision and detail fidelity.— OpenAI
Envie de l'histoire complète ?
Lire l'original sur SiliconANGLE ↗
Technology