Just now, OpenAI released GPT-Images-2.5:
Generation latency reduced by up to 50%, improved editing precision, new sketch hand-drawing input feature added, and API endpoints now feature separate fast and slow models for the first time.
https://x.com/OpenAI/status/2097394956457623964
Available to all users of ChatGPT, ChatGPT Work, and Codex, including free-tier users.
Translated at this rate, approximately 430 million images per day pass through this system, and the perceived improvement in generation speed far exceeds that of typical feature updates.
OpenAI announced that the combined generation volume of ChatGPT Images and the GPT-Image API reaches 3 billion images per week.
Demo惊艳：Chinese garbled text is gone
How powerful is GPT-Image-2.5? Let’s look at the demo directly.
The Chinese garbled text has disappeared!
This is the reference image:
Here is the output retro image:
Can you still tell the difference between real photos and AI-generated images?
Extracting and upgrading printed photos to high-definition digital photos is a breeze.
Want to try on clothes and see the visual effect? Just let ChatGPT handle it:
Input:
Output:
Upload the original image, the quilt is messy:
Outputted an image of a neatly folded quilt:
Extremely consistent scene, with no visible continuity errors.
Faster by half, revised to be orderly
The speed improvement is the most direct. OpenAI reports a reduction in generation latency of up to 50% compared to Images 2.0.
The rendering quality of lighting and texture has been improved simultaneously, resulting in higher fidelity to the subject in the reference photo.
Precise editing addresses a longstanding issue with image generation tools: when users request changes to a specific area, the model inadvertently alters parts that were meant to remain unchanged.
According to OpenAI, Images 2.5 can modify only specified areas while preserving the composition, lighting, and main subject characteristics of the rest of the image, with significantly improved stability in complex backgrounds and multi-subject scenes.
Users can also place comment annotations directly on the image to highlight areas that need modification, with an interface and workflow similar to design tools like Figma.
Consistency across multiple edits is the third improvement.
When repeatedly editing the same image in a long conversation, the edits from earlier rounds are preserved, and image quality does not degrade over iterations.
Higgsfield AI Product Lead Axultan Alimkulov commented on Flare:
What impressed us most was its understanding of what shouldn’t be changed.
Make a meaningful edit that preserves the original image’s characters, composition, and visual style.
Sketch and Templates: From Typing to Drawing
At the product level, Images 2.5 introduces four new features.
To access Sketch, type @Sketch in the ChatGPT chat box to open the sketch panel. Users draw a sketch and add a textual description of the style and details; ChatGPT then generates a final image based on the sketch as a reference.
OpenAI's examples include transforming a sketch of a room layout into an interior design rendering, or a character outline into an illustration. The barrier is not in drawing skill, but in capturing spatial relationships and approximate proportions to help the model understand the composition.
The template covers high-frequency formats such as posters, product images, and flyers.
Choose a template, enter your information and style preferences, and the model generates the image according to the preset structure, eliminating the trial and error of writing a prompt from scratch.
Prompt sharing allows users to share the complete prompt used to generate their image, enabling others to input their own photos and details to create personalized versions within the same framework.
A prompt for an "80s retro portrait" has been circulating on social media, allowing users to upload selfies and generate photos in the same style.
Four features point to the same change: the process of turning mental images into pictures no longer relies solely on typing!
Flare and Sunburst: API Split for the First Time
For developers, OpenAI is simultaneously launching two image models on the API: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst.
Flare emphasizes speed and batch generation, and OpenAI positions it as the default choice for most applications.
Suitable scenarios include social content, product experience images, rapid prototyping, and high-frequency image production.
Lucky Liao from the Manus review team provided more specific data: Flare achieved image generation speeds 2 to 4 times faster than GPT-Image-2 in their tests, and improvements in transparent background generation directly benefit the programmatic creation of brand assets, presentations, and web graphics.
Sunburst is designed for high-end creative workflows, trading longer generation times for finer editing control.
OpenAI's typical use case involves production-grade marketing materials and retouched product images for branding.
Adobe has confirmed the integration of the Images 2.5 model into Firefly.
Matt Chotin, Senior Director of Adobe Products, said that version 2.5 delivers faster generation speeds and improved resolution consistency, ensuring images remain sharp and realistic even after multiple rounds of refinement.
The separation of the two models corresponds to two distinct requirements: high-frequency, low-latency trading and high-precision customization.
Flare is suited for high-volume scenarios, while Sunburst is ideal for high-precision use cases—developers can choose based on their business needs without waiting for unnecessary precision.
This is the first time the OpenAI Image API has introduced product tiers based on speed and quality, following the same logic as the language model API tiers.
The naming of Images 2.5 continues the rhythm of OpenAI’s image product line: same architecture, with speed and accuracy taking a leap forward.
GPT-Image-1.5 at the end of 2024 used the same approach.
The interval between two 0.5 versions is less than a year, and the iteration frequency of image models is aligning with that of language models.
This is the world's most powerful text-to-image model, significantly leading the competition.
OpenAI's continuously improving multimodal capabilities are likely aimed at enhancing base models like GPT-7 by strengthening abilities such as computer use, in response to Anthropic's models.
Reference materials:
https://openai.com/index/introducing-chatgpt-images-2-5/
This article is from the WeChat public account "New Intelligence Yuan" (ID: AI_era), author: ASI Revelation, editor: Marco