ChatGPT Images 2.5 — AI image generation and editing
Bottom Line Up Front (BLUF): ChatGPT Images 2.5 is the most capable generation of the family for creating and editing images: details come out sharper, lighting looks more natural, and reference-based edits hold together far better. The interface gives you five rendering levels from a cheap draft to maximum, an exact canvas size, transparent background and up to sixteen references of 50 MB each. All in the browser, paid from your token balance, with a full history of every generation.
What is new compared to the previous generation
The new generation takes the main step where it matters: the picture itself. Texture, hair, small second-plane text are drawn more carefully; lighting stops over-brightening the scene and carries shadows honestly; a series of edits on one frame no longer falls apart. Editing is where the difference is most visible — background replacement, recolouring, object removal keep everything you did not ask to touch.
The control panel stays familiar: the same three canvas sizes, the same file formats, the same references and up to four variants per run. You pick the new model in the list and keep working as before — only the result for the same prompt comes out stronger.
Five rendering levels: from draft to maximum
On top of the familiar low, medium and high levels there are two upper ones: ultra high and max. Low is for trying ideas — fast and cheap, composition and angle are visible immediately. Medium is the working middle for social media and websites. High is a confident final for most tasks. Ultra high and max are for cases where fine texture and edges decide the outcome: print, large format, close crops.
The economical strategy is unchanged: cheap drafts at the low level, one expensive final from the best prompt. The difference is that the final now has two reserve tiers — if high is not enough, step up without switching models. The price is counted per picture and shown before the run, so there are no balance surprises.
Exact canvas size and custom proportions
Three proven sizes are offered: a 1024×1024 square, a 1536×1024 landscape and a 1024×1536 portrait — for them the price is known in advance and shown before the run. The model itself also accepts custom canvases: sides divisible by 16, proportions within 1:3…3:1 and a bounded pixel budget — so a non-standard frame for a specific platform is within reach.
Take the square for avatars, cards and catalogue tiles, the landscape for covers and headers, the portrait for stories and posters. Choose the size before you launch: re-cropping a finished picture almost always means losing an edge of the composition.
Transparent background, file format and compression
Transparent background is supported: an object without a backdrop is ready for a collage, layout or game asset straight away — no cutting out needed. The default file format is JPEG: a light file every platform takes without questions. Choose PNG when the picture will be edited further or needs transparency, WebP when page weight matters.
For JPEG and WebP a compression slider sits next to the format: the lower it goes, the smaller the file and the more visible the artefacts on smooth gradients. For clean scenes with crisp edges use stronger compression freely; for gradient skies and studio light keep the quality higher.
References and precise edits
You can attach up to sixteen images of up to 50 MB each. One shot plus an instruction is editing: replace the background, remove what is in the way, recolour an object. Several shots is assembling a new scene from your own sources. There is no separate editing switch: the moment references are attached, the model moves into it by itself.
This generation is strongest right here: edits respect boundaries — what you asked to change changes, the rest of the frame stays put. That saves iterations: instead of regenerating and hoping, you walk the frame to the target with a series of small precise edits, each step visible in the generation history.
Questions fréquentes
How is ChatGPT Images 2.5 different from GPT Image 2 Pro on the site?
It is the next generation of the same family: sharper details, more natural light, more precise reference-based edits. It also adds two upper rendering levels — ultra high and max — that Pro does not have. For quick and mid-tier pictures Pro is enough; for final quality and complex edits pick 2.5.
What do the ultra high and max levels give me?
They are two levels above the familiar high: the model spends more computation per frame, so fine texture, materials and edges come out cleaner. They cost more and take longer, so they are the choice for finals — print, large format, close crops — while drafts stay at the low level.
How much does a generation cost and what drives the price?
The price depends on the rendering level, the canvas size and the number of variants: the low level costs pennies, the upper ones noticeably more, and four pictures cost four times one. The exact amount is shown on the page before the run and never changes after confirmation; if the generation fails, tokens are refunded automatically.