GPT Image 2.5 Size and Aspect: 1:1, 2K, and 4K

GPT Image 2.5 size and aspect are controls on Create, not a tail on the prompt. Match the ratio to the slot first: 1:1 for avatars, 16:9 for landscape covers, 9:16 for vertical feeds. Explore on Flare at 1K, then raise to 2K or 4K after the composition holds. Sunburst starts at 2K on this site, which fits a finishing pass, not every trial.

Set the frame before 1K, 2K, or 4K

Create splits the decision in two:

Control	What GPT Image 2.5 offers here	Default
Aspect	auto, 1:1, 3:2, 2:3, 4:3, 3:4, 16:9, 9:16, 21:9, 9:21, 5:4, 4:5, 2:1, 1:2	auto
Resolution	1K, 2K, 4K	Flare 1K, Sunburst 2K

auto lets the model pick the ratio. If the channel already crops to a fixed shape, choose that shape so the subject is not cut off later. Resolution is an output tier: 1K checks composition, 2K and 4K check edges and materials. Leave 1024, 2048, and 4K out of the prompt. How to order the sentence itself is in the four-layer prompt.

Where 1:1, 16:9, and 9:16 belong

Pick the ratio from the slot, not from what looks “more premium.”

• 1:1 — avatars, square product tiles, stickers that must survive a center crop. Keep the subject centered with a little margin.
• 16:9 — blog heroes, landscape social, presentation openers. Shift the subject left or right and leave a title zone.
• 9:16 — phone feeds and short-video covers. Place the face or product in the upper half so captions can sit below.
• 4:5 or 3:4 — slightly tall posters and product detail. They crop less off the sides than 9:16.
• 21:9 — wide banners. Do not park the product on the far left or right, or the thumbnail becomes empty background.

When you only change the frame, keep the subject, light, and identity lines still. One variable per pass tells you whether the ratio helped. The lock list for later passes is in multi-turn edits.

When to leave auto

auto is fine for the first passes, while you are still hunting a composition. Switch to a fixed ratio when:

1. The next step crops to a known shape (round avatar, 16:9 hero, 9:16 feed).
2. The new still has to line up with the previous one as a series.
3. The reference is already square or tall and the delivery should keep that frame.

Official image-generation notes set hard limits on custom sizes: both edges are multiples of 16, the long-to-short ratio stays within 3:1, neither edge exceeds 3840 pixels, and the total pixel count stays between 655,360 and 8,294,400. Resolutions above 2560×1440 are marked experimental. Documented examples include 1024×1024, 2048×2048 (2K square), 2048×1152 (2K landscape), 3840×2160 and 2160×3840 (4K landscape and portrait), plus auto as the default. Source: OpenAI image generation guide. image25.xin is an independent create portal. The studio maps those choices to the aspect list and the 1K / 2K / 4K tiers, so you do not type raw pixels.

Explore on Flare at 1K, then raise resolution

A practical order:

1. Flare + 1K + the target aspect — check crop, margin, and whether the text zone still exists.
2. When the frame is right, move to 2K. Sunburst already defaults to 2K, which is the right moment to clean edges and materials.
3. Turn on 4K only for the still you will ship. Exploring entirely at 4K spends the loop on a composition that is not settled.

How Flare and Sunburst split the work is in Flare vs Sunburst. You can change aspect or resolution without also changing the model. Changing both at once hides which one moved the result. Once the resolution is set, the next decision is the quality tier, covered in GPT Image 2.5 quality tiers.

Pick one real slot. Set Create to 1:1, 16:9, or 9:16, make two Flare stills at 1K, then raise the keeper to 2K. If it drops into the layout without cutting the subject, the size choice is working.