How to change aspect ratio without cropping with AI

You shot a product photo as a clean 1:1 square. Now the paid social brief wants 9:16 for Reels, the YouTube thumbnail wants 16:9, and the site hero wants something closer to 21:9. The usual move is to crop, and the usual result is that something important gets cut — the logo, the model's hands, the negative space that made the composition work in the first place.

There is a better answer. Instead of removing pixels to reach a new shape, you generate the pixels that were never in the frame. That is how you change aspect ratio without cropping: treat the original as a reference, tell an image model what belongs in the new space, and let it build outward from what you already have. Used this way, an image model becomes an AI image extender rather than a from-scratch generator.

This guide walks through that workflow on OmniArt — picking the target ratio, feeding the original in, what to write in the prompt, and how to fix the two failure modes that show up most often: visible seams and repeated patterns.

Why cropping loses content, and when reframing wins

Cropping is subtractive. To turn a 1:1 image into 9:16, you keep a narrow vertical slice and throw away roughly 44% of the width; going the other way costs a similar share of the height. The math is unforgiving, and it always takes from the edges, which is exactly where framing information lives: the horizon line, the table edge, the product's shadow, the space a headline was supposed to sit in.

There is a second cost that is easier to miss. Cropping changes the subject's relationship to the frame: a portrait composed with comfortable headroom becomes a tight, slightly claustrophobic shot, and a product sitting one third into the frame ends up dead center. The result is technically the right shape and compositionally a different photo. Reframing avoids both problems because it adds instead of removes — you keep the entire original composition and generate believable surroundings to fill the new ratio.

Common target ratios and where they are used

Before prompting anything, decide the exact ratio you need. Guessing here wastes generations: a 4:5 asset delivered into a 9:16 slot gets auto-cropped by the platform anyway, and you are back where you started.

Ratio options in the OmniArt image workspace vary by model and quality setting, so the list updates when you switch models. Set the ratio first, then write the prompt around it — describing a wide panoramic composition while the ratio is still 1:1 produces a squeezed, unconvincing result.

The reframing workflow, step by step

1. Choose the target ratio and the direction of growth

Write down which dimension is expanding. Going from 1:1 to 16:9 means the image grows left and right; 1:1 to 9:16 means it grows up and down. That single sentence shapes the whole prompt, because you are describing new content in a specific direction, not a general "wider" image.

Also decide where the subject should sit in the new frame. Centered is the safe default, but off-center often reads better: put a product at the right third of a 21:9 banner and the left two thirds become clean space for a headline.

2. Upload the original as a reference image

In the OmniArt image workspace, attach the original file so the model has the actual pixels to work from rather than your description of them. Models that accept a single reference use it as the image-to-image base. Models that accept multiple references — Seedream 5.0 Pro takes a set of them — let you add supporting shots: another angle of the same product, a texture close-up, or a background plate you want the extension to match.

Two references are usually enough: the original, plus one wider shot of the same environment. The second one gives the model factual information about what the room, the table, or the sky actually looks like beyond the original crop, which is the single biggest quality lever in this workflow.

3. Prompt what fills the new space, not what is already there

This is the step most people get wrong. They re-describe the subject, and the model regenerates the subject. Instead, describe the extension: what continues, in which direction, with what lighting and perspective, and what must not change.

A reliable prompt shape has four parts:

1. Preserve — state that the reference subject and composition stay exactly as they are.
2. Extend — name the direction and the ratio you are filling.
3. Fill — describe the new content concretely: surfaces, background elements, depth.
4. Match — specify continuity: light direction, shadow behavior, focal length, grain, color temperature.

Here is that shape applied to a square product shot going ultra-wide:

"Use the reference photo as the exact center-right subject. Extend the scene horizontally to a 21:9 composition. Continue the same pale oak tabletop to the left with two soft out-of-focus ceramic props in the far background. Keep the identical soft window light from the upper left, matching shadow direction and softness, same shallow depth of field and same neutral color temperature. Do not alter the bottle, its label, or its existing shadow."

A portrait going vertical for Reels:

"Keep the subject from the reference image unchanged, positioned in the upper two thirds. Extend the frame vertically to 9:16, adding the continuation of the studio backdrop above the head and the subject's lower torso and hands resting naturally below. Maintain the same 85mm perspective, single softbox from camera left, and consistent film grain across the whole frame. Leave clear, uncluttered backdrop in the lowest quarter for caption text."

An interior shot going 16:9 for a blog cover:

"Treat the reference interior as the right half of a 16:9 frame. Extend to the left with the same room continuing: matching wall color and skirting board, one tall window with the same daylight quality, floorboards continuing in correct perspective toward the existing vanishing point. Keep all existing furniture and reflections exactly as shown."

4. Refine edges and continuity

Compare the output against the original at full size and look specifically at the boundary where old content meets new. If something is off, fix it with a targeted follow-up edit rather than a fresh generation — a re-roll usually costs you a good extension along with the flaw.

Seedream 5.0 Pro is a capable option here because it responds to region and anchor language, letting you name a spot and change only that. GPT Image 2 handles the same job through plain natural-language editing. Either way, keep the correction narrow: "Along the left seam roughly one third from the top, continue the tabletop edge as one straight unbroken line and remove the slight tonal step. Change nothing else."

5. Finish at the delivery resolution

Extension work is easy to judge at a working resolution, but delivery should happen at the size the destination needs. Raise the quality setting in the workspace for a final pass — GPT Image 2 outputs natively at 2K with a 4K option, and Seedream 5.0 Lite generates 4K directly — so a 21:9 hero does not arrive soft on a large display. Generate the final version at the target ratio and resolution together rather than scaling up afterward.

Troubleshooting seams and repeated patterns

Visible seams

A seam is a faint line, tonal step, or texture change where the original ends. It usually means the prompt described the new content but not the continuity between old and new.

• Name the light source and its direction explicitly. Most seams are lighting mismatches, not edge artifacts.
• Add grain, noise, and color temperature to the match clause. A clean synthetic extension beside a slightly grainy photo reads as a seam even when the geometry is perfect.
• If a straight edge breaks across the boundary — a table lip, a horizon, a wall line — name it and demand one unbroken line.

Repeated patterns and cloned objects

When a model has to invent a large area with little information, it often duplicates what it can already see: two identical plants, a tiled floor pattern that repeats too regularly, the same cloud twice.

• Give the extension explicit, different content. "Add two out-of-focus ceramic props" beats "extend the background," because a vague instruction invites copying.
• Supply a second reference showing the real wider environment. Facts beat guesses.
• Ask for less at once. Going 1:1 to 21:9 in one step means inventing more than half the image; 1:1 to 16:9, then 16:9 to 21:9, gives the model a stronger base each time.

Subject drifted or got re-rendered

If the product label changed, a face shifted, or the color is off, the model regenerated rather than extended. Strengthen the preserve clause, confirm the original is attached as a reference rather than merely described, and use region or anchor language to protect the original area where the model supports it.

The extension looks flat

New areas can come out with less depth than the original. Ask for depth explicitly: foreground occlusion, atmospheric falloff, matching depth of field, and shadows following the same direction and softness as those already in frame.

Getting started on OmniArt

Open the OmniArt image workspace, upload the image you need reframed, set the aspect ratio to your delivery target, and write a prompt using the preserve, extend, fill, match shape from step 3. Start with one ratio step and one direction of growth, then iterate on the seam with a narrow follow-up edit.

For deeper prompt control on the refinement pass, the Seedream 5.0 Pro prompt guide covers region-level and anchor-based editing plus multi-image fusion, both of which apply directly to extension work. Every image model sits in the same workspace, so you can reframe with one, refine with another, and deliver every ratio the campaign needs from a single original.