I needed a wide banner and only had a tall photo, so I tried outpainting

The campaign page needed a landscape banner. What I had was a portrait product shot. A reshoot would not land in time, so outpainting was the first thing I tried.

It works, with a boundary worth knowing before you spend an afternoon finding it yourself.

What outpainting actually does

Outpainting generates new picture content outside the original frame. The original pixels stay; the new region is inferred from the edges.

So success comes down to one question. Can what belongs in the new region be inferred from what is already there?

TangyuanAI lists image extension among its capabilities (product overview, checked 6 August 2026).

When it comes out well

Original	Result	Why
Plain or gradient background	Good	The new region only continues a colour
Simple repeating texture, wood or fabric	Fairly good	The pattern has a rule to follow
Subject centred with margin around it	Good	Nothing about the product needs inventing
Subject cropped by the frame	Poor	The model has to invent the part that was never photographed
Busy scene with several objects	Poor	Odd objects grow in the new region

The critical one is the fourth. If the product edge runs off the frame, the strip that gets generated is invented and will not match the physical item.

How far you can push it

In my testing, extending one side by up to about thirty percent of the original stays stable. Past that the generated share grows large enough to shift the visual centre, and the amount of invented content rises with it.

Portrait to landscape usually needs more than fifty percent on one side. For that I extend twice: run one pass, look at it, then extend again from the result. Two passes succeed noticeably more often than one large jump.

What to inspect afterwards

Start at the seam. Where the original meets the generated region, lighting and grain have to carry across. Where they do not, there is a faint visible line.

Then look for objects that grew in the new region. A blurred shape appearing in an empty background is the most common defect.

Finally check the visual centre. After extending, the subject may no longer sit where the composition wants it, which means one more crop.

When to skip it

If the product is cropped at the frame edge, reshoot or use a different photo.

If the job needs an exact ratio with no generated content at all, padding is safer than outpainting.

Generate at the target ratio when you can

The product image tool supports 1:1, 4:5, 3:4, 16:9 and 9:16 (product image page, same date). When the final ratio is known, generating at it directly is more reliable than generating and then extending.

Outpainting is a recovery move, not a first choice.

Sources

Capability and frame sizes from the TangyuanAI product overview and product image page, https://tangyuanai.vip/en/about and https://tangyuanai.vip/en/main-image , checked 6 August 2026. The ratio guidance comes from my own testing. Output depends on the uploaded material; nothing here promises an extension quality.