AI image extender: add canvas to a photo without cropping or reshooting

Outpainting is a narrow, specific job: grow the frame. You give it a photo and tell it how many pixels to add on which sides, and it fills only that new border, leaving the original photo untouched inside it. That is different from most of VdoBloom's other image tools, which change what is already inside the frame — FLUX Erase removes something from inside the photo, an instruction-edit model like Flux Kontext restyles or replaces things inside it. This one never edits the original pixels at all; it only builds new ones around them.

Because the model gets no prompt, it has no way to know what you want the new area to become beyond what it can infer from the photo's own edges, lighting and perspective. It extrapolates a wall, a sky, a floor or a background convincingly when the source photo gives it enough visual information to continue — a plain sky, a flat wall, a simple studio backdrop. It struggles the same way any outpainting model does when the edge it is extending from is busy, cluttered or cut mid-object: a person's arm at the frame edge, a sign with text running off the photo, or a tiled floor at an angle. In those cases the model has to guess, and guesses on complex detail are where seams and artefacts show up.

It also will not fix a badly composed photo by itself. If the subject is off-centre because the shot was cropped wrong, outpainting adds space beside it — it does not recompose or re-crop the existing content to recentre the subject. What it is good for is exactly what its name says: making a frame that is too tight into one that is not, without reshooting.

FLUX Outpainting is Black Forest Labs' dedicated outpainting model, and VdoBloom calls it through the Runware API rather than hosting it directly. The request VdoBloom sends is deliberately small: the source image, plus four integers — how many pixels to add on the top, right, bottom and left. There is no prompt, no mask, no style reference. The model's only job is to continue the photo believably into the new space you have defined.

That narrowness is also the strength. Because the interface is four sliders instead of a text box, there is no prompt-writing skill involved and nothing to get wrong beyond picking sensible pixel amounts for your source photo. The trade-off is that you have very little steering power over what appears in the new area — you cannot ask for "add a palm tree on the left" or "make the sky sunset orange." The model decides based on the image alone.

The backend re-checks every value you send: each side is clamped into the 0-512px range even if a request tried to send something outside it, and the combined output size on each axis is clamped to 2,048px. That means the worst outcome from an aggressive setting is a smaller expansion than you asked for, not a failed or malformed job.

The recurring reason people reach for an image extender is a mismatch between the photo they have and the aspect ratio they need. A portrait-orientation product photo needs to fill a widescreen banner; a tightly cropped headshot needs breathing room above the subject for a poster layout; a square Instagram photo needs to become a 16:9 YouTube thumbnail without cutting anything out of the original shot.

It is also useful for print and layout work where a design needs bleed — extra canvas around the edges that will be trimmed or covered by text and logos — and for simply undoing an overly tight crop from the original camera shot, where the alternative would be reshooting.

Every number below is enforced in the same request path that runs the job, not just shown as a suggestion in the UI — the backend clamps or rejects values outside these ranges before the model ever runs.

Start with smaller values than the maximum. A 64-128px expansion on a side with a simple background (sky, plain wall, solid-colour backdrop) almost always blends cleanly; pushing straight to 512px on a busy or detailed edge gives the model more space to fill with less to go on, and that is where visible seams show up. If you need a large expansion on a complex edge, it is often cleaner to run it once at a moderate amount, then run outpainting again on the result for a second pass.

Match your slider choice to what is actually at that edge of the photo. Extending a side that ends in open sky or an empty floor gives the model an easy, low-risk continuation. Extending a side that cuts through a person, a sign, or patterned detail (tiles, brickwork, text) gives it a harder problem, because it has to both continue the pattern and invent plausible detail at the same time.

Because the price is flat at 12 credits regardless of how much canvas you add, there is no cost reason to under-ask — if you are not sure how much space you will need for the final layout, it is fine to expand generously in one pass rather than running the job twice at smaller amounts.