2026-09-06

Cutting out a food photo — the menu board, the poster, and the one place you should not

Two very different jobs get filed under "fix the food photo." One is get a clean, honest picture of the dish — which is what every delivery platform's photo policy is actually about. The other is lift the dish off its table so it can sit on a poster, a menu board, or a promo card — a transparent-PNG cutout. They are not the same job, the results do not go to the same places, and doing the second one to a photo headed for the first is a documented way to get that photo rejected.

So this is a walk through where a food cutout genuinely earns its place in a small restaurant's or cafe's graphics, where it actively hurts, and what NearBG's on-device pass really manages with a plate of food — including the specific things it loses and cannot get back.

Why photos are the lever in the first place

The National Restaurant Association's 2026 State of the Restaurant Industry, published February 12, 2026, forecasts $1.55 trillion in total restaurant and foodservice sales and industry employment reaching 15.8 million — a large, crowded market in which your listing competes inside a scrolling grid of thumbnails. DoorDash's own merchant learning center says menus with item photos "saw an increase of up to 44% in monthly sales," and that menus with header images and logos get "up to 50% (header) and 23% (logo) more monthly sales than menus without these branding elements." The same page cites DoorDash's Delivery Trends Report finding that nearly half of Gen Z diners (46%) say food photos influence their decision to try a new restaurant.

Take those numbers for what they are: a platform's own marketing material, measured on its own merchants, with "up to" doing real work in every sentence. They are still the most concrete public figures on the question, and the direction is not seriously in dispute — a menu item with a picture outperforms one without. But what follows from that is "photograph your dishes," not "cut out your dishes." Those two diverge immediately.

The one place you should not send a cutout: the platform tile

Delivery-app item photos are fixed-aspect rectangles that the app fills with its own layout. Transparency buys you nothing there — and the platforms' written rules run against edited edges.

Read that last row twice if you sell in Korea. "Neutral background" (DoorDash) is a shooting instruction — put the plate on a plain surface — not an editing one, and Baemin names a badly cut food edge as grounds for rejecting the image outright. An automatic cutout pasted back onto a flat white rectangle is precisely the artifact those rules describe. The platform tile wants the original photograph.

Where the cutout actually pays for itself

Everywhere you control the layout yourself. On those surfaces the food sits on your background — a brand colour, a wood texture, a chalkboard, a photo of your own storefront — and a rectangular photo with a visible white box around it reads as pasted on. A cutout doesn't.

• Printed and laminated in-store menus — a few hero dishes floating in the margin beside their prices.
• A-boards, window posters, and table tents — one dish, large, on a solid colour.
• Digital menu boards and self-order kiosk promo screens, where a "today only" panel is usually a coloured background plus a dish plus a price.
• Instagram, KakaoTalk channel, and LINE announcements — a new-item graphic rather than a plain photo post.
• Takeout stickers, bag labels, coupons, and event flyers.
• Your own website's hero section, where the dish overlaps a headline.

Notice the pattern: none of these are platform submissions. The practical habit is to keep two versions of every hero shot — the untouched rectangle for the delivery apps, the PNG cutout for everything you design yourself.

What NearBG does with a plate of food, exactly

NearBG runs a segmentation model called U2Netp inside your browser tab. The image is never uploaded — the model file and the WebAssembly runtime come down to your device and the inference happens locally, which for a business photo you have not published yet is a meaningful difference from a cloud API. You drop in one image (or paste one with Ctrl + V), wait a few seconds — inference measured around three seconds single-threaded in our own model spike, faster on a browser with SIMD and threads — and get a PNG at the original photo's resolution with the background's alpha punched out. There is also an erase-only touch-up brush on the result canvas for stray leftovers, plus a reset that restores the automatic result.

Now the part that matters for food. U2Netp is a salient-object model. It was never trained to recognise food, or plates, or anything in particular — it predicts which pixels belong to "the thing this photograph is of." Three consequences follow, and all three show up on the first burger you try.

1. You get the plate too, because the plate is part of the subject

Shoot a bowl of ramen on a table and the salient object is the bowl with the ramen in it, not the noodles alone. Usually that is exactly what you want on a menu board — a floating bowl reads as food, floating noodles read as an accident. But if you genuinely want only the burger without the slate it is served on, that is manual brush work, and the brush only erases: there is no paint-back. The reset button returns you to the automatic result, so a bad erase costs you the session's touch-ups, not the photo.

2. Thin things disappear, and this is structural

The model's input is a fixed 320×320. Your 4000-pixel-wide photo is scaled down to fit that box, the mask is predicted at that size, and the mask is then scaled back up to the original resolution. Anything narrower than roughly one pixel at 320×320 has no reliable representation in the mask at all. In food photography that list is long and annoyingly specific:

• noodle strands lifted on chopsticks — and the chopsticks themselves
• skewer sticks, cocktail picks, and straws
• a drizzle of sauce or a thread of caramel
• herb sprigs, microgreens, chive batons, a single strand of shredded chilli
• steam, and the wire handle of a fry basket
• fries or breadsticks poking out past the rim of a cup

They don't come out slightly rough. They come out gone, or reduced to a soft smear. This is the same limitation that makes flyaway hair hard, and the touch-up brush does not fix it — the brush removes pixels, it cannot restore ones the mask never kept. If your hero shot depends on a skewer, plan to shoot a version of the dish that doesn't.

3. Glass and ice are not solvable this way

The mask sets each pixel's opacity for the subject. It has no way to express "this glass is here, and you can see the table through it." An iced americano, a beer, a soda, a jelly drink, a glass teapot — you will get either a solid, opaque-looking blob where the transparency was, or a half-eaten edge where the model couldn't decide. For drinks in clear glass the honest advice is to skip the cutout and design with the rectangular photo, or to shoot the cup against the colour you intend to place behind it.

Shooting so that the cutout is easy

Almost everything that makes the automatic pass clean is decided before you open any tool.

• Contrast the plate against the surface. A white plate on white marble is the hardest case; a white plate on dark wood is the easiest. This is the single biggest factor.
• One dish, well inside the frame. A subject that runs off the edge gets cut flat at the border. Uber Eats already asks for centred, single-item framing, so this costs you nothing you weren't already doing.
• Kill the props. A napkin, a second glass, a hand, a menu leaning in the background — each is a competing salient object, and the model may keep it, or half-keep it.
• Watch the shadow. A hard cast shadow attached to the plate often reads as part of the subject and comes along with it. Soft, indirect light — which Uber Eats' guidelines recommend anyway — leaves far less to argue with.
• Angle by dish shape. Uber Eats' own guidance is top-down for plates and bowls, 45 degrees for burgers, sandwiches, and other tall items. Top-down on a plain surface also happens to be the cleanest possible input for a cutout.
• Feed it the original file. Not a screenshot, not a copy re-saved by a messenger app. Output resolution equals input resolution, so a compressed 900-pixel copy gives you a 900-pixel cutout — and an A2 poster needs a great deal more than that.

What NearBG deliberately does not do

Worth knowing before you plan a menu refresh around it:

• No background replacement. You get transparency, not a new backdrop. Putting the dish on your brand colour happens in Canva, Miricanvas, Figma, PowerPoint, or whatever you already lay out with.
• No batch. One image per run. A 40-item menu is 40 runs — which is the honest reason this is a tool for your five or six hero dishes, not for re-cutting a whole catalogue.
• No resizing or cropping. It will not produce Baemin's 1280×960 minimum or Uber Eats' 5:4 frame for you — and those specs apply to the tile photo, which shouldn't be a cutout anyway.
• No templates, no cut lines, no print bleed. If the poster gets printed and die-cut, the cut path is your print shop's job or your design tool's.
• No paint-back on the brush. Erase-only, with a reset.
• Keep the download as PNG. Re-saving it as JPEG silently fills every transparent pixel with a white box, which is the exact failure you were trying to avoid.

A short checklist

• Shoot the dish on a plain surface that contrasts with the plate, in soft light, centred, one item, no props.
• Upload that original photo to the delivery apps. No cutout, no text, no overlay — Baemin can reject an awkwardly edited food edge, and DoorDash asks for a neutral background and no overlays.
• Run the same original through a cutout only for graphics you design yourself: board, poster, kiosk screen, sticker, social announcement.
• Check the result at 100% before you use it — specifically the garnish, the steam, the skewer, and anything made of glass.
• Keep both files. The rectangle and the PNG have different jobs, and neither replaces the other.

The honest summary

Sources, named: the $1.55 trillion sales forecast and 15.8 million employment figure come from the National Restaurant Association's February 12, 2026 press release for its 2026 State of the Restaurant Industry report; the 44% item-photo figure, the 50%/23% header-and-logo figures, and the 46% Gen Z figure all come from DoorDash's own merchant learning center and its Delivery Trends Report, which is a platform marketing source rather than independent research; the Uber Eats item-photo specification (5:4–6:4, jpg/png/gif, 10 MB, 440–10,000 px tall, 550–10,000 px wide, centred, single item, no text or watermarks) and the angle and lighting advice come from Uber's own merchant help pages; the "proper lighting, without text or other overlays… neutral background" wording is DoorDash's own help article; and the 1280×960 minimum, the 30 MB ceiling, the "avoid excessive compositing" rule, and the awkwardly-edited-edge rejection reason come from 배민외식업광장's published menu-image standards. Several widely repeated statistics were left out on purpose because no primary source could be found for them — a "50% photo coverage lifts sales 13%" figure, an "88% higher sales" Uber Eats figure, and various "professional photography lifts orders 20–35%" claims all trace back to marketing pages for AI photo tools rather than to any platform's or association's own publication. Everything said about the tool comes from the tool's code: one photo per on-device pass, a fixed 320×320 saliency mask scaled back up over your full-resolution image, an erase-only brush with a reset, no background replacement, no batch, no resizing — and honest failure anywhere a skewer, a steam wisp, or an iced glass meets a mask that was never built to describe them.