2026-09-10

Making a coscard, sticker, or portfolio shot from a cosplay photo — where the cutout fits, and what armor and props do to it

New York Comic Con 2026 runs October 8–11 at the Javits Center, the same show that has already drawn a record 250,000 fans in a single weekend. A number like that tells you something concrete about what a cosplay photo taken there actually looks like: not a clean backdrop, but a hallway packed with strangers, booth signage, and fluorescent show-floor light.

That photo is exactly what a cosplayer needs afterward for three completely ordinary things: a coscard (the cosplay community's own name for a business card that trades contact info for a fandom-appropriate photo instead), a con sticker to sell or trade at the table, and a portfolio shot for a cosplay Instagram or costuming site. None of those three want the hallway. All three want the costume, alone.

The coscard is older than the term "content creator," and the format hasn't changed

Coscards aren't new. Kotaku covered them plainly over a decade ago: cosplayer cards "look more like glossy trading cards, with full color pictures, than business cards," carrying photos, character names, and website info instead of a job title — popular enough that a Japanese printer built a service specifically around the format, at roughly $26 for a hundred cards. A cosplay-focused how-to guide is more specific about the contents: "the cosplay's name, at least one cosplay photo and contact or social media handles," plus a warning worth repeating for anyone laying one out — pay attention to the edges when building the card, because most print sites crop into whatever bleeds past the template, so leave the subject room to breathe.

Nothing in that guidance says the photo has to be a clean cutout instead of a normal rectangular photo. In practice, most cosplayers still want one, for the same reason a product shot or an album cover does: a subject on its own transparent background drops onto a card, sticker sheet, or Instagram highlight template without a second, mismatched rectangle of convention carpet riding along behind it.

Good news: the model doesn't know it's armor, not skin

NearBG's segmentation model, U2Netp, belongs to a family built for salient object detection — finding the one prominent thing in a photo, whatever material it happens to be made of — rather than for recognizing human bodies or faces specifically. A cosplayer in full costume is, to that model, exactly the same kind of input as a product photo or a pet on a rug: one clearly bounded subject against a background it is not part of. Face paint, a full-coverage mask, foam armor, a solid-color bodysuit — none of it confuses the model the way it might confuse a face-detection tool, because nothing here is trying to detect a face in the first place.

The honest hard case: what a convention floor does to the edges

Two separate problems stack on top of each other at a con, and neither is specific to costumes — they are the same segmentation limits this blog has already documented elsewhere, just showing up in armor and props instead of a wedding veil or a broomstick.

Reflective and translucent materials. A mirrored visor, chrome-painted armor, or glossy vinyl catches and reflects whatever is actually behind it — other attendees, booth lights, the carpet — so the pixel at the edge of a helmet is genuinely ambiguous between "subject" and "reflection of the background," the same ambiguity this blog's piece on cutout edges describes for glass. There is no clean automatic fix for it; a matte prop photographs more predictably than a mirrored one.
Thin props at a fixed mask size. NearBG's mask is computed at a fixed 320×320 and then scaled back up to the photo's real size, which this blog's Halloween post walks through for broom handles and wand tips: anything narrower than roughly 1/320th of the photo's long side has no reliable representation in the mask at all. A spear shaft, a bowstring, cat-ear wire, antennae, a katana's blade edge — on a typical 3,000–4,000 px event photo that is anything under about 10–13 px wide, and it comes out faint, broken, or missing rather than merely rough.

Lighting compounds both problems. Convention halls run on high-contrast, often colored show lighting, and even photographers who do this for a living say the floor itself works against a clean shot. One recurring complaint in a photography how-to about cosplay convention shoots put it bluntly: "trying to find a spot away from the crowd on the con floor is a nightmare." The article's own fix is to physically move the cosplayer against a corner or a plain patch of wall before the shutter clicks, precisely so there is less clutter for anything — a person editing by hand, or a model running in a browser — to cut around afterward. A background remover helps after the fact; it does not replace choosing a better wall.

One photo in, one transparent PNG out. Which coscard printer, sticker die-cut, or portfolio grid the cutout lands in afterward is a different tool's job entirely.

Where NearBG fits, stated narrowly

On-device, in any browser. The photo is processed inside the tab and never uploaded — no account, no server copy. That happens to line up with a norm a lot of cosplayers already keep on purpose: a coscard trades a costumed photo and a handle for contact info, not a legal name, and plenty of cosplayers separate their persona from their identity deliberately. A tool that never sends the photo anywhere fits that habit better than one that asks you to sign up first.
The touch-up brush only subtracts. It erases alpha under the cursor, with a size slider and an all-or-nothing reset back to the untouched automatic result. It can clean up a patch of carpet the model kept along a cape's hem or a prop's base. It cannot repaint a mirrored visor's reflection into something else, and it cannot restore a spear tip the mask never captured — the brush only takes pixels away, it never adds detail back.
One photo at a time, no batch. A coscard set for two characters, or a sticker sheet with three poses, means running the tool three separate times. There is no bulk mode.
Original-resolution PNG out, nothing else. The only output option is a transparent PNG at the source photo's own dimensions — no built-in resize to a card printer's spec, no crop to a sticker die line, no template. Input is PNG, JPEG, or WebP, up to 500 MB on desktop or 50 MB on a phone, which matters here because event photography files, especially from a real camera rather than a phone, run large.

Honest summary

The claims about the convention itself are sourced and dated: NYCC 2026's October 8–11 dates at the Javits Center come from the show's own site, and the 250,000-attendee record comes from Publishers Weekly's coverage of the 2018 event. The claims about coscards come from Kotaku's decade-old but still-accurate description of the format and from a current cosplay how-to guide's list of what actually goes on one. The convention-floor lighting complaint is a direct quote from a working photographer's how-to piece, not a paraphrase. And everything said about NearBG itself — the salient-object model, the fixed 320×320 mask and what it does to thin props, the erase-only touch-up brush, the lack of batch processing or resizing — comes from the tool's own shipped behavior, most of it already documented in more technical detail in this blog's edge-halo and Halloween posts.