NearBG

2026-09-04

Halloween cutouts from your own photos — what the on-device pass does, and what it loses

On September 1, 2026 — three days before this post — the National Retail Federation published a piece on its own site about why Halloween keeps arriving earlier, and put two numbers in it: "one-quarter of consumers saying they had already purchased Halloween items such as decorations, costumes, candy or greeting cards as of early August," and, looking back a year, "nearly half (49%) of consumers began shopping for Halloween before October, a number that had risen from just 34% 10 years prior." NRF's full 2026 Halloween survey doesn't land until late September, so the most recent complete figures are still last year's: its September 18, 2025 press release, built on a survey of 8,045 consumers fielded September 2–9 with a margin of error of ±1.1 points, put 2025 spending at a record $13.1 billion, $114.45 per person, with 73% of consumers planning to take part. The category split in that release is the interesting part for anyone who makes things: costumes at $4.3 billion (71% buying), decorations at $4.2 billion (78% buying) — and NRF's own October 1, 2025 follow-up noted plainly that "more people are buying decorations than costumes," with 45% of decorators putting things out in the first half of October and nearly half of the early shoppers starting in September. Greeting cards, the category that quietly covers party invitations, drew 38% of shoppers and $700 million.

Which is to say: if you are making any of it yourself, early September is when the making window actually opens, not when it starts feeling early. This post is about one small step that sits underneath a surprising share of homemade Halloween graphics — getting a person (or a pet, or a jack-o'-lantern) out of a photo and onto a transparent background — and about exactly where NearBG's part of that job begins and ends.

Where a cutout actually shows up in a homemade Halloween

Halloween is an unusually shape-driven holiday. Almost nothing in its visual language is a rectangle: it is silhouettes in windows, figures on a door, faces on a banner, stickers cut along an outline. That is the whole reason the transparency step keeps reappearing:

Every one of those is a layout job done somewhere else, in a template tool or a cutting-machine app or a print shop's uploader. In all of them, the design already owns the background. So the photo cannot arrive as a photo.

What the NearBG pass actually is

One photo per pass. Drop it in, paste it, or share it in from your phone's share sheet; the segmentation model runs inside your browser tab, so nothing is uploaded, there is no account, and no server ever holds a copy of the picture. On a desktop browser that pass takes a few seconds — the honest caveat, straight from the product notes, is that the measurements behind that number were all taken on desktop hardware, and low-end mobile has never been re-verified. What comes back is a PNG at the original photo's full resolution, plus an erase-only touch-up brush with a size slider and a reset button that reverts to the untouched automatic result if a stroke goes wrong. Then you download or share it. That is the entire feature set for this job.

Halloween is genuinely the hard season for this model

Two things about the engine matter here, and both are in the code rather than the marketing.

First: the model underneath is a salient-object detector, not a person detector. It is asking "what stands out in this frame," not "where is the human." That cuts in your favor more often than people expect — a full-body monster costume, a masked face, a carved pumpkin on a porch rail, a skeleton propped in a chair are all perfectly reasonable subjects, because the model was never looking for a face in the first place. It also sets up the season's specific failure mode: salient means "stands out from its surroundings," and a black cloak photographed in a dim room against a dark wall is the definition of not standing out. Halloween photos are disproportionately shot at night, under orange light, in front of dark decor. If a cutout is what you need, take one deliberate photo in daylight against a plain wall that contrasts with the costume, and keep the atmospheric night shots for the things that don't need cutting out.

Second: there is a resolution bottleneck, and it explains almost every "why did my broom disappear" question at once. The model's input is fixed at 320×320 pixels. Your photo is scaled down to fit inside that square, the mask is computed at that size, and the mask is then scaled back up and used as the alpha channel on your full-resolution image.

1. Your photo 3000 × 4000 px broom handle ≈ 12 px wide 2. Fit into 320 × 320 aspect kept, padding on the right handle now ≈ 1 px 3. Saliency mask computed at 320 × 320 nothing left to keep 4. Scaled back up alpha applied at 3000 × 4000 — minus the broom The 320 × 320 input is fixed by the model NearBG runs (U2Netp) — it is not a quality setting you can raise. Rule of thumb: anything thinner than roughly ten pixels in your original lands on about one mask pixel.
Why thin things vanish: the decision about every pixel is made on a 320×320 version of your photo, then stretched back up.

Do that arithmetic on a typical phone photo and the consequences fall out immediately. A 3000×4000 image is scaled by 0.08 to fit the square, so a feature twelve pixels wide in the original is about one pixel wide in the mask. That is the budget you are working inside, and Halloween props sit right at its edge: broom bristles, a wand, a scythe shaft, the wire ribs of a pair of wings, a spider's legs. Props held out away from the body are doubly exposed, because they are both thin and, to a saliency model, arguably a separate object from the person.

The soft-edge cases are the same ones this blog's hair post and pet post already cover, only worse: a costume wig sheds more stray strands than real hair, and tulle, chiffon capes, cobweb stretch, and fog are semi-transparent by nature. Every one of those is a blend of subject and background in the same pixel, and a mask has to make a hard in-or-out call. Preserving see-through fabric would take matting, which NearBG does not do. Expect a wig or a cape to come back as a solid shape with a somewhat crunchy outline.

And one thing the touch-up brush cannot rescue: it only erases. If the automatic pass left a patch of wall between your arm and your side, the brush clears it. If the pass dropped the wand, the brush cannot paint it back, because it removes pixels rather than recovering detail the mask never caught. Practical version: shoot in good light, against a contrasting wall, and hold thin props against the body rather than out at arm's length.

What NearBG deliberately doesn't do here

The honest summary

Sources, named: the early-August quarter of consumers and the 49%-before-October figure (up from 34% a decade earlier) come from the National Retail Federation's own blog post of September 1, 2026, which does not name the underlying survey for the early-August number; the $13.1 billion total, $114.45 per person, 73% participation, and the costume/decoration/greeting-card splits come from NRF's September 18, 2025 press release for its 2025 Halloween survey of 8,045 consumers fielded September 2–9, 2025 (±1.1 points); "more people are buying decorations than costumes" and the October-decorating timing come from NRF's October 1, 2025 follow-up post; NRF says its 2026 survey publishes in late September. The Print Then Cut maximum of 9.25″ × 6.75″ is Cricut's own published figure. 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 roughness wherever a wig, a cape, or a broom handle meets a hard-edged mask.