SMF·PIXELCUT
Can a prompt replace Pixelcut?
AI image generation — AI image and video generation
Exhibit tracking slip
Verdict
The pieces are all open source and all good: rembg cuts the product out, a compositor drops it onto a background, and a resize pass produces every size Amazon, Etsy and Shopify ask for. Assembling that into a batch pipeline with a preview is genuinely a sitting. What you lose is the phone. Pixelcut is used standing over a table with the product in one hand, and a desktop batch tool is a different job even when the output is identical.
Exhibit A — The prompt
Received on31.07.2026Build a batch product-photo pipeline: cutout, composite, resize. No image generation.
Input: a folder of photos, or drag and drop. Process the whole set in one run with a visible queue.
Cutout: run a local segmentation model (rembg with the u2net or isnet weights, downloaded on first use) to produce an alpha matte. Then refine it: erode the matte by a pixel to kill the halo that segmentation models leave on light backgrounds, and feather the edge slightly. Expose both as sliders, because the right values differ between a matte fabric and a glass bottle, and a fixed default will look wrong on half a catalogue.
Composite: place the cutout on a background chosen per preset — a solid colour, a gradient, or an image the user supplies. Scale the subject so it occupies a configurable percentage of the frame and centre it against its own bounding box, not the original image's, so items photographed off-centre still line up with each other. Add an optional contact shadow generated from the alpha matte, offset and blurred.
Presets: a preset holds a background, a target aspect ratio, an output size, a subject-fill percentage and a margin. Ship presets for square marketplace listings, a tall social format, and a wide banner, and let the user define their own. One run applies every enabled preset to every input.
Review: a contact sheet of results with the source thumbnail alongside, so a bad cutout is obvious at a glance. Clicking one opens a full-size comparison and lets that single image be re-run with adjusted matte settings without redoing the batch.
Output: write to a folder per preset, keep the original filename, and write a sidecar JSON per image recording the model, matte settings and preset used, so a batch can be reproduced.
Out of scope: generating backgrounds with a diffusion model, any hosted API, mobile apps, and retouching tools beyond the matte controls.Opening prefills the prompt — press enter to run it.
Exhibit B — What you lose
- B.1 the phone app, which is where product photos are actually taken
- B.2 generated backgrounds — this composites onto colours and images you supply
- B.3 the shared team template library
- B.4 one-tap results on a photo you just took, with no folder involved
Prior art
Exhibit C — Why people still pay: frontier models, compute, and data
Because ten dollars a month against standing at a desk building a pipeline is not a close call for someone whose job is selling things, not processing images.
Questions
Can I bring anything over from Pixelcut?
Only the images themselves. Pixelcut's templates and saved backgrounds are not exportable, so presets get rebuilt here, which takes a few minutes each. Nothing else is worth migrating.
How good is the cutout compared to Pixelcut's?
Close on solid subjects with clear edges, and worse on hair, fur, mesh and anything transparent. The erode-and-feather controls in the prompt exist specifically because the default output has a visible halo on light backgrounds, which is the difference you will notice first.
What does it cost to run?
Nothing after the model downloads. Everything runs locally on CPU at a few seconds per image, faster with a GPU. For a catalogue of a few hundred products this is a coffee break, once.
What is the one thing that does not survive the rebuild?
Doing it on the phone. Pixelcut's real workflow is photograph, tap, post, without the image ever reaching a computer. A folder-based desktop pipeline is a different habit, and for a lot of sellers it is the wrong one.
Related tools
Receipt