The job most people hire Pixelcut for is a cutout, an eraser and an upscaler, and open models do all three on an ordinary laptop: rembg or BiRefNet for the background, an inpainting model for the eraser, and Real-ESRGAN for upscaling. An agent wraps them in a small web page with batch export in one sitting. What stays behind is the generative side: AI backgrounds, generative fill and the third-party image and video models that the credits pay for, which you would call through an API and pay per image, plus the phone app you shoot and edit in.
Before you buildPixelcut Free already removes backgrounds and exports without a watermark within daily limits. Build only if you want batches, privacy or no daily cap.
Build me a local photo cleanup and product-shot tool to replace Pixelcut. Requirements: - Python 3.12 with FastAPI and one HTML page with a little vanilla JS: `python -m cutout` serves http://localhost:4730 with tabs for Remove background, Erase, Upscale and Batch. - Remove background: `rembg` with the `birefnet-general` model (use `isnet-general-use` under 8 GB of RAM). Preview on a checkerboard, then export a PNG with alpha, or composite on white, a hex colour or an uploaded image with a soft shadow made from the blurred alpha mask (Pillow). - Erase: paint a mask on a canvas, then remove the marked area with `simple-lama-inpainting`, falling back to OpenCV `cv2.inpaint` (Telea) when the model isn't downloaded. - Upscale: run the `realesrgan-ncnn-vulkan` binary at 2x and 4x with a progress bar; without Vulkan, fall back to Pillow Lanczos and say so on screen. - Batch: drop a folder or many files, pick a preset (remove background on white at 2000x2000, square 1080, marketplace 1600 on the long edge), run a worker pool sized to the CPU, and write ./out/<date>-<preset>/ plus a zip. - Optional AI backgrounds: when IMAGE_BASE_URL, IMAGE_MODEL and IMAGE_API_KEY are in .env, a prompt box asks an OpenAI-compatible images endpoint for a background and composites the cutout over it; the product pixels are never regenerated. Warn about per-image cost before the first call. - Never modify originals; save each batch's settings in job.json next to its output. No accounts, no telemetry; the only network calls are model downloads and the optional image endpoint. - Out of scope: generative fill, AI video, a phone app, templates, teams. - README: model sizes, rough seconds per image on CPU and GPU, and how to install the upscaler binary.
Opens your agent with the prompt filled in, you press enter. Or copy it and paste it into an empty folder.
The first prompt builds the core. Run these one at a time in the same folder to grow it into your own Pixelcut.
Extend the photo tool in this folder. Requirements: - `python -m cutout watch ./inbox --preset white-2000` watches a folder with `watchdog` and processes each new image once it has stopped growing for 2 seconds. - Write results to ./out/watch/ with the original name plus the preset, and move the original into ./inbox/done/. - Log failures to ./out/watch/errors.log and keep going. - Update the README with a launchd and a systemd example.
Add custom export presets to the photo tool in this folder. Requirements: - presets.json holds name, width, height, fit (contain or cover), padding percent, background and format (png, jpg, webp) with a quality setting. - A preset editor in the page, and a `--preset all` batch option that writes one folder per preset. - Refuse presets with a side under 100 px or over 8000 px. - Update the README.
Add a comparison view to the photo tool in this folder. Requirements: - After each operation show a draggable before and after slider, with the result on a checkerboard when it has transparency. - Keyboard shortcuts: left and right arrows move between images in a batch, `a` accepts, `r` re-runs with the other model. - Mark rejected images in job.json so a re-run only touches those. - Update the README.
Small sellers and creators pay for speed on a phone: shoot a product, drop the background and get a clean listing image in seconds, with AI backdrops and a commercial licence included. The credit pool also saves juggling accounts with several image-model providers.
Estimate · Everything runs on your own machine for free; the optional AI backgrounds cost a few cents per image at an API provider. The build is token usage on your agent plan.
Checked 2026-10-19 against Pixelcut's docs: 1, 2, 3, 4, 5, 6, 7
| Plan | Monthly | Annual, per month | What you get |
|---|---|---|---|
| Free | $0 | $0 | Limited background removal, magic eraser, fast upscaler, generative fill and expand image; export without watermark. |
| Pro | $10 | $8 | Unlimited background removal and upscale, 600 AI credits a month for advanced models, generative fill, expand and AI backgrounds up to about 300 a day, 1,000 batch exports a month, 3-person team, commercial license. |
| Business | $30 | $24 | Everything in Pro with 3,600 credits a month, generative tools up to about 1,200 a day, 2,000 batch exports a month and a 10-person team. |
Free gives limited background removal, magic eraser and upscaling, and exports without a watermark.
Yes. An AI coding agent such as Claude Code, Codex or Cursor can build a usable personal version in one session with the prompt on this page, running on your own machine.
About $10 a month on the Pro plan, $120 a year (checked 2026-10-19).
Generative fill, expand image and AI backgrounds on a daily allowance; One credit balance that reaches many third-party image and video models; The iOS and Android apps for shooting and editing on the go; Team seats, shared work and monthly batch quotas; A commercial-use licence that covers the output. If any of those matter to you, keep paying.
Yes: rembg (open source), BiRefNet (open source), Upscayl (open source).