Sharp Stock

English

简体中文

English

简体中文

AI Smart Matting ​

AI Smart Matting is a background removal tool designed specifically for game artists. Using local BiRefNet Lite, BiRefNet, or their Mac-only CoreML editions, you can remove complex image backgrounds in one click and generate high-quality transparent PNG assets.

Platform Availability & Download Speed

  • Cross-Platform Support: Since this feature uses fully offline local inference, it does not rely on any third-party cloud APIs. Therefore, it is enabled by default across Microsoft Store, App Store, and Steam builds.
  • First-Time Download: Before first use, open [Settings] → [AI Model Configuration] and click Download Now. The download endpoint is built in, and the model can be used offline after the download finishes.

1. How It Works & Optimization ​

AI Smart Matting runs securely on your local computer using the following mechanisms:

  • Local Neural Network Inference: Windows offers BiRefNet Lite and BiRefNet with DirectML GPU acceleration enabled by default. Mac can run those models on the CPU and also offers the native-optimized BiRefNet Lite CoreML and BiRefNet CoreML. All processing stays on the device and never uploads source images.
  • Independent Asset Generation: AI Smart Matting is a fully independent image tool. Once processed, a new transparent PNG image with a -matting suffix is created in the same folder, leaving your original asset intact.

2. Operating Guide ​

You can trigger AI Smart Matting via two separate entry points:

Option 1: Right-Click Context Menu ​

  1. Select Assets: In the asset grid, select one or more images to process (PNG, JPG, JPEG, WebP, and BMP), then right-click the selection. The command appears only when every selected asset supports AI matting.
  2. Trigger Feature: Choose “AI Smart Matting (model name)” from the context menu. Sharp Stock first checks whether each {filename}-matting.png output already exists. If conflicts are found, it lists the target files and asks whether to overwrite them.
  3. Prepare the Model: If the model has not been downloaded, the app prompts you to open Settings. In AI Model Configuration, choose the model and download source, then click Download Now. Sharp Stock displays the download, extraction, verification, and installation stages. Wait until progress reaches 100% and the status changes to Downloaded.
  4. Run Matting Again: After the model download finishes, choose AI Smart Matting again. A full-screen progress display shows the current filename, processed count, and overall progress until all images finish.
  5. View Results: Once complete, the progress display disappears. Each successful source image gets a transparent {filename}-matting.png asset in its original folder.

Option 2: Detail Panel Edit Section ​

  1. View Details: Select the target image asset to display the “Detail Panel” on the right.
  2. Click Matting: Under the “Image Edit” section, click the “AI Smart Matting” button (with the AI magic wand icon).
  3. Complete: The button tooltip, processing progress, and completion notification show the model used for this operation. Wait a moment for the transparent asset to be created.

3. Notes & FAQ ​

  • Initial Model Download: The compressed download sizes are approximately 78 MiB for BiRefNet Lite, 78 MiB for BiRefNet Lite CoreML, 388 MiB for BiRefNet, and 388 MiB for BiRefNet CoreML. Sharp Stock extracts and validates the model locally after downloading, so its installed size is larger than the displayed download size. A model archive or internal file that fails validation is not installed.
  • Free Space During Installation: Sharp Stock temporarily keeps both the archive and extracted model while installing. Make sure the disk has enough free space before downloading. If it does not, the task stops and displays an error.
  • Switching Models: Each model must be downloaded separately. You can switch to any downloaded model without restarting Sharp Stock.
  • Mac CoreML Editions: The three CoreML options are shown only on Mac and require macOS 14 or later. Larger models usually produce better results, but take longer to run and use more resources. On Mac, CoreML models are recommended because they usually perform better. Sharp Stock downloads precompiled models directly and does not compile them on the device. If loading or inference fails, Sharp Stock reports the error and does not switch to the corresponding ONNX model or another model.
  • Sandbox Compliance: On Mac App Store sandbox builds, the model is cached inside the app's designated container, respecting macOS sandbox permissions.
  • Batch Processing: Select multiple images to start one batch. Sharp Stock processes them sequentially, and each image runs as an independent background task so model memory is released before the next image starts. Single-image matting uses the same workflow, including output conflict checks and progress reporting.
  • Lite / Demo asset limit: Before inference starts, Sharp Stock checks whether the output adds a new asset. Replacing an already registered output uses no additional slot, while keeping both with an automatically renamed output adds one asset. If the final count would exceed 100, the app asks you to upgrade before running the model and does not leave an unregistered output file.
  • Fine-Tuning Edges: For extremely complex semi-transparent edges (like very fine hair or smoke), you can use AI Smart Matting to remove the bulk background first, then use the built-in "Auto Cutout / Slice" color dropper tool for micro-adjustments.