Skip to content

v1.0.3

Choose a tag to compare

View all tags
@github-actions github-actions released this 14 Jun 06:39
· 27 commits to main since this release
v1.0.3
26dbd56

Full Changelog: v1.0.2...v1.0.3

Added

  • Automatic GPU selection. DirectML now targets the highest-VRAM adapter instead of blindly using DXGI adapter 0, so machines with switchable graphics (laptop iGPU + discrete GPU, or a desktop with the iGPU enabled) use the right card automatically. Override with STICKER_DML_DEVICE if needed.
  • CPU fallback for heavy models. When a model runs out of GPU memory — during inference or while loading — Sticker first frees other warm GPU sessions and retries on the GPU, then offers to retry on the CPU (same result, slower) before giving up. STICKER_FORCE_CPU=1 always skips the GPU.
  • STICKER_BIREFNET_SIZE to lower BiRefNet's input resolution (e.g. 768, 512) and cut its memory footprint.
  • Model download progress. The first-use download now streams with a live progress bar (MB / percent) instead of an opaque wait.
  • On-sticker busy spinner shown while (re-)matting — on GPU and during the slower CPU pass — so a long run never looks like a frozen app.
  • "Clear matte cache…" tray item that frees cached cutouts without touching your session, clipboard captures, or downloaded models.
  • Cache-aware model picker. The right-click "Matte with" entries reuse a previously computed result instantly; "Re-process current (ignore cache)" forces a fresh pass.

Changed

  • Context menu now covers all supported image types (.jpg .jpeg .png .webp .bmp .gif). Previously it registered under the image PerceivedType, which .webp (and others) often lack — so it only appeared for .jpg. It's now registered per extension.
  • README: rewritten GPU / troubleshooting guidance, a full environment-variable reference, and clearer context-menu and caching docs.

Fixed

  • Atomic model downloads. Downloads land via a temp file and rename, so an interrupted download no longer leaves a corrupt .onnx that crashes on every launch.
  • Failed matte no longer poisons later ones. A model that errors mid-run is now disposed and evicted, freeing its reserved VRAM so subsequent re-mattes (even of other models) aren't starved.
  • Oversized images are rejected with a clear message instead of overflowing the pixel-buffer allocation and crashing.
  • Concurrent re-mattes of the same image+model are serialized — no more duplicate inference or racing writes to the same cache file.
  • Clearer clipboard errors. "Paste as sticker" distinguishes a busy/locked clipboard from a genuinely empty one.
  • Session restore is resilient to a malformed entry — it skips the bad one and restores the rest instead of dropping the whole session.
Assets 3
Loading