[WirePortLabs]

Summary

Replaced Depth-Anything-V2 and SAM 2 with Depth-Anything-V3 Small and SAM 3.1. Added Rust/ONNX Runtime inference for SAM 3 image, text, and geometry/mask models, using text prompts for foreground and sky masks and box prompts for subject selection. SAM 3 image features are cached so different prompts can reuse the encoder result.

Depth estimation now uses the DA3 Small ONNX model, preserves image aspect ratio, and inverts depth values to maintain the editor’s near-brighter mask behavior. Models are downloaded lazily and separately on first use; downloads stream to disk, use pinned SHA-256 checksum, and remove legacy cache files only after replacement weights load successfully. The executable remains separate from model weights.

Added a detected display-resolution readout and matching preview-size option while retaining WGPU, and added ICC for the display

Testing

Manually tested on Windows 11 with an Intel Core i3-1315U, 16 GB DDR4 RAM, and Intel UHD Graphics. No automated tests were added because the project does not currently have a test suite. Actual executable and installer sizes still require a platform build.

AI Use

The code was mostly AI-generated, then edited, integrated, tested, and largely corrected by a human.

[CyberTimon]

Thanks for the PR. Just my thought, I think we first have to implement good GPU acceleration for SAM / depth / normal / denoise models before moving on to larger models. I will try to do this in the next days :)

[WirePortLabs]

I'm working on something that makes it so you can use Depth-Anything-V2 and SAM 2 if the user wants to and as a fall back if needid and i added CPU + GPU and fallback for cpu only I'm doing some testing on the code and yes it dose work on igpus

[WirePortLabs]

im working on it but its 65-85% done its can use the gpu + cpu it Windows NVIDIA CUDA first, then DirectML, then CPU fallback Windows Intel and AMD GPUs DirectML acceleration Apple Silicon CoreML, with CPU fallback Linux and Android CPU inference for now

[WirePortLabs]

rn im working on Phones and tablets: investigate Android and iOS support, including touch-friendly UI, photo import/export, and realistic mobile AI performance.

Integrated GPUs (iGPUs): Intel UHD/Iris Xe/Arc integrated graphics, AMD Radeon integrated graphics, and Apple Silicon's unified GPU.

Linux: native build, packaging, photo access, GPU backends, and CPU fallback.

Low-end hardware: detect available acceleration, avoid assuming a dedicated GPU exists, and fall back safely when memory or drivers can't handle a model.

More hardware combinations: keep the editor usable even when AI acceleration isn't supported.

[WirePortLabs]

added Intel GPU and compatible Intel NPU OpenVINO

[WirePortLabs]

testing is happing

[WirePortLabs]

retesting