New Local Photo Editor Kokon Introduces Layered Editing Capabilities

A developer has released a new open-source photo editing tool, Kokon, which utilizes the Qwen-Image 2.1 Turbo model to enable layer-based editing on local hardware. According to documentation provided on GitHub, the software aims to provide users with the ability to isolate objects within an image and manipulate them as independent layers, a feature typically associated with proprietary professional software.

Technical Architecture and Workflow

The Kokon software operates through a multi-step pipeline designed to circumvent the inherent limitations of the underlying Qwen-Image 2.1 model, which does not natively support layer output. As reported by the developer on Reddit, the workflow consists of four primary stages:

  1. Selection: The user selects an object, and the MobileSAM model (running on the CPU at 44 MB) generates a mask.
  2. Extraction: A matting model on the CPU performs the cutout. Users can opt for a higher-quality edge refinement via Qwen for complex textures like hair or fur.
  3. Inpainting: The software uses a masked Qwen-Image edit to repaint the background where the object was removed. Performance benchmarks shared by the developer indicate this step takes approximately 15–25 seconds on an NVIDIA RTX 5070 Ti, while an M2 Pro processor requires one to two minutes per object.
  4. Blending: An optional step that calculates contact shadows and reflections to integrate moved objects back into the scene.

The developer has noted several technical limitations in the current version. Reported issues include the model occasionally generating new objects during the fill process, difficulty handling crowds or touching people, and an inability to reconstruct the hidden parts of an object behind another.

Hardware Compatibility and Performance

Kokon is released under the AGPL license and is compatible with Windows, macOS 15+, and Linux. It supports NVIDIA hardware (requiring 12 GB of VRAM, with an 8 GB beta available), as well as AMD and Intel via Vulkan.

For Apple Silicon users, the developer has implemented a custom engine written in Swift and Metal. On an M2 Pro, the engine reportedly processes a 1 MP image in approximately one minute, which is four times faster than the generic backend. On an M5 Pro, performance is estimated at 29 seconds per image.

Network and Integration Features

Recent updates to the software have introduced several connectivity features. Version 0.11 transitioned the application to a local server model (127.0.0.1), allowing for remote use across a Local Area Network (LAN) via a password-protected service. Additionally, a Docker image is available for NVIDIA-based headless Linux systems, and an MCP server (kokon-mcp) has been released to allow AI agents to perform image generation and editing tasks.

While the software provides a local alternative to cloud-based editing suites, the developer has categorized the current release as experimental, noting that the tool is intended to gather user feedback on specific failure cases to inform future updates.