Scumble Brings Open-Source AI Inpainting to Desktop

Overview

Scumble has emerged as a notable entry in the growing landscape of desktop-based AI image editing tools, offering users a free, open-source solution for inpainting tasks. Developed by DenRakEiw, the application is currently in an early 0.1.x state but provides a robust framework for integrating various generative AI models directly onto a user’s desktop.

The software distinguishes itself by prioritizing local control and layer-based editing, allowing artists and designers to manipulate AI-generated results with the precision of traditional graphic design workflows. By supporting both local rendering via ComfyUI and a wide array of cloud API providers, Scumble aims to bridge the gap between experimental AI capabilities and practical, editable output.

How Scumble Works

The core functionality of Scumble revolves around a simple yet powerful workflow: select an area, describe the desired change, and generate. Users can create selections using a variety of tools, including brush, rectangle, ellipse, lasso, magic wand, or object hover detection powered by the in-app SAM2 model. Additionally, text-based selection allows users to type prompts like "the handbag" to automatically isolate objects.

Once a region is selected, the generated result is applied as a new layer over the original image. This layer can be blended, color-matched to its surroundings, erased in specific parts, or regenerated multiple times without altering the base image. The editor supports stacking filter and text layers, with the ability to export the entire project as a PSD or OpenRaster file, ensuring compatibility with existing professional pipelines.

A unique feature is the integrated assistant, which allows users to drive the editor through natural language commands. This assistant acts as an external agent, processing requests one step at a time and requesting confirmation before executing any actions that incur costs. This approach provides transparency and control, particularly important when working with paid API services.

The rendering engine is highly flexible. Users can run models locally on their own ComfyUI instance or leverage remote services through providers such as Google, OpenAI, Black Forest Labs, ByteDance, and Qwen. The application also supports object masks and background removal internally using ONNX Runtime models like BiRefNet and RMBG, reducing the need for external dependencies for basic editing tasks.

Supported Models and Providers

Scumble’s strength lies in its extensive support for state-of-the-art image generation models. The software currently integrates with FLUX 3 Image, FLUX.2, GPT Image, Nano Banana, Seedream, Qwen Image Edit, and more. Each model can be accessed through its native API or via aggregator services such as ToAPIs, fal.ai, Replicate, WaveSpeedAI, Comfy Cloud, and OpenRouter.

Specific provider integrations include Google’s Nano Banana 2 and Pro models, OpenAI’s GPT Image 2.5 Flare and Sunburst, Black Forest Labs’ FLUX series, ByteDance’s Seedream 5 and 4.5, and Qwen Image Edit. For models like Seedream, Scumble connects directly through BytePlus ModelArk, while others may route through third-party inference platforms.

The application also introduces a novel "boxes in the prompt" feature for compatible models like FLUX 3 Image and Ideogram 4. Users can draw boxes on the image to specify exactly where elements should be added, kept, moved, or removed. This spatial guidance allows for more precise inpainting, transforming the selection into structured instructions for the AI model.

Recipes play a central role in managing these complex configurations. Instead of building node graphs, users select a model and provider from predefined recipes. Scumble ships with thirteen templates from Comfy Cloud, covering both image edits and new image generation. Users can also import their own ComfyUI workflows as custom recipes, provided they include the Inpaint Canvas node.

Availability and Development Status

Scumble is released under the GPL-3.0 license, making it free and open-source software. Currently, the application is available for Windows through the Microsoft Store and a direct installer. A Linux build in AppImage and .deb formats is also provided, though testing on Linux systems is still in progress. macOS support is planned for the future.

As an early-stage project, Scumble acknowledges that not all features have been fully tested end-to-end. Users are encouraged to keep backups of their images and report any issues through the project’s GitHub repository. Verified features include local rendering through ComfyUI, in-app helper models, the command core, the MCP server, and the tile engine for handling large documents.

Among API providers, FLUX 3 Image on Black Forest Labs and OpenRouter, GPT Image 2.5 through OpenRouter, and Ideogram 4 on fal.ai have been successfully tested. However, many other provider integrations and the assistant’s model calls remain untested against live services, indicating that the project is still in a rapid development phase.

For those interested in following the project’s progress, manual documentation, videos, and a developer blog are available at denrakeiw.com/scumble. The repository also includes a detailed changelog and invites community contributions through its active issue tracker.