Paint.NET ships internal Direct2D rewrite for WINE via AI-assisted reverse engineering

Paint.NET’s reliance on Direct2D has historically prevented it from running properly on WINE, a compatibility layer for Windows applications on Linux. The project’s solution, a from-scratch rewrite of Direct2D, was generated with AI assistance, bypassing the need for WINE’s incomplete implementation. This approach sidesteps the dependency entirely but introduces a new, unproven codebase that must now be maintained alongside the rest of the application. The decision reflects a growing trend of using AI to accelerate reverse engineering, though the long-term viability of such code remains uncertain.

The AI-generated Direct2D implementation spans 180,000 lines, a substantial addition to Paint.NET’s existing 700,000-line codebase. The scale of this rewrite suggests a high degree of complexity, yet the author explicitly describes it as "vibe coded," implying minimal formal review or testing. This raises concerns about latent bugs, performance bottlenecks, or edge cases that may only surface under real-world usage. Engineers adopting similar AI-generated solutions should anticipate significant manual oversight to address gaps in correctness, especially in low-level systems programming.

The rewrite required extensive manual intervention to fix critical flaws, such as missing reference counting for COM objects. These issues highlight the limitations of AI in understanding nuanced systems-level requirements, even when trained on large codebases. While the AI demonstrated capability in reverse-engineering Direct2D’s effects library, its output still demanded human expertise to align with the project’s architectural standards. This underscores the need for rigorous validation when integrating AI-generated code into production systems, particularly for foundational components like graphics libraries.

The solution is currently triggered only when running Paint.NET on WINE, limiting its exposure to a subset of users. This reduces immediate risk but also means the rewrite’s stability and performance are untested in broader scenarios. If the implementation proves reliable, it could pave the way for similar AI-assisted compatibility fixes in other projects. However, the lack of transparency around the AI’s training data or methodology complicates efforts to audit or reproduce the results. Engineers should treat this as a case study in rapid prototyping, not a template for scalable solutions.