Style > gpt-image-2-mspaint

This GPT Image 2 prompt is going insanely viral right now.

“Redraw the attached image in the most clumsy, scribbly, and utterly pathetic way possible. Use a white background, and make it look like it was drawn in MS Paint with a mouse. It should be vaguely similar but also not really, kind of matching but also off in a confusing, awkward way, with that low-quality pixel-by-pixel feel that really emphasizes how ridiculously bad it is. Actually, you know what, whatever, just draw it however you want.”

• https://x.com/arrakis_ai/status/2049689793118998717

Training Notes

This LoRA behaved very differently from a typical style LoRA.

My initial training used dim16 / alpha16 / lr=5e-5 / cosine, which only introduced subtle MS Paint-like characteristics while preserving the base model's strong anatomy, facial quality, lighting, and rendering.

After several experiments, I found that "bad drawing" behaves very differently from a normal art style during LoRA training.

Main observations:

• Lower dimensions tended to damage anatomy before fully learning the style.

• Higher dimensions preserved anatomy while separating the "bad drawing" characteristics more effectively.

• Surprisingly, dim32 handled a much higher learning rate without collapsing.

The final training settings were:

• dim32 / alpha8

• lr = 1e-3

• 575 constant steps

• 325 cosine steps

This produced the best balance between:

• rough MS Paint-like outlines,

• jagged pixel-like edges,

• intentionally messy coloring,

• while still keeping recognizable human anatomy.

One interesting observation is that the face remains relatively attractive compared to the original reference images. The base model still preserves much of its facial prior, while the LoRA mainly affects line quality, coloring, and overall rendering style.

This turned out to be closer to "an AI intentionally drawing in MS Paint" rather than simply degrading image quality, which was exactly the goal of this project.