Design unique icons with AI

AI image generators are already a practical way to create distinctive SVG-style icons for knowledge work—solving a common frustration with stock icon libraries where the same graphics show up everywhere. The core pitch is simple: instead of relying on Flat icon’s searchable, recolorable SVG catalog, use Midjourney to generate icons that are more likely to be unique to a specific project, then refine them into transparent, reusable assets.

The motivation comes from repeated exposure to identical icons across YouTube, websites, posters, and office materials. Even with a strong workflow—Flat icon’s fast search, multiple styles, recoloring, and SVG downloads for paying members—the creator’s work started to feel “copycat” because the same icons were being used by others. AI generation is positioned as a workaround: it removes the need to learn drawing/inking techniques while still producing visuals tailored to a person’s thinking and project context.

Midjourney is used as the main tool, accessed through a Discord bot. Prompts begin with “/imagine” and then specify what kind of output is wanted—often “simple SVG icon”—followed by the subject (e.g., “New Year celebration”), optional color guidance, and style words. A key operational detail is quality control: setting “q 0.5” halves processing quality to save GPU time and quota, and the transcript claims that half-quality is often “perfectly okay” for icons. For convenience, Midjourney settings can be adjusted once (including selecting “version 4” and half quality) so the prompt doesn’t need to repeat those parameters.

Turning AI output into usable icon assets requires post-processing. Generated images are first saved, then background removal is handled by PhotoRoom, which the transcript praises for fast, free background removal without size limits. For cases where unwanted elements appear (like corner artifacts), PhotoRoom’s “Magic retouch” is used to remove specific parts before background removal and cropping.

Color control is treated as a prompt engineering problem. The workflow uses palette tools in Obsidian (Paletin) to generate named color schemes such as monochromatic, analogous, complementary, split complementary, triadic, and tetriadic. Since Midjourney doesn’t reliably accept hex codes, the transcript recommends using a color picker (via Coolors) to obtain human-readable color names (e.g., “Viridian Green,” “coffee brown”). Negative prompts (“--no …”) are used to exclude specific terms like “coffee” or “text,” and the transcript notes that excluding “person” can prevent unwanted figures when using artist-style prompts.

To increase variety, the transcript recommends generating “sets” of icons (rather than one-off results) and then upscaling with “U” commands. Upscaling can change details, sometimes requiring multiple upscales to get a preferred look. A practical warning is that half-quality plus upscaling can yield blurrier results, so sets are better generated at full quality.

Finally, the transcript argues that AI-generated icons can outperform stock libraries for uniqueness, while Flat icon still remains useful for inspiration and for ideas like searching “New Year” to quickly browse variations—especially since Midjourney struggles with correct numbers (e.g., “2023”). The takeaway is a repeatable prompt structure: output type + subject + palette + style + exclusions + optional performance switches—followed by background removal and cleanup—so icons can be integrated into documents and mind maps with a more personal visual identity.