Icons Are Drawn to a Grid, and This Draws Them One at a Time
Open any icon set that feels professional and the individual drawings are unremarkable. What makes it work is that every icon in it was built to the same rules: the same stroke width, the same corner radius, the same optical size inside the same invisible box, the same decision about whether shapes are outlined or filled. Those rules are the product. A tool that draws one lovely icon at a time is solving the part that was never the hard part.
Use this without the search next time. Prathom Workbench puts Prathom's tools in your toolbar.
Add to Chrome — freeDrop an image here, or click to browse
Up to 25 MB. Uploaded temporarily, deleted after processing.
What it does
- One centerd subject with one stroke weight
- Even margin on all four sides
- Square 512 pixel output
- Built to stay legible at 24 pixels
How to use Icon Generator
- 1
Give it one object
An icon is a single idea. A reference containing two objects produces an icon that is trying to say two things, which at small size says neither.
- 2
Run the redraw
The subject is redrawn as flat solid shapes with one stroke weight, centerd with margin, and the result is squared to 512 pixels.
- 3
View it at 24 pixels among other icons
Paste it beside the icons it will sit next to and reduce them all. The question is never whether it looks good alone.
How it works
The reference is scaled to about one and a fifth megapixels and repainted at 0.85 denoise, then squared to 512 pixels with a center crop, so what you download has the aspect ratio an icon is used at.
The instruction is a list of drawing conventions rather than a description of a subject: one centerd subject, one consistent stroke weight throughout, flat solid shapes, an even margin on all four sides, and a silhouette that stays legible at twenty-four pixels.
The negative prompt names the specific ways icons fail rather than generic artifacts — varying stroke weight, hairline strokes, multiple subjects, an off-center subject, a tight margin.
Below 24 pixels, only the silhouette exists
An icon is used small, and small changes what an image is made of.
At twenty-four pixels, a stroke that is two percent of the width is half a pixel and disappears into a gray wash. Two shapes separated by a thin gap merge into one. Interior detail — the handle inside the cup, the buttons on the shirt — stops being detail and becomes noise that softens the outline.
What survives is the outer shape and the largest one or two internal holes. So the design question for an icon is never what it depicts, it is whether the outline alone is distinguishable from the outlines it will sit beside. A camera and a briefcase are unmistakable at any size. A tablet and a notebook are the same rounded rectangle, and no amount of internal detail fixes that.
The reduction test is therefore the only test worth running, and it should be run against the neighbors rather than alone.
A set is a system, not a pile
The instruction here pins the conventions it can — one stroke weight, one subject, even margin — and pinning them within a single image is not the same as holding them across a set.
Between two runs, the model resamples. One icon comes back with slightly heavier strokes, another with more internal detail, another with the subject drawn a little larger inside its box. Individually all three are fine. Together they read as three icons from three sources, which is exactly the impression a set exists to avoid.
Real icon systems solve this with a keyline grid — a fixed template of circle, square, and rectangle guides that every icon is drawn against, plus a stated stroke width and corner radius. The grid is the reason the set coheres, and it is a rule applied by a person across many drawings rather than a property of any one of them.
Which makes the honest use of this page singular: generate one icon, for one purpose, or generate candidates to hand to whoever is building the system.
Where it sits against the logo page
These two produce similar-looking images and answer different questions, and confusing them wastes time.
A logo identifies an organization, is used at wildly varying sizes from a business card to a building, and carries legal weight — the similarity check on that page matters. An icon labels a function inside an interface, is used at two or three fixed small sizes, and is judged against the other icons around it.
A logo that fails is a rebrand. An icon that fails is a tooltip nobody needed to read.
Publication gate
This page ships once the workflow has been run against a single object photograph, a reference containing several objects, a thin outlined object, and a solid rounded object, with every output viewed at twenty-four pixels alongside a standard system icon set and its stroke weight compared between runs.
Examples
Photograph of a single object
The intended case. One silhouette in, one silhouette out, and the margin is preserved so it does not crowd its box.
Reference containing three objects
Predictably weak. The pass has to choose a subject or combine them, and a combined icon is unreadable at the size icons are used.
Frequently asked questions
Can I generate a matching set?
Not reliably. Each run samples independently, so stroke weight, optical size, and the level of detail drift between them. You can get a set that shares a subject matter and not one that shares a system, and the second is what makes a set look designed.
What is optical size?
The observation that shapes of the same measured size do not look the same size. A circle looks smaller than a square with the same bounding box, so icon systems scale circular forms up slightly to compensate. It is a per-icon judgment made against the rest of the set, which is precisely the information a single generation does not have.
Is a 512 pixel PNG enough for an app icon?
For a first look, yes. For shipping, no. Platforms want several sizes with different amounts of detail, some want a vector or a layered source, and app store icons have their own masking and safe-area rules. Treat the output as a design to redraw, not as an asset to submit.
Why does it refuse to draw text?
Because letterforms in an icon are almost always a mistake and generated ones are unreliable anyway. If a concept only reads with a word in it, it is not working as an icon, and at twenty-four pixels the word will not be readable regardless.
Is the AI Icon Generator tool free, and do I need an account?
There is no paid tier for Icon Generator, which means there is no free version being held back against one. You do not sign in, you do not enter a card, and the file you download carries no mark. What the page does ask of you is patience with the queue, because the job runs on a shared GPU host and finishes in its own time.