Arrow
Real SVG vectors for icons, marks, diagrams and illustrations you can scale and edit.
Last verified
Best for
Arrow 2, by QuiverAI, writes vector files. Most models here paint pixels; Arrow draws shapes and paths, so the result is an SVG you can scale to any size, open in Figma or Illustrator, and edit point by point. The archer above is one file, shown at 32, 128 and 480 px.
Reach for it when the file will be scaled or edited later: icons, logos, stickers, badges, diagrams and flat illustrations. Recraft V4.1 SVG and Recraft V4.1 Pro SVG also make vector files, if you want a second opinion.
It also does bold, flat compositions in a few colours, like a poster or a cut-paper print. Name the colours and say "large simple shapes".
Arrow doesn't do photographs. Ask for one and you get a flat illustration of it. For photos, use Seedream. For a poster with lots of type, use Ideogram. Which model for what shows all five families on the same prompts.
Write the prompt
Use QuiverAI's template. It covers what a vector file needs that a picture doesn't:
Create [subject] for [use and size]. Use [style], [colors], and [composition]. Include [required details]. Avoid [unwanted details].
Name the size it will be used at. A 24 px toolbar icon needs far less than a 5 cm sticker, and Arrow draws to match. QuiverAI's fix for a cluttered result is to name the display size and drop texture. Asked for "a chocolate bar icon", Arrow drew a glossy, shaded chocolate bar. Told it was for a 24 px toolbar button, it drew an icon.
Say how it's built. Fills or outlines, stroke weight, rounded or square line ends, background. "One cream outline weight on a cobalt square, rounded ends, no fills" is what made the six icons above one family.
Give colors as hex codes. "Sunset colors" gave us a busy, shaded lynx in oranges and purples. Four hex codes gave a flat lynx in exactly those four.
Make a set from one prompt. Write the prompt once, then change only the subject for each icon. Keep the size, stroke and fill words identical, and the set will match. QuiverAI suggests checking the whole set at its real size and fixing only the icons that don't read.
For diagrams and line art, set the line weights. Name the viewpoint and say which lines are heavier. QuiverAI's advice is to keep major outlines heavier than construction details, as in the patent drawing above.
Give each reference one job. "Take the two inks, the misregistration and the die-cut border from Image 1; the subject is new." That's how the heron sticker took its style from our lynx.
Turn a sketch into a vector. Attach a drawing and say what the finished file should be. Asked to turn our sketch into "a clean vector illustration", Arrow traced the pencil texture and the paper too. Describing the finished file worked: "solid black fills and bold, even strokes ... no pencil texture, no construction lines and no paper". It still kept a few thin guide lines. The sketch here was made on Nano Banana 2.
Write the exact words, and check every letter. Arrow spelled "HALVARD" correctly and drew it as shapes, so it scales like the rest of the file. Short labels, like the patent drawing's "Fig. 1", came out as real text you can edit. Keep text to a short name or label: in our comparison Arrow misprinted two of five diagram labels.
Change one thing at a time. To fix a result, rerun the prompt with one change and name what stays. QuiverAI's pattern: "Change only [one element] to [new requirement]. Keep [silhouette, layout, colors, text, and other fixed details] unchanged."
Settings
Arrow has no settings of its own. Aspect ratio sets the shape of the file, and the prompt does the rest.
Every Arrow image costs 59 credits, a single icon or a busy scene, and takes about 90 seconds. Set Images per prompt to 1 until the prompt is right, then ask for the full set.
| Model | Credits per image | Other settings |
|---|---|---|
| Arrow 2 | 59 | Same price |
Limits
- Up to 14 reference images.
- No photographs. Text is best kept to a short name or label.
Sources
QuiverAI's own guides: Arrow 2 prompt guide and Text to SVG.
Images made in Batch Banana on October 2, 2026. Select Prompt on an image to read the prompt and model that made it. Side-by-side pairs and comparison rows show the first result.