See it as a diagram
Everything below, as a diagram you can edit. Describe yours and see it in seconds.
No account needed · Editable canvas, not a picture
What each format actually is
An SVG is a description of shapes. It has no resolution, so it renders at whatever size the viewer asks for, and a 4 KB file is as sharp on a 5K display as in a thumbnail. A PNG is a grid of pixels, which means the size decision was made when the file was written and every later resize is either interpolation or waste.
That is the whole argument, and the vendors have already acted on it. Microsoft is the clearest case: the Azure Architecture Center ships the set as general purpose SVG files, states there is no Visio stencil format and no plan for one, and offers no raster alternative at all, so on Azure the format question does not exist. AWS publishes an icon package alongside its PowerPoint toolkits. Google offers SVG and PNG. Datadef mirrors the vendor SVGs rather than converting them, which is why the AWS collection is SVG only.
Where SVG wins, and the three places it breaks
SVG wins in every editing tool, in slides, in print, on high-density screens, and anywhere the diagram might be zoomed. PowerPoint and Keynote both keep an inserted SVG as vector artwork, so the size you drag it to stops mattering; a raster inserted from a small size folder and dragged up is an upscale and looks like one on a projector.
It breaks in three specific places. Destinations that strip or refuse SVG, which still includes several wiki and ticketing surfaces and most email clients. Files that reference fonts the viewer does not have, where labels silently fall back to something else and your careful spacing collapses. And embedded raster content, because an SVG containing a screenshot is a PNG wearing a costume.
The font trap is the one that catches generated diagrams. If an SVG is going anywhere you do not control, either convert text to paths or accept a fallback font, and look at it in the destination before it ships. Aspect ratio is the other habit worth fixing: hold Shift while dragging a corner handle, which is also the Azure rule stated as a requirement rather than a preference, since the published terms ask you not to distort or change icon shape in any way.
When PNG is the right answer, and at what size
Choose PNG when the destination cannot render vectors, when you need one file that behaves identically everywhere, or when the diagram is being pasted into a ticket or a chat thread and will never be zoomed. Then pick the size on purpose: export at roughly twice the width the image will be displayed at, so it survives a high-density screen.
That gives concrete numbers. A GitHub README renders its content column at around 900 pixels on a desktop, so 1800 pixels wide is the floor for a README diagram. A slide that fills a 1920-wide projected frame wants an export at least that wide. A Confluence page body is usually narrower than the README, so the same 1800 pixel export covers both.
JPEG is only correct when the image is photographic or when file size dominates, because it has no transparency and it puts compression artefacts around exactly the thing a diagram is made of: hard edges and thin text. Datadef exports PNG and JPEG; for diagrams, use PNG.
Destination Format --------------------------- ------------------------------------- Editing, slides, print SVG (scales, stays sharp, small file) README, wiki, Confluence live embed image, re-rendered on change Ticket, chat, email PNG at ~2x the displayed width Photographic background JPEG (no transparency, lossy edges)
The format question hides a freshness question
Arguing about SVG and PNG is arguing about the last step. The failure people actually feel is not a blurry icon, it is a crisp diagram that describes an architecture retired two quarters ago, because the export was a one-off and nobody re-ran it.
A live embed sidesteps the whole thing: one markdown image line, the image re-renders from the canvas, and viewers need no account. When the source is a connected repository, repository sync regenerates the diagram from the branch on a daily sync.
The mechanism that makes a daily regeneration safe rather than annoying is a structure fingerprint: a hash of the parsed draw plan, not of the commit. A commit that edits a README, a test or a comment moves the sha and leaves the fingerprint alone, and a sync whose fingerprint is unchanged skips generation outright. So the exported picture only moves when the architecture moves, which is the property a hand-run export can never have.
FAQ
Should I use SVG or PNG for AWS architecture icons?
Why does my SVG icon look wrong after resizing in PowerPoint?
What resolution should I export an architecture diagram at?
Can I export a diagram as a PDF?
Do SVG icons work inside a GitHub README?