UnHTML vs AI Screenshot to Figma Tools: Fidelity and Editability

AI screenshot to Figma tools infer layers from pixels. UnHTML reads the page's own HTML and CSS. Here is how that changes text, auto layout and cleanup.

COMPARISONDESIGNERPublished

Introduction

An AI screenshot to Figma tool looks at an image and guesses the layers. UnHTML reads the page's own HTML and CSS and maps them to layers. That one difference decides how faithful the result is, how much of the text and spacing is real, and how much cleanup you do afterward.

Neither approach is better in every case. If all you have is an image, a screenshot tool is the only option. If you have the page itself, reading its source removes most of the guessing. Below: what each one reads, how text and fonts differ, where auto layout comes from, and what happens to your data. Facts about the screenshot tools come from their own documentation, as of November 2026.

What an AI screenshot converter has to guess

A screenshot has no structure. Every button, card and heading is a block of pixels. A converter has to detect where components begin and end, work out which text is a heading, and estimate the spacing between things. image.to.design describes this as an analysis step that identifies component types, layout patterns and spacing, text elements and a font hierarchy, then outputs organized layers, auto layout groups and descriptive layer names.

That's useful. You don't need access to the source. It works on mockups, on competitor screens and on pages that no longer exist. Figma's own agent also converts a screenshot into editable layers (text, shapes and frames), and Figma says it is available on all paid plans and did not use AI credits during the open beta.

The limits are in the vendors' own wording. Figma says accuracy depends on the screenshot's clarity and complexity, and that custom or ambiguous elements may need more guidance. image.to.design says complex or unusual designs may need some manual cleanup. Every value in the output is an estimate. A 16 px gap might come back as 15 px or 18 px, and nothing in the file tells you which. If you have tried this before, you know the other version of the same problem: a screenshot with an image trace gives you shapes, not a structure.

What UnHTML reads instead of pixels

UnHTML is a Figma plugin that takes the page's source. You give it a self-contained .html file, a Safari .webarchive or a scene.json. For a live site, remote assets or a page built by JavaScript, you first capture a scene.json with the open-source CLI or a bookmarklet, as described on the product page.

Because it starts from the document, the text is the page's real text, the colors are the computed values, and the sizes are the rendered sizes. Nothing is detected from an image. The result is native frames, text, images and vectors.

It is not a pixel-exact clone. Some CSS has no Figma equivalent and falls back to something close. When a frame cannot reproduce the captured geometry, UnHTML falls back to absolute positioning rather than inventing a layout. And it has a hard requirement: you need the page's HTML. If you only have an image, this tool does not apply.

Text and fonts

Text is where the two approaches differ most. A screenshot tool reads text with OCR and font detection. Codia's documentation states plainly that tiny or blurry text hurts OCR and font detection, and that fonts must be installed locally or Figma falls back to its default. Misread characters are easy to miss, so Codia's own checklist tells you to confirm wording and font matches after every conversion.

An HTML import never has to read characters from an image. The text string and the font-family are in the page. The remaining problem is a font that is not installed on your machine. UnHTML substitutes it (Inter is the default) and lists each swap in the import report, so you know exactly which text styles to fix. The mechanics are covered in how missing fonts get replaced and logged.

In practice: with a screenshot tool you audit every text layer for errors. With an import you audit a short list of font swaps.

Auto layout

Both approaches can produce auto layout. They get there differently.

A screenshot tool infers it. Codia's docs say auto layout is applied only where a regular grid or stack is detected, and that very free-form layouts may stay as absolute positioning. That is a reasonable rule, but the padding and gap values are measured from the image.

UnHTML maps it. Spacing, padding, wrap and stretch come from the page's flex and grid CSS, so gap, padding and flex-wrap become the equivalent auto layout settings. When the geometry cannot be reproduced, the frame uses absolute positioning instead.

AI screenshot toolUnHTML
ReadsPixels of an imageThe page's HTML and CSS
TextOCR and font detectionThe page's own text and font names
Missing fontsFalls back to Figma's defaultSubstituted, each swap logged
Auto layoutInferred where a stack or grid is detectedMapped from flex and grid CSS
FallbackAbsolute positioning for free-form areasAbsolute positioning when geometry cannot be reproduced
NeedsAny imageA saved .html, .webarchive or scene.json
Typical cleanupText, spacing, component groupingFont swaps, unsupported CSS

The table summarizes the claims above. It is not a benchmark: it comes from vendor documentation and has not been tested head to head.

Privacy and input trade-offs

Where the work happens matters for confidential screens. Codia's documentation says the cropped image goes to Codia's cloud and comes back as a layer tree, and its plans set limits on resolution, file size and monthly conversions. Other tools handle this differently, so read each vendor's privacy page before you upload an unreleased design.

With default settings, an UnHTML import makes zero network requests. Fetching remote assets is an opt-in toggle, and scripts are stripped from the markup before rendering. For a longer checklist, see what to check about plugin privacy.

The trade is convenience. An image is one drag away. A faithful import needs a saved file or a capture step first.

Which one to pick

Use a screenshot tool when the image is all you have: a mockup, a screen from a presentation, a page that is gone. Budget time to check the text and spacing, because both are inferred.

Use an HTML import when you can save the page, such as your own site, a staging build or a page you are allowed to study. You trade a setup step for real text, real font names and spacing taken from the CSS. The two are not exclusive. You can use a screenshot tool for references and an import for pages you own. If you are also weighing code-first tools, UnHTML vs Anima covers which direction each conversion runs.

Frequently Asked Questions

Can an AI screenshot to Figma tool produce real auto layout?

Yes, where it detects a regular stack or grid. Codia's docs say very free-form layouts may stay absolutely positioned. The auto layout is inferred from the image, so check the padding and gap values.

Do I need the page's source to use UnHTML?

Yes. UnHTML reads a self-contained .html file, a Safari .webarchive or a scene.json capture. If all you have is an image, a screenshot tool is the right category.

Do screenshot converters upload my image?

Codia's docs say the cropped image goes to its cloud. Other tools vary, so read each vendor's privacy page before uploading a confidential screen.

Which gives more accurate text?

A screenshot tool reads text with OCR and font detection, so blurry text hurts it. An HTML import reads the text and font names from the page itself, and logs substitutions when a font is missing.

Questions

Can an AI screenshot to Figma tool produce real auto layout?
Yes, where it detects a regular stack or grid. Codia's docs say very free-form layouts may stay absolutely positioned. The auto layout is inferred from the image, so check the padding and gap values.
Do I need the page's source to use UnHTML?
Yes. UnHTML reads a self-contained .html file, a Safari .webarchive or a scene.json capture. If all you have is an image, a screenshot tool is the right category.
Do screenshot converters upload my image?
Codia's docs say the cropped image goes to its cloud. Other tools vary, so read each vendor's privacy page before uploading a confidential screen.
Which gives more accurate text?
A screenshot tool reads text with OCR and font detection, so blurry text hurts it. An HTML import reads the text and font names from the page itself, and logs substitutions when a font is missing.