Real SVG, fake SVG : Recraft V4 and the end of the vector designer

Take five services advertising themselves as “AI SVG generators” in the top Google results. Ask them all the same thing: a minimalist logo, thin lines, transparent background, one word of your choice. Download the five .svg files they return. Open them one by one in a text editor: VS Code, Sublime, or simply Notepad on Windows.

Four files out of five will start with something like this:

<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1024 1024">
  <image xlink:href="data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAA..."
</svg>
Code language: JavaScript (javascript)

Followed by tens of thousands of unintelligible characters. It’s a PNG. A PNG in an SVG costume. The root tag is valid, the file extension is honest, but the content is a raster image encoded in base64. On screen, it scales exactly like a PNG — which is to say, badly. Zoom in and it pixelates. In Figma or Illustrator, it’s a flat image, not a manipulable vector object. You paid for vector, you were delivered disguised raster.

The fifth file, however, starts like this:

<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 200 200">
  <path d="M150 150 C 175 100, 125 50, 100 75" fill="#1e4878"/>
  <circle cx="100" cy="100" r="40" fill="none" stroke="#1e4878" stroke-width="2"/>
</svg>
Code language: HTML, XML (xml)

Real vector. Explicitly declared Bézier curves, hexadecimal colours, numeric coordinates. Zoom in: no degradation. Edit it: every stroke remains a manipulable object.

This article is about that dividing line. The majority of tools that sell themselves as “AI SVG generators” are cheating — they dress a raster in a .svg extension and count on no client ever checking. In 2026, Recraft V4 is the only mainstream tool that reliably generates native SVG from a text prompt. This technical distinction has significant economic consequences for three identifiable professions: brand graphic designer, icon designer, vectoriser. And it directly extends the analytical framework laid out in the previous article on AI image generation in 2026 — the pixel as commodity — into a terrain where commoditisation is more subtle, more partial, and where technical deception is widespread.

Why vector is not an image

The distinction is not merely technical. It structures two different professional markets, two different production pipelines, two different economic regimes. Before discussing the tools, we need to clarify what we’re actually talking about — because a significant part of the confusion surrounding AI-generated SVG stems from a poor understanding of what a vector file actually is.

A raster image is a grid of pixels. Each pixel has a position and a colour. When you zoom in, the display software reproduces each pixel at a larger size — that’s what produces the characteristic staircase effect of pixelation. Raster formats are PNG, JPG, WebP, GIF, TIFF. They store information directly as a pixel matrix.

A vector file is a sequence of mathematical instructions. No pixels, no grid, no matrix. Instead, descriptions like “draw a Bézier curve from point A to point B with these two control points”, “fill this area with this colour”, “apply this gradient effect between these two coordinates”. Rendering happens at display time, by the software or browser, which interprets the instructions and calculates the final image at the desired resolution. Vector formats are SVG (Scalable Vector Graphics, the W3C open standard), EPS, AI (Illustrator), and PDF in its vector usage.

Worth noting in passing — to head off a frequent confusion with another meaning of “vector” in computing — the “vector” we’re discussing here is not the embedding vector used in machine learning, which is a numerical representation of meaning in a space of several hundred or thousands of dimensions. The term means two different things depending on context. This article deals exclusively with vector graphics.

The practical consequences of the difference between raster and vector are numerous:

For a button icon in an interface, for a logo that needs to display at 32px on a favicon and 5 metres tall on a storefront sign, for a set of pictograms manipulated via CSS — vector is the only correct answer. And that’s precisely why the “fake SVG” deception is so damaging: it robs the client of the technical benefits they specifically chose the format for.

The text editor test: unmasking fake SVG in ten seconds

This is the most useful demonstration in this article, and one nobody seems to explain clearly in plain language. Verifying that an SVG is a real SVG is trivial. The only obstacle is cultural — 90% of users don’t know they should check, or can’t imagine that the content of a .svg file could be anything other than what the extension promises.

The method is simple. Open the file in any text editor. On macOS and Linux, the command line head -c 500 logo.svg shows the first 500 characters. On Windows, right-click the file, choose “Open with” and select Notepad or VS Code.

You’re looking for two opposite signatures.

Real SVG contains explicit descriptive tags. You’ll see <path>, <circle>, <rect>, <polygon>, <line>, <g> (a group of elements), and attributes like d="M..." (a curve path), fill="#hexcode", stroke-width="2". The code is readable, structured, manipulable. A human or a script can change a colour, move a point, add an element. It’s code, not binary data.

Example of real SVG, simplified:

<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 200 200">
  <circle cx="100" cy="100" r="80" fill="#1e4878"/>
  <path d="M60 100 L90 130 L140 70" 
        stroke="white" stroke-width="6" 
        fill="none" stroke-linecap="round"/>
</svg>
Code language: HTML, XML (xml)

Twenty-two lines, a few hundred bytes, perfectly readable. It’s a checkmark inside a blue circle. Open it in Figma: every element is individually manipulable.

Fake SVG contains a single <image> tag followed by an xlink:href that begins with data:image/png;base64, and then a string of several thousand — often tens of thousands — apparently random characters. That string is the base64 encoding of a PNG file. The browser or software decodes the string, reconstructs the PNG in memory, and displays it. You have a file with a .svg extension, with a <svg> root tag, but whose actual content is a raster image.

Example of fake SVG:

<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1024 1024">
  <image width="1024" height="1024" 
         xlink:href="data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAA[...50,000 characters...]"/>
</svg>
Code language: HTML, XML (xml)

In use, this file behaves exactly like a PNG. Zoom in and it pixelates. Edit it and it’s non-editable as a vector object — Figma treats it as an imported image. Its file size is typically larger than an equivalent PNG because of the base64 encoding overhead (33% inflation), and it benefits from none of the advantages of the vector format.

A non-exhaustive list of tools producing fake SVG in 2026: several SaaS services with generic names ending in ai-svg-generator.xxx, some image generation functions within broader platforms that expose a .svg export that’s only SVG in the container sense, and several Chrome extensions or free applications that actually act as wrappers around raster models.

List of tools that produce real native SVG: Recraft (V3 SVG and V4), Adobe Illustrator with Firefly Text-to-Vector, Iconify AI specifically for icons, and specialised vectorisers like Vectorizer.AI that convert a supplied raster to clean SVG via algorithmic tracing.

The text editor test should be a systematic reflex before any purchase or production integration.

Recraft V4: architecture, strengths, limits

Recraft is, in April 2026, the only mainstream tool that reliably generates native SVG from a text prompt at a visual quality level that makes it usable in professional production. It’s worth understanding why, because this specificity stems from an architectural choice that structurally distinguishes it from its competitors.

Most image generation models are diffusion models trained on image-text pairs where the image is a raster. If you ask them for SVG export, two strategies are available: encapsulate the generated raster in an <image> tag (fake SVG), or trace the raster after the fact with a classical vectorisation algorithm. The first is a deception. The second produces a technically real SVG but one that’s visually inferior, because vectorisation by tracing a raster always loses curve quality.

Recraft made a different architectural choice: training a model whose native output consists of vector paths. The model doesn’t generate a raster that then needs converting — it directly generates Bézier curves, coordinates, colours and groups. This specialisation explains the visual cleanliness of the outputs: the curves are smooth because they’re native, not because they were reconstructed from pixels.

Four variants are available in the Recraft V4 interface: Standard (1024px raster), Pro (2048px raster), SVG, and SVG Pro. All share the same prompt base and the same stylistic consistency. The two vector modes differ in acceptable complexity and detail precision — SVG Pro tolerates more complexity and produces more precise paths.

Where Recraft V4 excels:

• Flat icons in outline or filled style, perfectly sharp, exportable directly into an icon library or design system.
• Minimalist logos, marks in geometric or simple typographic styles, monograms.
• Flat illustrations of the kind popular on SaaS websites, onboarding pages, editorial thumbnails.
• Consistent brand kits thanks to Style Lock — the function that lets you train a custom style on 5 to 10 reference images and then generate hundreds of assets consistent with that style.

Where Recraft V4 fails or produces uneven results:

• Complex vector illustrations with sophisticated multi-layer gradients.
• Very detail-dense vector scenes, which often come out with artefacts or spatial inconsistencies.
• Highly customised typographic lettering — typography remains a weakness, as with most image generation models.
• Photorealistic vectorised illustrations — a marginal but requested use case.

The Style Lock feature deserves dedicated attention. For an agency or brand team that needs to produce a high volume of assets consistent with an existing visual identity, it’s an extraordinary production multiplier. Upload ten or so brand visuals — colour palette, existing treatments, iconographic language. The model learns the signature and all subsequent generations respect it: same proportions, same stroke weight, same corner treatment, same palette. Where the difficulty of brand design historically lay in maintaining consistency at volume, Style Lock solves the production problem while leaving the difficulty of the initial design entirely intact.

Recraft’s Vectorize function converts a supplied raster (PNG, JPG) to clean SVG. It’s more reliable than Illustrator’s Image Trace on simple logos and pictograms, but falls short on complex illustrations. For pure vectorisation from raster, Vectorizer.AI remains more precise.

Pricing: freemium with public gallery (your images are visible to all users), $12/month for the private commercial plan, $33/month for the Pro plan with unlimited access to Pro and SVG Pro generations, API available for integration into automated workflows.

Adobe responds: Firefly Text-to-Vector in Illustrator

Recraft isn’t alone. Adobe integrated a Text-to-Vector Graphic function into Illustrator since 2024, powered by Firefly. Generation happens directly in the Illustrator canvas, from a prompt, and the resulting paths are immediately editable within the professional environment.

It’s a different proposition, worth comparing point by point.

Adobe’s advantage lies in native integration into the existing professional workflow. For a studio already producing in Illustrator, not having to leave the application to generate assets and adjust them is a genuine productivity gain. Adobe’s contractual indemnification for enterprise clients — Firefly is trained on a licensed corpus, which legally covers commercial use — is also a decisive argument for the legal departments of large organisations.

Adobe’s disadvantage lies in the entry price and relative visual quality. A full Illustrator subscription, or Creative Cloud All Apps, is a significant cost if the Text-to-Vector function is the sole need. This is a subject I’ve already addressed in a broader piece on Adobe’s pricing capture — the same analytical framework applies strictly here. On pure visual quality, blind tested, Recraft V4 outputs are regularly preferred by designers for non-corporate styles, while Firefly excels on more institutional styles.

Synthetic comparative verdict:

• Adobe Firefly Text-to-Vector for companies already in the Adobe ecosystem, with legal exposure, working on institutional styles.
• Recraft V4 for visual quality across varied styles, pricing flexibility, and Style Lock as a brand consistency mechanism at volume.

For an agency or freelancer, combining both is defensible: Recraft for rapid production, Illustrator for final adjustments and integration into print deliverables or existing client workflows.

The real losers of generated vector

The effect of vector generation models on professional markets is mechanical and localised. A cold inventory, in the register established in the previous article on image generation.

The junior UI icon designer. Producing a consistent set of 50 icons for an application — same style, same grid, same stroke weight — was a multi-day deliverable. Recraft V4 with Style Lock does it in ten minutes, with files directly editable in Figma. The entry segment of the profession is being compressed. What remains: the design system architect who defines the grid, the consistency, the usage rules. Not the executor who produced them.

The budget freelance logo designer. Fiverr, 99designs, and their equivalents relied in part on producing simple logos at low prices — between €50 and €200 for a basic logo, delivered in a few days. This entry segment is disappearing: a private client or small business can now generate a hundred logo variations for €5 to €12 on Recraft, pick three or four they like, and adjust them themselves in free Inkscape or Figma. The compression is brutal, and already measurable in platform order volumes.

The vectoriser. Converting a photo, a scanned drawing, or an old raster into clean SVG was a standalone service, priced at several tens of euros per file on specialist platforms. Recraft’s Vectorize function, and even more so Vectorizer.AI for pure conversions, reduces it to a click and a few cents. The profession disappears as a distinct category. It survives as a quick step integrated into broader workflows, but no longer as an autonomous service.

The vector editorial illustrator. The flat thumbnails that dress corporate blog articles, the unDraw-style illustrations popularised in the 2018–2022 period, SaaS onboarding visuals, friendly 404 error page pictograms. This entire segment, which had been a stable market for specialised illustrators for a decade, is collapsing. The cost-to-quality ratio tips entirely toward generation.

The mechanism is rigorously parallel to the one I documented for developers last January: compression of the intermediate segment, displacement of value toward upstream (architecture, vision, definition) and downstream (technical integration, specialist finishing), elimination of the intermediate execution profession.

What gains value in vector design

The inventory of destructions only makes sense if accompanied by an inventory of reclassifications. Here is what gains value, or retains it, in a regime of commoditised vector generation.

The senior brand designer. Their value is not in asset production, which is commoditising — it’s in defining a visual universe precise, structured and differentiated enough to be parameterisable in a Style Lock and then generate hundreds of consistent assets. This definition requires deep knowledge of the client, the market, the visual semiotics of the sector, and what resonates with a given audience. It’s a profession that strengthens, not disappears.

The design system architect. Designing consistent icon libraries, design tokens, reusable components, scaling and animation rules — this profession was already growing before the arrival of vector generation. It accelerates, because automated generation requires solid prior rules to produce consistent output at scale. Without architecture, Style Lock produces noise. With architecture, it produces consistent volume.

The auteur vector illustrator. The vector equivalent of author photography discussed in the previous article. A recognisable visual signature, a distinctive stylistic universe, a narrative consistency that can’t be reproduced by a Style Lock trained on ten images. The market for signed editorial illustration — magazine covers, illustrated books, author campaign visuals — holds, and may even strengthen in contrast to the commoditisation of the generic.

The SVG technical integration specialist. File optimisation (SVGO and its derivatives), CSS and SMIL animation, lazy loading, web performance, accessible integration (ARIA roles, contrast ratios, text alternatives). An emerging profession where pure technical expertise is rare and well compensated. Generation produces usable files, not optimised ones. The distance between the two is work.

The specialist typographer. Vector generation models still struggle with complex custom lettering, calligraphy, sophisticated typographic marks. This is a stable niche, possibly durably stable, because the technical difficulty of lettering and the craft dimension of the profession resist models trained on generalist corpora particularly well.

Practical case: an e-commerce production workflow with Recraft V4

From theory to applied workflow. The use case is concrete: a WooCommerce store needs an icon set for its category pages, guarantee and service pictograms, and onboarding visuals for the checkout funnel. Conventionally, that’s two to three days of freelance work billed between €600 and €1,200 for a complete set. Here’s how that restructures in 2026.

First step, before any generation: define a Style Lock from the client’s brand guidelines. Three to five reference images — main logo, existing brand treatments if there’s an established identity, or a simple moodboard if the project is starting from scratch. The model learns the visual signature. This step is crucial and remains highly human: the choice of references conditions everything that follows.

Second step, generating the icon set. A loop of prompts — “shopping cart icon, monoline style, transparent background”, “delivery truck icon, monoline style”, “payment security padlock icon”, and so on. Each prompt produces four variations in native SVG. Selection is done by eye, keeping one icon per concept and regenerating those that don’t work. For a set of twenty icons: forty to sixty minutes of actual work.

Third step, adjustments in Figma or Illustrator. SVGs produced by Recraft are clean but not always perfectly aligned with your design system grid. The final adjustment — standardising stroke weight to exactly 2px, aligning corners to a 24×24 grid, harmonising corner radii — takes about twenty minutes.

Fourth step, optimisation and WordPress integration. Run files through SVGO to reduce file size (typically 40 to 60% gain), import into the WordPress media library, integrate into Gutenberg blocks or via shortcodes depending on target pages. For a WooCommerce store managed with a custom child theme, integration via dedicated CSS classes is the cleanest solution.

Indicative total cost: approximately two hours of work at a developer’s rate, plus the monthly Recraft subscription. In practice, between €80 and €150 for a deliverable equivalent to what was billed at €600 to €1,200 two years ago. A cost reduction in the order of 80 to 90%.

Important caveats:

• Verify the commercial licence of the Recraft plan being used. The freemium plan does not cover private commercial use.
• Test each file systematically with the text editor method before integration. Recraft V4 produces native SVG, but a quick check costs nothing and protects against future regressions.
• Keep a record of the prompts used, if only for later regeneration should the client request adjustments or a complementary set in the same style.
• For legally sensitive brands, prefer Adobe Firefly Text-to-Vector for its training licence framework, even at the cost of some perceived visual quality.

Honest vector as a technical act

The distinction between real and fake SVG is not a technical detail. It’s an economic dividing line. Tools that produce fake SVG — a PNG in a vector costume — are selling snake oil to clients who don’t check. The text editor test, which takes ten seconds, should be a systematic reflex before any production integration. Its simple adoption would be enough to eliminate the majority of dishonest providers from the market.

Tools that produce real native SVG — Recraft V4 leading, Adobe Firefly Text-to-Vector as the more legally protected alternative — don’t close the vector graphic design profession. They shift its value. Execution commoditises, as with raster. What gains value is upstream — identity definition, system architecture — and downstream — technical integration, optimisation, consistency at scale.

What is fading in 2026 is not the graphic designer’s profession. It’s its intermediate version — execution without vision — which was already under pressure for a decade and is now watching its last protection collapse. As with photography, the commoditisation of the deliverable revalues what cannot be commoditised: judgement, stylistic universe, intention.

The pixel was a commodity. The vector is too. The honest Bézier curve remains a human act.