Ideogram Vector 1.0 vs Krea 3: What You Hand Over at the End

Every design brief that starts with "make me a logo" ends, eventually, with somebody asking for the file. That is the moment when Ideogram Vector 1.0 and Krea 3 stop looking like two ways to do the same job. Ideogram Vector 1.0, released on October 22, 2026, emits an SVG you can open in Illustrator or Figma and keep editing as geometry. Krea 3, released a week earlier on October 15, emits a 2K raster image that is finished the instant it arrives.

Both are good at what they do. The problem is that the difference does not show up in a side-by-side of sample outputs, which is how almost every comparison of these two tools is written. It shows up three weeks later, in a file that arrives as a 2736-pixel PNG and gets sent back.

Two definitions of finished

Krea 3 is an instruction-based image editor. Its job is to take an image you already have and change part of it — swap the product, restyle the character, replace the background — while leaving the pixels outside the edit untouched. Text-to-image works, but it is the secondary path. Everything about the model is tuned for the loop: generate, look, adjust, regenerate. Krea's own median for a typical edit is under eight seconds.

Ideogram Vector 1.0 is not an editor at all. It has two endpoints and neither one takes an instruction about an existing composition. One generates SVG artwork from a prompt of up to 9,500 characters, optionally conditioned on reference images. The other converts a JPEG, PNG or WEBP of up to 50 MB into SVG. There is no inpainting, no region selection, no "keep everything else the same." If the file is wrong, you generate another file.

So the honest framing is not "which model is better." It is: at what point in the job does each one stop being useful?

What Krea 3 is actually good at

The case for Krea 3 is iteration speed against a reference you control. The parts that matter when you are choosing:

• Reference handling — up to six reference images in the app, with the API documentation allowing as many as eight. Above four or five, references get downscaled to 1,536 pixels on the long side, so the practical ceiling for detail-critical work is lower than the number suggests.

• Regional editing — a bounding-box edit changes the region and copies the rest verbatim. This is genuinely different from a full regeneration, and it is what makes product photography work viable.

• Thinking modes — a fast default and a slower high mode for multi-reference or multi-step compositions. High mode costs extra and is where the good results usually live.

• LoRA tuning — Krea trains small adapters for a product, a person or a house style. For a brand with a consistent look, that is a real moat, and it is the feature vector models have no answer to.

• Output — native 2K, with a 1K option. No vector path, no layer structure, no editable text.

What Ideogram Vector 1.0 is actually good at

Ideogram's pitch is that the SVG arrives as artwork rather than as a traced silhouette. The vendor describes individually movable, rotatable and recolourable elements, with gradients and typography preserved as editable objects rather than baked into paths. This is a launch claim and, as of this writing, nobody outside Ideogram has published a structured test of it.

The API surface is unusually plain, which is a point in its favour:

• Text-to-vector — POST /v2/svg/generate/ideogram-vector-1, prompt 1 to 9,500 characters, optional reference images, aspect ratio defaulting to auto.

• Image-to-vector — POST /v2/svg/vectorize/ideogram-vector-1, one image up to 50 MB, a mode of simple (fewer, simpler shapes) or default (finer detail).

• Quality tiers — low, medium and high, defaulting to medium. Price scales with the tier.

• Dry run — passing dry_run: true validates and prices the request without generating, storing or billing it, returning a price quote instead of an image. For anyone building a cost guardrail, this is the single most useful line in the docs.

• Async by default — both endpoints return a generation ID, with Ed25519-signed webhooks available.

There is one caveat worth knowing before you plan a pipeline around it. The vectorize endpoint exposes a group boolean documented as "request grouping of shapes into layers," followed immediately by "Grouping is not yet available." Upload a flat PNG and you get geometry back, but not the layer organisation the marketing language implies. The layer story is strongest on the text-to-vector path, where the model is composing rather than inferring.

The handoff test

Three situations separate these models faster than any benchmark.

Test one: the twelve-metre banner

Take a mark generated at 2K and scale it to a wall. The raster version is resampled; edges soften; the vendor-supplied upscaler helps but does not invent detail that was never there. The SVG version is re-rendered at whatever resolution the output device wants. For signage, embroidery, screen printing, laser cutting or vinyl, this is not a quality preference — it is the difference between usable and not.

Test two: the colourway

A client asks for the same mark in four brand colours. With a raster file, that is four edits or four regenerations, each of which may move a curve slightly, and none of which you can guarantee are geometrically identical. With an SVG whose fills are separate objects, it is a find-and-replace on a hex value. If the fills are not separate objects, you are back to redrawing. Which of those you get is exactly what the layer claim is about, and it is the first thing to verify on a real file.

Test three: the legal edit

Legal wants the registered trademark symbol moved two points right and the wordmark tightened. In a raster, that is a fresh generation with a hope. In an SVG with live text, it is a nudge. In an SVG where the text has been converted to outlines, it is a redraw — which is the quiet failure mode nobody mentions: vector does not automatically mean editable text.

What the file is worth, in money

Per-generation pricing is close enough to be a distraction. Krea 3 is $0.03 for a 1K output and $0.06 for a 2K output, before reference-image line items (roughly $0.02 for a 1K reference and $0.07 for a 2K one) and before high thinking mode. Ideogram Vector 1.0 is $0.13, $0.16 and $0.20 for low, medium and high text-to-vector, $0.32 at high quality with reference images, and a flat $0.10 for a conversion. At the tier a professional would actually use, the two are within a few cents of each other per asset.

The number that dwarfs both is the redraw. A designer converting a raster logo into clean vector geometry takes somewhere between twenty minutes and an hour depending on complexity. At a loaded rate of $60 an hour, that is $20 to $60 of labour against a $0.20 generation. Run it across a twenty-asset brand refresh and the raster path costs between $400 and $1,200 in retracing that the vector path does not.

That arithmetic also explains where OrcaRouter fits in a mixed workflow. Ideogram's vector endpoints are vendor-direct — we do not route them, and this piece would be misleading if it implied otherwise. What we do route is the raster side of the same job: models like GPT Image 2 and Gemini 3.1 Flash Image are in the catalogue on one key at provider list price with 0% markup, which is useful when the vector step is one box in a pipeline that also does moodboards, upscales and cutouts.

Where each one fails

Krea 3 fails at exactly the moment the asset leaves the screen. Hand it to a printer, a sign maker or a packaging vendor and you will be asked for a vector version you do not have. It also cannot give you a file whose text is text. For a poster that will only ever be a poster, none of that matters. For anything that becomes a system — a logo, an icon set, a packaging line — it is disqualifying.

Ideogram Vector 1.0 fails at the front of the job. It cannot iterate on your existing composition. It will not take your product photo and change the label. It has no regional edit, no reference-driven consistency training, and no way to say "keep this, change that" about an image you already like. It is also, as of October 23, 2026, unscored: Artificial Analysis's editing leaderboard ranks Krea 3 Ultra fifth at 1,152 Elo across 5,440 samples, and Ideogram Vector 1.0 does not appear on it at all, because the board measures instruction-based editing and this model does not do that.

A rule of thumb that survives contact with clients

Ask one question before picking: does the deliverable ever get bigger, recoloured, or printed on something other than a screen? If yes, generate the SVG, even if the exploratory work was done in a raster model. If the answer is no — a social post, a mockup, a hero image for a landing page — then a raster editor with regional control and a LoRA trained on the brand will get you there faster and cheaper, and nobody will ever ask for the file.

The mistake to avoid is doing the exploration in the vector model. Ideogram Vector 1.0 gives you one shot per generation with no regional correction, so a composition you are still deciding on is expensive to steer. Explore where you can edit. Ship where the file is right.