Magnific: the model shelf, not the render button

In April 2026, Freepik renamed itself after a tool it had bought two years earlier. The upscaler became the company. For an architect who bookmarked magnific.ai to sharpen a competition board, that is more than a branding note — the thing behind the bookmark is now a shelf holding roughly forty image, video and audio models under one subscription, and what it asks of you has changed from operating a tool to choosing between engines.

What it is and who makes it

Magnific is a general creative platform run by Freepik Company S.L.U., a Spanish firm trading since 2010. It acquired Magnific — then a standalone AI upscaler — in 2024, and on 28 April 2026 renamed the entire platform after it. The old magnific.ai address now points to magnific.com, with the original upscaler preserved at a legacy URL. Expect to meet both names in documentation, support articles and search results for some time yet.

Almost nothing on the shelf is Freepik's own work. A model is the engine underneath; a tool is the vehicle wrapped around it. Magnific licenses engines and sells the vehicle: Nano Banana 2 and Nano Banana Pro from Google, Flux.2 Pro and Flux 3 Video from Black Forest Labs, plus Kling, Seedream, Veo, Wan, Recraft, GPT Image, Ideogram, and ElevenLabs for audio. What Freepik actually built is the wrapper — the Magnific and Magnific Precision upscalers, an editing suite, a stock library of 250 million assets, and Spaces, a node canvas for chaining operations into repeatable workflows.

That distinction is the whole product. You are not buying a renderer. You are buying access, billing and orchestration across engines you could otherwise reach separately.

Where it sits in your week

Late. Magnific has no idea what a Rhino file is. It takes a picture and returns a picture, which places it firmly after geometry, on the presentation side of the project.

Three addresses in a real week. The first is rescue: an image exists at the wrong resolution and the board prints on Thursday. The second is volume — concept-stage option testing, where the point is to see thirty wrong images quickly so the right direction gets the week. The third is motion: a finished still becomes a five-second clip for a pitch that previously had no film in it at all.

None of that touches design. Everything it produces is representation.

What it is genuinely good at

The upscaler is still the reason to be here. It does not interpolate pixels; it invents plausible detail as it enlarges, which is why brick, foliage, fabric and weathered concrete come back looking photographed rather than stretched. It runs from 2× to 16×.

The second strength is duller and probably more valuable to a practice: one bill. Testing four models against the same massing image normally means four accounts, four credit systems and four sets of terms. Here it is one login, and the models that matter to architects — Nano Banana Pro, Flux.2 Pro, Seedream — sit next to each other. For a firm deciding what to standardize on, that comparison is the actual purchase.

Spaces earns a mention for the same reason. Once a chain works — upscale, then relight, then extend to a 16:9 board crop — it becomes a workflow you can rerun rather than a sequence you re-remember. And the platform can be driven from outside the browser: there is an API, an MCP server, plugins, and a ComfyUI integration, so a practice with any scripting capacity can automate the repetitive half.

Where it breaks

The upscaler invents, and that is not a quirk to work around — it is the mechanism. A 4× pass across a facade will add mortar joints, glazing bars and panel divisions that were never drawn, and it will add them convincingly. On a mood image, fine. On a planning submission, you have just published detail nobody designed, and the failure is invisible, because invented detail looks exactly like resolution. Every upscaled image needs an architect's eye on it before it leaves the office.

Nothing here reads your model. There is no source-file rendering, no geometry lock, no notion of a datum. If holding your massing is the requirement, this is the wrong category of tool.

Credits are an abstraction over thirty-odd separately-priced models, and the prices move. Magnific reserves the right to change credit costs without notice, and "unlimited" applies per model and shifts between plans. Forecasting a monthly spend is genuinely difficult, which matters if you are trying to put a line in a fee proposal.

And it is a general creative platform. Nothing in it is shaped by architecture — no material vocabulary, no entourage that understands scale figures, no concept of a design review. You are the architecture-specific part.

What it costs, what it asks

Web app, with desktop and mobile clients, an API, an MCP server, plugins and ComfyUI support.

Prices from Magnific's own pricing page, checked September 2026, excluding VAT. Free: attribution required, personal use only. Premium, $20/month or $14.50 billed annually, 240,000 credits a year. Premium+, $45 or $33.75 annually, 600,000 credits. Pro, $280 or $210 annually, 4,000,000 credits. Business is per seat for teams of 2 to 100 with a shared pool; Enterprise is custom.

Translating credits into work: Magnific Precision costs 90 credits for a 2K image, Nano Banana Pro 75 credits at 1K–2K and 150 at 4K, and Flux 3 Video runs 2,750 credits for five seconds at 1080p. So a Premium seat is thousands of stills a year, or a few dozen short clips — not both.

On the question a practice actually has to ask: Magnific lists "no model training with your data" as a feature of every paid plan, and its Enterprise tier carries GDPR, ISO/IEC 27001 and SOC 2 Type I. That is the company's published position, and it is a reasonable one to rely on for paid accounts. The free plan is a different matter — personal use only, attribution required. Client work should never go through it.

Who should bother

The principal: treat this as a procurement question, not a design one. If the practice already pays for stock imagery, the bundle argument is strong. If it does not, you are buying a lot of platform to reach one upscaler.

The project architect: worth an afternoon for the upscaler alone, particularly if boards are printing at sizes your renders were never made for. Skip the rest until you need it.

The graduate: this is the highest-value account of the three, because it is the cheapest way to build a real opinion about which model does what. Run the same image through four engines and you will learn more in an hour than from a month of comparison articles.

Ignore it if what you need is a render that respects the model you built. That is a different tool.

The rebrand was honest about what the company had become: a merchant, not a manufacturer. What it sells is convenient access to other people's engines, and that convenience is real. What it does not sell — and cannot — is the judgment about which engine to use, or whether the detail it just added to your facade was ever there.