Three years ago the confident position was that architectural visualisation would be among the first professional services to be automated away. Type a description, get a photorealistic building. The studios producing those images for a fee would go the way of stock photography.
That is not what happened, and the way it failed to happen is more interesting than the prediction. Rendering did absorb a great deal of machine learning, quite deeply in places. The market for outsourced rendering did not shrink. Prices at the bottom of it collapsed while prices at the top barely moved, which tells you something specific about what these models are good at.
Here is where the line actually fell, and the ten studios the industry settled on either side of it.
The AI that was already there
The part that gets missed in most of these discussions is that production rendering was quietly running on neural networks before anyone was typing prompts.
Denoising is the clearest case. Path tracing produces a grainy image that gets cleaner the longer you compute it, and the last few per cent of that noise historically cost more than everything before it. Learned denoisers, trained on pairs of noisy and converged frames, now do that final step almost instantly. It is standard in the major renderers and has been for years. Nobody markets it as AI because it arrived as a checkbox rather than a product.
Machine-learned upscaling followed the same path, along with denoising for animation sequences where flicker between frames used to be the hard problem. Material capture is a third: photograph a surface, get a physically plausible set of maps back.
All of that is real, all of it is in daily production, and none of it changed who does the work. It made the same studios faster. That distinction is the whole story.
What generative models did take
Concept and mood work, genuinely and permanently.
The early phase of a project, where somebody needs twenty atmospheres to argue about before anything is decided, used to cost days of an artist’s time. Diffusion models do it in an afternoon, and the output is good enough for the purpose, because the purpose is agreeing on a direction rather than describing a building.
Studios that were selling that phase as billable work lost it. Anyone claiming otherwise is not being straight with clients. The freelance end of the market, where the product was one attractive image with no accountability attached, took the same hit from the same direction.
What they did not take
Everything downstream of a decision.
A production render has to match a specific building. Specified materials, specified glazing ratio, a mullion where the drawings put the mullion, at four in the afternoon in March at that latitude. It has to be consistent with the eleven other views of the same building, and it has to survive an architect who will notice if the balustrade height changed between images.
Generative models are not built for that and, at time of writing, do not do it. They produce something plausible rather than something specified, and plausible is the opposite of the product here. The failure mode is subtle: an image that looks right, is wrong in a way the client cannot see and the planning authority can.
That accountability gap is sharpest in verified views for planning submissions, where the image is effectively a legal statement about what a building will look like from a fixed camera position. There is no version of that which a prompt produces.
The commercial consequence is a squeeze from below rather than a collapse. Concept work left. Accuracy work stayed and, if anything, became a larger share of what gets paid for.
The models hiding behind one word
Outsourcing rendering is not one transaction, and AI pressure lands differently on each version of it.
The production studio runs on volume. Large teams, defined pipelines, published turnaround, predictable economics per image. This is where learned denoising and upscaling paid off most, because throughput is the product.
The boutique runs on judgment. Small senior team, direct access to whoever decides, fewer projects at once. Largely untouched, because what is being bought was never the pixels.
The agency runs on scope. Rendering sits inside branding, film, sales tools and strategy, and you buy the campaign.
The freelance marketplace runs on price, and it is the part of the market that generative tools genuinely disrupted, because a single unaccountable image is exactly what they produce.
ArchiCGI
San Francisco. Exteriors, interiors, animation, 3D floor plans, virtual tours and virtual staging, alongside a library of more than 70,000 furniture and decor models that clients draw on.
Running since 2011, and on the company’s own count 600 artists and 144 modelers producing north of 760,000 renders a year. Zaha Hadid Architects, Scott Brownrigg, Compass and Grand Peaks appear on the client list, which is worth registering if you assume firms outsource 3D rendering only when they cannot afford talent in house.
At that volume the machine-learning gains in the pipeline compound in a way they do not for a five-person studio. Suits continuous demand rather than one landmark project.
Hayes Davidson
London, based on Weston Street in Southwark. Creative strategy, verified images for planning, visualisation, film, and real-time work handled by in-house app developers.
Founded over thirty years ago by Alan Davidson and now employee owned. It built the verified view methodology that the planning industry still runs on, and its portfolio stretches from the Shard to the High Line.
The clearest example of work that generative tools cannot approach, because a planning inspector is the reader.
Binyan Studios
Sydney, with offices in Melbourne, Brisbane, New York, London and Dubai. Content strategy, CGI stills, art direction, storyboarding, photography, live action film, animation, display suite planning and digital sales applications.
Founded in 2007 out of real estate rather than architecture, with a project count the company puts above three thousand. The display suite and sales app services tell you where the emphasis sits, which is selling apartments rather than winning competitions.
A fit for large residential launches that need a whole sales environment, not a set of files.
Mir
Bergen, on the west coast of Norway. Stills and films for architecture and landscape, and effectively nothing else.
Mats and Trond founded it in 2000 while still at the Norwegian Academy of Art and Design and still run it. The team is about ten artists, hired only to work from the studio, with no freelancers and no remote staff. Museums, operas, bridges, airports and stadiums make up much of the back catalogue, and the studio describes its own role as the invisible partner on them.
The purest boutique here, and the clearest case of a practice selling judgment rather than output. Book early and give it room.
Luxigon
Paris, on rue Montmartre, with studios in Los Angeles and Milan. Competition imagery, concept design, commercial work, events, and real-time 3D with AR and VR.
Twenty years of work, and the client index reads like a European architecture directory: Ateliers Jean Nouvel, BIG, 3XN, Aedas, Barozzi Veiga, Baumschlager Eberle, Carlo Ratti, alongside Bouygues Immobilier, BNP Paribas Immobilier and BMW.
The competition specialist, and the right call when an image has to win the commission.
RealSpace 3D
Vancouver, on Richards Street. Exterior and interior rendering, animation, floor plans and virtual tours, sold with a direct to principal workflow rather than a project coordinator.
Nineteen years in, and unusually blunt about the positioning: an outsourced partner for mid market developers offering the reliability of a large agency with the access of a boutique. Bentall Green Oak, Oppenheim Group, Four Seasons and Miracon Development are on the list. Turnaround is stated at two to three weeks, and they validate against CAD files instead of exports.
That CAD habit is the accuracy discipline this whole argument turns on, and it earns the fee on its own.
ZOA Studio
Budapest. Cinematic stills, signature animations, and a fixed package of fifteen high resolution renders delivered in a week.
Gensler, Snohetta and UNS sit on the architecture side of the client list, with Emaar, Brabus and Not A Hotel among the developers. The framing is motion first, and the animation reel is where the studio clearly puts its energy.
Animation remains the most stubbornly manual corner of this field, because temporal consistency is exactly what generative models handle worst.
xoio
Berlin. Architectural visualisation, product visualisation, illustration, animation, interactive work, and a strand they call future visions.
Mercedes-Benz, Daimler, Tishman Speyer, Solarlux, FSB, Feldberg Capital and the Fraunhofer Institute have all published testimonials, an unusual client mix for an archviz studio that explains the range. The Berlin projects, LXK campus and Am Tacheles among them, are the property side of it.
Worth a call when the brief crosses from a building into a product, a brand or a scenario nobody can photograph.
Beauty & The Bit
Madrid. Stills, animation, interactive work, VR and 360, plus original films and IP of their own.
Fifteen years in, run by Victor Bonafonte and Lina Garau, and openly a boutique: a collective of architects, photographers and designers rather than a production line. The phrase they use for what they sell is architectural communication, and the cinematic side of the portfolio backs it up.
Choose them when the project needs a story told well once, not a catalogue filled.
Pixelflakes
London. Architectural visualisation, animation, interactive work, and sketch studies, alongside workshops, CPD sessions and supplier presentations.
WATG, Tishman Speyer, HOK and PRP Architects are among the collaborators of the past decade. The firm calls itself an architectural marketing company rather than a rendering supplier, and running CPDs is the behaviour of a studio that wants to sit inside practices rather than take briefs by email.
The pick for architects who want a partner in the room during design instead of a vendor at the end of it.
What to ask a studio about AI now
The question is not whether they use it. Everyone does, and the ones who say otherwise are either describing a marketing position or have not looked inside their own renderer.
Better questions. Where in the pipeline does it sit, and is any of it generating content rather than cleaning it up? Denoising an image you specified and inventing an image you described are different products with the same output format.
What happens to entourage, meaning the people, the trees and the cars? Generated entourage is now common and mostly harmless, though it is worth knowing whether the building sits in a real surveyed context or a plausible one.
Does anything generated end up in a verified view or a planning submission? The answer should be a flat no.
And who carries it if an image turns out to be wrong? That is the question the whole distinction reduces to, and it is the reason this work is still bought from organisations rather than from prompts.
Picking the model before picking the studio
Continuous volume with deadlines points to a production studio. One project that has to be exceptional points to a boutique. A campaign with film, branding and a sales tool attached points to an agency.
The automation argument did not change that structure. It removed the cheapest tier of it, absorbed a lot of unglamorous computation into the tools, and left the part that was always about accountability exactly where it was.
Most outsourcing failures worth studying are still not quality problems. They are people who bought one model while needing another.
