In the first part of this article I described the impact of new LLM-based image tools like GPT-Image-1 and Gemini 2.0.Flash (Experimental Image Mode).

Now, in this second part I turn my focus to Midjourney, a tool that has recently undergone a few pivotal changes that I think are going to have a big impact on the fundamental design culture of practices. That means that they are worthy of critical reflection as practices begin testing and adopting:

1) Retexture – Reduces randomness and brings “control net” functionality to Midjourney (MJ). This means rather than starting with random form and composition, we give the model linework or 3D views to work from. Previously, despite the remarkable quality of image outputs, this was not possible in MJ.

2) Moodboards – Make it easy to very quickly “train your own style” with a small collection of image references. Previously we have had to train “LoRAs” in Stable Diffusion (SD) or Flux, taking many hours of preparation and testing. Moodboards provide a lower fidelity but much more convenient alternative.

3) Personal codes – Tailors your outputs to your taste profile using ‘Personalize’ (US spelling). You can train your own “–p” code by offering up hundreds of your own A/B test preferences within your account – you can then switch to your ‘taste’ profile extremely easily. In short, once you’ve told MJ what you like, it gets a whole lot better at giving it back to you each time.

A model that instantly knows your aesthetic preferences

Personal codes (or “Personalization” codes to be more precise) allow us to train MJ on our style preferences for different kinds of image material. To better understand the idea, in Figure 1 below you’ll see a clear example of running the same text prompt both with and without my “–p” code. For me there is no contest, I consistently massively prefer the images that have applied my –p code as compared to those that have not.

When enabled, Personalization substantially improves the average quality of your output, everything goes quickly from fairly generic ‘meh’ to ‘hey!’. It’s also now possible to develop a number of different personal preference codes for use in different settings. For example, one studio group or team may have a desire to develop a slightly different style code of preferences to another part of the studio, because they work in a different sector with different methods of communication.

Find this article plus many more in the July / August 2025 Edition of AEC Magazine

Subscribe FREE here

Midjourney vs Stable Diffusion / Flux

In the last 18 months, many heads have been turned by the potential of new tools like Stable Diffusion in architecture, because they have allowed us to train our own image styles, render sketches and gain increasingly configurable controls over image generation using Ai – and often without even making a 3D model. Flux, a new parallel opensource model ecosystem has taken the same methods and techniques from SD and added greater levels of quality.

We may marvel at what Ai makes possible in shorter time frames, but we should all be thinking – “great, let’s try to make a bit more profit this year” not “great let’s use this to undercut my competitor

But for ease of use, broad accessibility and consistency of output, the closed-source (and paid product) Midjourney is now firmly winning for most practices I speak to that are not strongly technologically minded.

Anecdotally, when I do Ai workshops, perhaps 10% of attendees really ‘get’ SD, whereas more like 75% immediately tend to click with Midjourney and I find that it appeals to the intuitive and more nuanced instincts of designers who like to discover design through an iterative and open-ended method of exploration.

While SD & Flux are potentially very low cost to use (if you run them locally and have the prerequisite GPUs) and offer massive flexibility of control, they are also much much harder to use effectively than MJ and more recently GPT-4o.

For a few months now Midjourney now sits within a slick web interface that is very intuitive to use and will produce top quality output with minimal stress and technical research.

Before we reflect on what this means for the overall culture of design in architectural practice going forwards, here are two notable observations to start with:

1) Practices who are willing to try their hand with diffusion models during feasibility or competition stage are beginning to find an edge. More than one recent conversation is suggesting that the use of diffusion models during competition stages has made a pivotal difference to recent bid processes and partially contributed to winning proposals.

2) I now see a growing interest from my developer client base, who want to go ahead and see vivid imagery even before they’ve engaged an architect or design team – they simply have an idea and want to go directly to seeing it visualised. In some cases, developers are looking to use Ai imagery to help dispose of sites, to quickly test alternative (visual) options to understand potential, or to secure new development contracts or funding.

Make of that what you will. I’m sure many architects will be cringing as they read that, but I think both observations are key signals of things to come for the industry whether it’s a shift you support or not. At the same time, I would say there is certainly a commercial opportunity there for architects if they’re willing to meet their clients on this level, adjust their standard methods of engagement and begin to think about exactly what value they bring in curating initial design concepts in an overtly transparent way at the inception stage of a project.

Text vs Image – where are people focused?

While I believe focusing on LLM adoption currently offers the most immediate and broadest benefits across practice and projects – the image realm is where most architects are spending their time when they jump into Generative Ai.

If you’re already modelling every detail and texture of your design and you want finite control, then you don’t use an Ai for visualisation, just continue to use CGI

Architects are fundamentally aesthetic creatures and so perhaps unsurprisingly they assume the image and modelling side of our work will be the most transformed over time. Therefore, I tend to find that architects often really want to lean into image model techniques above alternative Ai methods or Generative Design methods that may be available.

In the short term, image models are likely to be the most impactful for “storytelling” and in the initial briefing stages of projects where you’re not really sure what you think about a distinctive design approach, but you have a framework of visual and 3D ideas you want to play with.

Mapping diffusion techniques to problems

If you’re not sure what all of this means, see table below for a simple explanation of these techniques mapped to typical problems faced by designers looking to use Ai image models.

Changes with Midjourney v7

Midjourney recently launched its v7 model and it was met with relatively muted praise, probably because people were so blown away by the ground breaking potential of GPT-image-1 (an auto-regression model) that arrived just a month before.

This latest version of the MJ model was trained entirely from scratch so as a result it behaves differently to the familiar v6.1 model. I’m finding myself switching between v7 and 6.1 more regularly than with any previous model release.

One of the striking things about v7 is that you can only access the model when you have provided at least 200+ “image rating” preferences which points to an interesting new direction for more customised Ai experiences. Perhaps Midjourney has realised that the personalisation that is now possible in the platform is exactly what people want in an age of abundant imagery (increasingly created with Ai).

I for one, much prefer using a model that feels like it’s tuned just for me – more broadly, I suspect users want to feel like only they can produce the images they create and that they have a more distinctive style as a result. Leaning more into “Personalize” mode is helping with that and I like that MJ gating access to v7 behind the image ranking process.

I have achieved great results with the new model, but I find it harder to use and you do need to work differently with it. Here is some initial guidance on best use:

• v7 has a new function called ‘draft’ mode which produces low-res options very fast. I’m finding that to get the best results in this version you have to work in this manner, first starting with draft mode enabled and then enhancing to larger resolution versions directly from there. It’s almost like draft mode helps v7 work out the right composition from the prompt and then enhance mode helps to refine the resolution from there. If you try to go for full res v7 in one rendering step, you’ll probably be confused by the lower-par output.
  –
• Getting your “personalize” code is essential for accessing v7 and I’m finding my –p code only begins to work relatively effectively from about 1,000+ rankings, so set aside a couple of hours to train your preferences in.
  –
• You can now prompt with voice activation mode, which means having a conversation about the composition and image type you are looking for. As you speak v7 will start producing ideas in front of you.

Letting the model play

Image models improvise and this is their great benefit. They aren’t the same as CGI.

The biggest psychological hurdle that teams have to cross in the image realm is to understand that using Ai diffusion models is not like rendering in the way we’ve become accustomed to – it’s a different value proposition. If you’re already modelling every detail and texture of your design and you want finite control, then you don’t use an Ai for visualisation, just continue to use CGI.

However, if you can provide looser guidance with your own design linework before you’ve actually designed the fine detail, feeding inputs for the overall 3D form and imagery for textures and materials, then you are essentially allowing the model to play within those boundaries.

This means letting go of some control and seeing what the model comes back with – a step that can feel uncomfortable for many designers. When you let the model play within boundaries you set, you likely find striking results that change the way you’re thinking about the design that you’re working on. You may at times find yourself both repulsed and seduced in short order as you search around through one image to the next, searching for a response that lands in the way you had hoped.

A big shift that I’m seeing is that Midjourney is making “control net” type work and “style transfer” with images accessible to a much wider audience than would naturally be inclined to try out a very technical tool like SD.

I think that Midjourney’s decision to finally take the tool out of the very dodgy feeling Discord and launching a proper new and easy to use UI has really made the difference to practices. I still love to work with SD most of all, but I really see these ideas are beginning to land in MJ because it’s just so much easier to get a good result first time and it’s become really delightful to use.

Midjourney has a bit more work to do on its licence agreements (it is currently setup for single prosumers rather than enterprise) and privacy (they are training on your inputs). While you may immediately rule the tool out on this basis, consider – in most cases your inputs are primitive sketches or Enscape white card views, do you really mind if they are used for training and do they give away anything that would be considered privileged? With Stealth mode enabled (which you have to be on pro level for), your work can’t be viewed in public galleries. In order to get going with Midjourney in practice, you will need to allay all current business concerns, but with some basic guardrails in place for responsible use I am now seeing traction in practice.

Looking afresh at design culture

The use of “synthetic precedents” (i.e. images made purely with Ai already) is also now beginning to shape our critical thinking about design in early stages. Midjourney which has an exceptional ability to tell vivid first-person stories around projects, design themes and briefs, with seductive landscapes, materials and atmosphere. From the evidence I’ve seen so far, the images very much appeal to clients.

We are now starting to see Ai imagery be pinned up on the wall for studio crits and therefore I think we need to consider the impact of Ai on the overall design culture of the profession.

If we put Ai aside for a moment – in architectural practice, I think it’s a good idea to regularly reflect on your current studio design culture by considering first;

• Are we actually setting enough time aside to talk about design or is it all happening ad-hoc at peoples’ desks or online?
  –
• Do we share a common design method and language that we all understand implicitly?
  –
• Are we progressing and getting better with each project?
  –
• Are all team members contributing to the dialogue or waiting passively to be told what to do by a director with a napkin sketch?
  –
• Are we reverting to our comfort zone and just repeating tired ideas? • Are we using the right tools and mediums to explore each concept?

When people express frustration with design culture, they often refer specifically to some aspect of technological “misuse”, for example;

1. “People are using SketchUp too much. They’re not drawing plans anymore”
   –
2. “We are modelling everything in Revit at Stage 3, and no one is thinking about interface detailing”
   –
3. “All I’m seeing is Enscape design options wall to wall. I’m struggling to engage”
   –
4. “I think we might be relying too heavily on Pinterest boards to think about materials”, or maybe;
   –
5. “I can’t read these computer images. I need a model to make a decision”.

… all things I’ve heard said in practice.

Design culture has changed a lot since I entered the profession, and I have found that our relationship with the broad category of “images” in general has changed dramatically over time. Perhaps this is because we used to have to do all our design research collecting monograph books and by visiting actual buildings to see them, whereas now I probably keep up to date on design in places like Dezeen or Arch Daily – platforms that specifically glorify the single image icon and that jump frenetically across scale, style and geography.

One of the great benefits of my role with Arka Works is that I get to visit so many design studios (more than 70 since I began) and I’m seeing so many different ways of working and a full range of opinions about Ai.

I recently heard from a practice leader who said that in their practice, pinning up the work of a deceased (and great) architect was okay, because if it’s still around it must have stood the test of time and also presumably it’s beyond the “life plus 70 year Intellectual Property rule” – but in this practice the random pinning up of images was not endorsed.

Other practice leads have expressed to me that they consider all design work to be somehow derivative and inspired by things we observe – in other words – it couldn’t exist without designers ruminating on shared ideas, being enamoured of another architects’ work, or just plain using peoples’ design material as a crib sheet. In these practices, you can pin up whatever you like – if it helps to move the conversation forward.

Some practices have specific rules about design culture – they may require a pin up on a schedule with a specific scope of materials – you might not be allowed to show certain kinds of project imagery, without a corresponding plan, for example (and therefore holistic understanding of the design concepts). Maybe you insist on models or prefer no renders.

I think those are very niche cases. More often I see images and references simply being used as a shortcut for words and I also think we are a more image-obsessed profession than ever. In my own experience so far, I think these new Ai image tools are extremely powerful and need to be wielded with care, but they absolutely can be part of the design culture and have a place in the design review, if adopted with good judgement.

This is an important caveat. The need for critical judgment at every step is absolutely essential and made all the more challenging by how extraordinary the outputs can be – we will be easily seduced into thinking “yes that’s what I meant”, or “that’s not exactly what I meant, but it’ll do”, or worse “that’s not at all what I meant, but the Ai has probably done a better job anyway – may as well just use Ai every time from now on.”

Pinterestification

This shortening of attention spans is a problem we face in all realms of popular culture, as we become more digital every day. We worry that quality will suffer as people’s attention spans cause more laziness around design idea creation and testing – which would cause a broad dumbing down effect. This has been referred to as the ‘idiot trap’, where we rely so heavily on subcontracting thinking to various Ais, that we forget how to think from first principles.

You might think as a reaction – “well let’s just not bother using Ai altogether” and I think that’s a valid critique if you believe that architectural creativity is a wholly artisanal and necessarily human crafted process.

Probably the practices that feel that way just aren’t calling me to talk about Ai, but you would be surprised by the kind of ‘artisanal’ practices who are extremely interested in adopting Ai image techniques because rather than seeing them as a threat, they just see it as another way of exercising and exploring their vision with creativity.

Perhaps you have observed something I call “Pinterestification” happening in your studio?

I describe this as the algorithmic convergence of taste around common tropes and norms. If you pick a chair you like in Pinterest, it will immediately start nudging you in the direction of living room furniture, kitchen cabinets and bathroom tiles that you also just happen to love.

They all go so well on the mood board…

It’s almost like the algorithm has aggregated the collective design preferences of millions of tastemakers and packaged it up onto a website with convenient links to buy all the products we hanker after and that’s because it has.

Pinterest is widely used by designers and now heavily relied upon. The company has mapped our clicks; they know what goes together, what we like, what other people with similar taste like – and the incentives of ever greater attention mean that it’s never in Pinterest’s best interest to challenge you. Instead, Pinterest is the infinite design ice cream parlour that always serves your favourite flavour; it’s hard to stop yourself going back every time.

Learning about design

I’ve recently heard that some universities require full disclosure of any Ai use and that in other cases it can actually lead to disciplinary action against the student. The academic world is grappling with these new tools just as practice is, but with additional concerns about how students develop fundamental design thinking skills – so what is their worry?

The tech writer Paul Graham once said “writing IS thinking” and I tend to agree. Sure, you could have an LLM come up with a stock essay response – but the act of actually thinking by writing down your words and editing yourself to find out where you land IS the whole point of it. Writing is needed to create new ideas in the world and to solve difficult problems. The concern from universities therefore is that if we stop writing, we will stop thinking.

For architects, sketching IS our means of design thinking – it’s consistently the most effective method of ‘problem abstraction’ that we have. If I think back to most skilful design mentors I had in my early career, they were ALL expert draftspeople.

That’s because they came up with the drawing board and what that meant was they could distil many problems quickly and draw a single thread through things to find a solution, in the form of an erudite sketch. They drew sparingly, putting just the right amount of information in all the right places and knowing when to explore different levels of detail – because when you’re drawing by hand, you have to be efficient – you have to solve problems as you go.

Someone recently said to me that the less time the profession has spent drawing by hand (by using CAD, Revit, or Ai), the less that architects have earned overall. This is indeed a bit of a mind puzzle, and the crude problem is that when a more efficient technology exists, we are forced into adoption because we have to compete for work, whether it’s in our long term interests or not – it’s a Catch 22.

But this observation contains a signal too; that immaculate CAD lines do a different job from a sketching or hand drawing. The sketch is the truly high-value solution, the CAD drawing is the prosaic instructions for how to realise it.

I worry that “the idiot trap” for architects would be losing the fundamental skills of abstract reasoning that combines spatial, material, engineering and cultural realms and in doing so failing to recognise this core value as being the thing that the client is actually paying for (i.e. they are paying for the solution, not the instructions).

Clients hire us because we can see complete design solutions and find value where others can’t and because we can navigate the socio-political realm of planning and construction in real life – places where human diplomacy and empathy are paramount.

They don’t hire us to simply ‘spend our time producing package information’ – that is a by-product and in recent years we’ve failed to make this argument effectively. We shouldn’t be charging “by the time needed to do the drawing”, we should be charging “by the value” of the building.

So as we consider things being done more quickly with Ai image models, we need to build consensus that we won’t dispense with the sketching and craft of our work. We have to avoid the risk of simply doing something faster and giving the saving straight back to the market in the form of reduced prices and undercutting. We may marvel at what Ai makes possible in shorter time frames, but we should all be thinking – “great, let’s try to make a bit more profit this year” not “great let’s use this to undercut my competitor”.

Conclusion: judicious use

There is a popular quote (by Joanna Maciejewska) that has become a meme online:

I want Ai to do my laundry and dishes, so that I can do art and writing, not for Ai to do my art and writing so that I can do my laundry and dishes

If we translate that into our professional lives, for architects that would probably mean having Ai assisting us with things like regulatory compliance and auditing, not making design images for us.

Counter-intuitively Ai is realising value for practices in the very areas we would previously have considered the most difficult to automate: design optioneering, testing and conceptual image generation.

When architects reach for a tool like Midjourney, we need to be aware that these methods go right to the core of our value and purpose as designers. More so, that Ai imagery forces us to question our existing culture of design and methods of critique.

Unless we expressly dissuade our teams from using tools like Midjourney (which would be a valid position), anyone experimenting with it will now find it to be so effective that it will inevitably percolate into our design processes in ways that we don’t control, or enjoy.

Rather than allow these ad-hoc methods to creep up on us in design reviews unannounced and uncontrolled, a better approach is to consider first what would be an ‘aligned’ mode of adoption within our design processes – one that fits with the core culture and mission of the practice and then to make more deliberate use of it with endorsed design processes that create repeatable outputs that we really appreciate.

If you have a particularly craft-based design method, you could consider how that mode of thinking would be applied that to your use of Ai? Can you take a particularly experimental view of adoption that aligns with your specific priorities? Think Archigram with the photocopier.

We also need to question when something is pinned up on a wall alongside other material, if it can be judged objectively on its merits and relevance to the project, and if it stands up to this test – does it really matter to us how it was made? If I tell you it’s “Ai generated” does it reduce its perceived value?

I find that experimentation with image models is best led by the design leaders in practice because they are the “tastemakers” of practice and usually create the permission structures around design. Image models are often mistakenly categorised as technical phenomena and while they require some knowledge and skill, they are actually far more integral to the aesthetic, conceptual and creative aspects of our work.

To get a picture of what “aligned adoption of Ai” would mean for your practice, it should feel like you’re turning up the volume on the particular areas of practice that you already excel at, or conversely to mitigate aspects of practice that you feel acutely weaker in.

Put another way – Ai should be used to either reinforce whatever your specialist niche is or to help you remedy your perceived vulnerabilities. I particularly like the idea of leaning into our specialisms because it will make our deployment of Ai much more experimental, more bespoke and more differentiated in practice.

When I am applying Ai in practice, I don’t see depressed and disempowered architects – I am reassured to find that the most effective people at writing bids with Ai, also tend to be some of the best bid writers. The people who end up becoming the most experimental and effective at producing good design images with Ai image models, also tend to be great designers too and this trend goes on in all areas where I see Ai being used judiciously, so far – without exception.

The “judicious use” part is most important because only a practitioner who really knows their craft can apply these ideas in ways that actually explore new avenues for design and realise true value in project settings. If you feel that description matches you – then you should be getting involved and having an opinion about it. In the Ai world this is referred to as keeping the “human-in-the-loop” but we could think of it as the “architect-in-the-loop” continuing to curate decisions, steer things away from creative cul de sacs and to more effectively drive design.

Recommended viewing

Keir Regan-Alexander is director at Arka Works, a creative consultancy specialising in the Built Environment and the application of AI in architecture. At NXT BLD 2025 he explored how to deploy Ai in practice.

CLICK HERE to watch the whole presentation free on-demand

Watch the teaser below

The post Ai & design culture (part 2) appeared first on AEC Magazine.

AI is transforming industries across the globe, and architectural visualisation (let’s call it ‘Arch Viz’) is no exception. Today, generative AI tools play an increasingly important role in an arch viz workflow, empowering creativity and efficiency while maintaining the precision and quality expected in high-end visuals.

In this piece I will share my experience and best practices for how AI is actively shaping arch viz by enhancing workflow efficiency, empowering creativity, and setting new industry standards.

Streamlining workflows with AI

AI, we dare say, has proven not to be a bubble or a simple trend, but a proper productivity driver and booster of creativity. Our team at Panoptikon and others in the industry leverage generative AI tools to the maximum to streamline processes and deliver higher-quality results.

Tools like Stable Diffusion, Midjourney and Krea.ai transform initial design ideas or sketches into refined visual concepts. Platforms like Runway, Sora, Kling, Hailuo or Luma can do the same for video.

With these platforms, designers can enter descriptive prompts or reference images, generating early-stage images or videos that help define a project’s look and feel without lengthy production times.

This capability is especially valuable for client pitches and brainstorming sessions, where generating multiple iterations is critical. Animating a still image is possible with the tools above just by entering a descriptive prompt, or by manipulating the camera in Runway.ml.

Sometimes, clients find themselves under pressure due to tight deadlines or external factors, while studios may also be fully booked or working within constrained timelines. To address these challenges, AI offers a solution for generating quick concept images and mood boards, which can speed up the initial stages of the visualisation process.

In these situations, AI tools provide a valuable shortcut by creating reference images that capture the mood, style, and thematic direction for the project. These AI-generated visuals serve as preliminary guides for client discussions, establishing a strong visual foundation without requiring extensive manual design work upfront.

Although these initial images aren’t typically production-ready, they enable both the client and visualisation team to align quickly on the project’s direction.

Once the visual direction is confirmed, the team shifts to standard production techniques to create the final, high-resolution images that would accurately showcase the full range of technical specifications that outline the design. While AI expedites the initial phase, the final output meets the high-quality standards expected for client presentations.

Dynamic visualisation

For projects that require multiple lighting or seasonal scenarios, Stable Diffusion, LookX or Project Dream allow arch viz artists to produce adaptable visuals by quickly applying lighting changes (morning, afternoon, evening) or weather effects (sunny, cloudy, rainy).

Additionally, AI’s ability to simulate seasonal shifts allows us to show a park, for example, lush and green in summer, warm-toned in autumn, and snow-covered in winter. These adjustments make client presentations more immersive and relatable.

Adding realism through texture and detail

AI tools can also enhance the realism of 3D renders. By specifying material qualities through prompts or reference images in Stable Diffusion, Magnific, and Krea, materials like wood, concrete, and stone, or greenery and people are quickly improved.

The tools add nuanced details like weathering to any surface or generate intricate enhancements that may be challenging to achieve through traditional rendering alone. The visuals become more engaging and give clients a richer sense of the project’s authenticity and realistic quality.

This step may not replace traditional rendering or post-production but serves as a complementary process to the overall aesthetic, bringing the image closer to the level of photorealism clients expect.

Bridging efficiency and artistic quality

While AI provides speed and efficiency, the reliance on human expertise for technical precision is mandatory. AI handles repetitive tasks, but designers need to review and refine each output so that the visuals meet the exact technical specifications provided by each project’s design brief.

Challenges and considerations

It is essential to approach the use of AI with awareness of its limitations and ethical considerations.

Maintaining quality and consistency: AI-generated images sometimes contain inconsistencies or unrealistic elements, especially in complex scenes. These outputs require human refinement to align with the project’s vision so that the result is accurate and credible.

Ethical concerns around originality: There’s an ongoing debate about originality in AI-generated designs, as many AI outputs are based on training data from existing works. We prioritise using AI as a support tool rather than a substitute for human creativity, as integrity is among our core values.

Future outlook: innovation with a human touch: Looking toward and past 2025, AI’s role in arch viz is likely to expand further – supporting, rather than replacing, human creativity. AI will increasingly handle technical hurdles, allowing designers to focus on higher-level creative tasks.

AI advancements in real-time rendering are another hot topic, expected to enable more immersive, interactive tours, while predictive AI models may suggest design elements based on client preferences and environmental data, helping studios anticipate client needs.

AI’s role in arch viz goes beyond productivity gains. It’s a catalyst for expanding creative possibilities, enabling responsive design, and enhancing client experiences. With careful integration and human oversight, AI empowers arch viz studios – us included – to push the boundaries of what’s possible while, at the same time, preserving the artistry and precision that define high-quality visualisation work.

About the author

Tudor Vasiliu is an architect turned architectural visualiser and the founder of Panoptikon, an award-winning high-end architectural visualisation studio serving clients globally. With over 18 years of experience, Tudor and his team help the world’s top architects, designers, and property developers realise their vision through high-quality 3D renders, films, animations, and virtual experiences. Tudor has been honoured with the CGarchitect 3D Awards 2019 – Best Architectural Image, and has led industry panels and speaking engagements at industry events internationally including the D2 Vienna Conference, State of Art Academy Days, Venice, Italy and Inbetweenness, Aveiro, Portugal – among others.

Main image caption: Rendering by Panoptikon for ‘The Point’, Salt Lake City, Utah. Client: Arcadis (Credit: Courtesy of Panoptikon, 2025)

The post AI and the future of arch viz appeared first on AEC Magazine.

One can’t help but be impressed by the current capabilities of many AI tools. Standout examples include Gemini from Google, ChatGPT from OpenAI, Musk’s Grok, Meta AI and now the new Chinese wunderkind, DeepSeek.

Many billions of dollars are being invested in hardware. Development teams around the globe are racing to create an artificial general intelligence, or AGI, to rival (and perhaps someday, surpass) human intelligence.

In the AEC sector, R&D teams within all of the major software vendors are hard at work on identifying uses for AI in this industry. And we’re seeing the emergence of start-ups claiming AI capabilities and hoping to beat the incumbents at their own game.

However, beyond the integration of ChatGPT frontends, or yet another AI renderer, we have yet to feel the promised power of AI in our everyday BIM tools.

The rendering race

The first and most notable application area for AI in the field of AEC has been rendering, with the likes of Midjourney, Stable Diffusion, Dall-E, Adobe Firefly and Sketch2Render all capturing the imaginations of architects.

While the price of admission has been low, challenges have included the need to specify words to describe an image (there is, it seems, a whole art to writing prompting strategies) and then somehow remain in control of its AI generation through subsequent iterations.

In this area, we’ve seen the use of LoRAs (Low Rank Adaptations), which implement trained concepts/styles and can ‘adapt’ to a base Stable Diffusion model, and ControlNet, which empowers precise and structural control to deliver impressive results in the right hands.

For those wishing to dig further, we recommend familiarising yourself with the amazing work of Ismail Seleit and his custom-trained LoRAs combined with ControlNet. For those who’d prefer not to dive so deep into the tech, SketchUp Diffusion, Veras, and AI Visualizer (for Archicad, Allplan and Vectorworks), have helped make AI rendering more consistent and likely to lead to repeatable results for the masses.

However, when it comes to AI ideation, at some point, architects would like to bring this into 3D – and there is no obvious way to do this. This work requires real skill, interpreting a 2D image into a Rhino model or Grasshopper script, as demonstrated by the work of Tim Fu at Studio Tim Fu.

It’s possible that AI could be used to auto-generate a 3D mesh from an AI conceptual image, but this remains a challenge, given the nature of AI image generation. There are some tools out there which are making some progress, by analysing the image to extract depth and spatial information, but the resultant mesh tends to come out as one lump, or as a bunch of meshes, incoherent for use as a BIM model or for downstream use.

Back in 2022, we tried taking 2D photos and AI-generated renderings from Hassan Ragab into 3D using an application called Kaedim. But the results were pretty unusable, not least because at that time Kaedim had not been trained on architectural models and was more aimed at the games sector.

Of course, if you have multiple 2D images of a building, it is possible to recreate a model using photogrammetry and depth mapping.

AI in AEC – text to 3D

It’s possible that the idea of auto-generating models from 2D conceptual AI output will remain a dream. That said, there are now many applications coming online that aim to provide the AI generation of 3D models from text-based input.

The idea here is that you simply describe in words the 3D model you want to create – a chair, a vase, a car – and AI will do the rest. AI algorithms are currently being trained on vast datasets of 3D models, 2D images and material libraries.

While 3D geometry has mainly been expressed through meshes, there have been innovations in modelling geometry with the development of Neural Radiance Fields (NeRFs) and Gaussian splats, which represent colour and light at any point in space, enabling the creation of photorealistic 3D models with greater detail and accuracy.

Today, we are seeing a high number of firms bringing ‘text to 3D’ solutions to market. Adobe Substance 3D Modeler has a plug-in for Photoshop that can perform text-to-3D. Similarly, Autodesk demonstrated similar technology — Project Bernini — at Autodesk University 2024.

However, the AI-generated output of these tools seems to be fairly basic — usually symmetrical objects and more aimed towards creating content for games.

In fact, the bias towards games content generation can be seen in many offerings. These include Tripo, Kaedim, Google DreamFusion  and Luma AI Genie.

There are also several open source alternatives. These include Hunyuan3D-1, Nvidia’s Magic 3D and Edify.

AI in AEC – the Schleusner viewpoint

When AEC Magazine spoke to Greg Schleusner of HOK on the subject of text-to-3D, he highlighted D5 Render, which is now an incredibly popular rendering tool in many AEC firms.

The application comes with an array of AI tools, to create materials, texture maps and atmosphere match from images. It supports AI scaling and has incorporated Meshy’s text-to-AI generator for creating content in-scene.

That means architects could add in simple content, such as chairs, desks, sofas and so on — via simple text input during the arch viz process. The items can be placed in-scene on surfaces with intelligent precision and are easily edited. It’s content on demand, as long as you can describe that content well in text form.

Schleusner said that, from his experimentation, text-to-image or image-tovideo tools are getting better, and will eventually be quite useful — but that can be scary for people working in architecture firms. As an example, he suggested that someone could show a rendering of a chair within a scene, generated via text to AI. But it’s not a real chair, and it can’t be purchased, which might be problematic when it comes to work that will be shown to clients. So, while there is certainly potential in these types of generative tools, mixing fantasy with reality in this way doesn’t come problem-free.

It may be possible to mix the various model generation technologies. As Schleusner put it: “What I’d really like to be able to do is to scan or build a photogrammetric interior using a 360-degree camera for a client and then selectively replace and augment the proposed new interior with new content, perhaps AI-created.”

Gaussian splat technology is getting good enough for this, he continued, while SLAM laser scan data is never dense enough. “However, I can’t put a Gaussian splat model inside Revit. In fact, none of the common design tools support that emerging reality capture technology, beyond scanning. In truth, they barely support meshes well.”

AI in AEC – LLMs and AI agents

At the time of writing, DeepSeek has suddenly appeared like a meteor, seemingly out of nowhere, intent on ruining the business models of ChatGPT, Gemini and other providers of paid-for AI tools.

Schleusner was early into DeepSeek and has experimented with its script and code-writing capabilities, which he described as very impressive.

LLMs, like ChatGPT, can generate Python scripts to perform tasks in minutes, such as creating sample data, training machine learning models, and writing code to interact with 3D data.

Schleusner is finding that AI-generated code can accomplish these tasks relatively quickly and simply, without needing to write all the code from scratch himself.

“While the initial AI-generated code may not be perfect,” he explained, “the ability to further refine and customise the code is still valuable. DeepSeek is able to generate code that performs well, even on large or complex tasks.”

WIth AI, much of the expectation of customers centres on the addition of these new capabilities to existing design products. For instance, in the case of Forma, Autodesk claims the product uses machine learning for real-time analysis of sunlight, daylight, wind and microclimate.

However, if you listen to AI-proactive firms such as Microsoft, executives talk a lot about ‘AI agents’ and ‘operators’, built to assist firms and perform intelligent tasks on their behalf.

Microsoft CEO Satya Nadella is quoted as saying, “Humans and swarms of AI agents will be the next frontier.” Another of his big statements is that, “AI will replace all software and will end software as a service.” If true, this promises to turn the entire software industry on its head.

Today’s software as a service, or SaaS, systems are proprietary databases/silos with hard-coded business logic. In an AI agent world, these boundaries would no longer exist. Instead, firms will run a multitude of agents, all performing business tasks and gathering data from any company database, files, email or website. In effect, if it’s connected, an AI agent can access it.

At the moment, to access certain formatted data, you have to open a specific application and maybe have deep knowledge to perform a range of tasks. An AI agent might transcend these limitations to get the information it needs to make decisions, taking action and achieving business-specific goals.

AI agents could analyse vast amounts of data, such as a building designs, to predict structural integrity, immediately flag up if a BIM component causes a clash, and perhaps eventually generate architectural concepts. They might also be able to streamline project management by automating routine tasks and providing real-time insights for decision-making.

AI agents could analyse vast amounts of data, such as a building designs, to predict structural integrity, immediately flag up if a BIM component causes a clash, and perhaps eventually generate architectural concepts

The main problem is going to be data privacy, as AI agents require access to sensitive information in order to function effectively. Additionally, the transparency of AI decision-making processes remains a critical issue, particularly in high-stakes AEC projects where safety, compliance and accuracy are paramount.

On the subject of AI agents, Schleusner said he has a very positive view of the potential for their application in architecture, especially in the automation of repetitive tasks. During our chat, he demonstrated how a simple AI agent might automate the process of generating something as simple as an expense report, extracting relevant information, both handwritten and printed from receipts.

He has also experimented by creating an AI agent for performing clash detection on two datasets, which contained only XYZ positions of object vertices. Without creating a model, the agent was able to identify if the objects were clashing or not. The files were never opened. This process could be running constantly in the background, as teams submitted components to a BIM model. AI agents could be a game-changer when it comes to simplifying data manipulation and automating repetitive tasks.

Another area where Schleusner feels that AI agents could be impactful is in the creation of customisable workflows, allowing practitioners to define the specific functions and data interactions they need in their business, rather than being limited by pre-built software interfaces and limited configuration workflows.

Most of today’s design and analysis tools have built-in limitations. Schleusner believes that AI agents could offer a more programmatic way to interact with data and automate key processes. As he explained, “There’s a big opportunity to orchestrate specialised agents which could work together, for example, with one agent generating building layouts and another checking for clashes. In our proprietary world with restrictive APIs, AI agents can have direct access and bypass the limits on getting at our data sources.”

Conclusion

For the foreseeable future, AEC professionals can rest assured that AI, in its current state, is not going to totally replace any key roles — but it will make firms more productive.

The potential for AI to automate design, modelling and documentation is currently overstated, but as the technology matures, it will become a solid assistant. And yes, at some point years hence, AI with hard-coded knowledge will be able to automate some new aspects of design, but I think many of us will be retired before that happens. However, there are benefits to be had now and firms should be experimenting with AI tools.

We are so used to the concept of programmes and applications that it’s kind of hard to digest the notion of AI agents and their impact. Those familiar with scripting are probably also constrained by the notion that the script runs in a single environment.

By contrast, AI agents work like ghosts, moving around connected business systems to gather, analyse, report, collaborate, prioritise, problem-solve and act continuously. The base level is a co-pilot that may work alongside a human performing tasks, all the way up to fully autonomous operation, uncovering data insights from complex systems that humans would have difficulty in identifying.

If the data security issues can be dealt with, firms may well end up with many strategic business AI agents running and performing small and large tasks, taking a lot of the donkey work from extracting value from company data, be that an Excel spreadsheet or a BIM model.

AI agents will be key IP tools for companies and will need management and monitoring. The first hurdle to overcome is realising that the nature of software, applications and data is going to change radically and in the not-too-distant future.

Main image: Stable Diffusion architectural images courtesy of James Gray. Image (left) generated with ModelMakerXL, a custom trained LoRA by Ismail Seleit. Follow Gray on LinkedIn

The post Artificial horizons: AI in AEC appeared first on AEC Magazine.

Stable Diffusion is a powerful text-to-image AI model that generates stunning photorealistic images based on textual descriptions. Its versatility, control and precision have made it a popular tool in industries such as architecture and product design.

One of its key benefits is its ability to enhance the conceptual design phase. Architects and product designers can quickly generate hundreds of images, allowing them to explore different design ideas and styles in a fraction of the time it would take to do manually.

This article is part of AEC Magazine’s 2025 Workstation Special report

Stable Diffusion relies on two main processes: inferencing and training. Most architects and designers will primarily engage with inferencing, the process of generating images from text prompts. This can be computationally demanding, requiring significant GPU power.

Training is even more resource intensive. It involves creating a custom diffusion model, which can be tailored to match a specific architectural style, client preference, product type, or brand. Training is often handled by a single expert within a firm.

There are several architecture-specific tools built on top of Stable Diffusion or other AI models, which run in a browser or handle the computation in the cloud. Examples include AI Visualizer (for Archicad, SketchUp, and Vectorworks), Veras, LookX AI, and CrXaI AI Image Generator. While these tools simplify access to the technology, and there are many different ways to run vanilla Stable Diffusion in the cloud, many architects still prefer to keep things local.

Running Stable Diffusion on a workstation offers more options for customisation, guarantees control over sensitive IP, and can turn out cheaper in the long run. Furthermore, if your team already uses real-time viz software, the chances are they already have a GPU powerful enough to handle Stable Diffusion’s computational demands.

While computational power is essential for Stable Diffusion, GPU memory plays an equally important role. Memory usage in Stable Diffusion is impacted by several factors, including:

• Resolution: higher res images (e.g. 1,024 x 1,024 pixels) demand more memory compared to lower res (e.g. 512 x 512).
• Batch size: Generating more images in parallel can decrease time per image, but uses more memory.
• Version: Newer versions of Stable Diffusion (e.g. SDXL) use more memory.
• Control: Using tools to enhance the model’s functionality, such as LoRAs for fine tuning or ControlNet for additional inputs, can add to the memory footprint.

For inferencing to be most efficient, the entire model must fit into GPU memory. When GPU memory becomes full, operations may still run, but at significantly reduced speeds as the GPU must then borrow from the workstation’s system memory, over the PCIe bus.

This is where professional GPUs can benefit some workflows, as they typically have more memory than consumer GPUs. For instance, the Nvidia RTX A4000 professional GPU is roughly the equivalent of the Nvidia GeForce RTX 3070, but it comes with 16 GB of GPU memory compared to 8 GB on the RTX 3070.

Inferencing performance

To evaluate GPU performance for Stable Diffusion inferencing, we used the UL Procyon AI Image Generation Benchmark. The benchmark supports multiple inference engines, including Intel OpenVino, Nvidia TensorRT, and ONNX runtime with DirectML. For this article, we focused on Nvidia professional GPUs and the Nvidia TensorRT engine.

This benchmark includes two tests utilising different versions of the Stable Diffusion model — Stable Diffusion 1.5, which generates images at 512 x 512 resolution and Stable Diffusion XL (SDXL), which generates images at 1,024 x 1,024. The SD 1.5 test uses 4.6 GB of GPU memory, while the SDXL test uses 9.8 GB.

In both tests, the UL Procyon benchmark generates a set of 16 images, divided into batches. SD 1.5 uses a batch size of 4, while SDXL uses a batch size of 1. A higher benchmark score indicates better GPU performance. To provide more insight into real-world performance, the benchmark also reports the average image generation speed, measured in seconds per image. All results can be seen in the charts below.

Key takeaways

It’s no surprise that performance goes up as you move up the range of GPUs, although there are diminishing returns at the higher-end. In the SD 1.5 test, even the Nvidia RTX A1000 delivers an image every 11.7 secs, which some will find acceptable.

The Nvidia RTX 4000 Ada Generation GPU looks to be a solid choice for Stable Diffusion, especially as it comes with 20 GB of GPU memory. The Nvidia RTX 6000 Ada Generation (48 GB) is around 2.3 times faster, but considering it costs almost six times more (£6,300 vs £1,066) it will be hard to justify on those performance metrics alone.

The real benefits of the higher end cards are most likely to be found in workflows where you can exploit the extra memory. This includes handling larger batch sizes, running more complex models, and, of course, speeding up training.

Perhaps the most revealing test result comes from SDXL, as it shows what can happen when you run out of GPU memory. The RTX A1000 still delivers results, but its performance slows drastically. Although it’s just 2 GB short of the 10 GB needed for the test, it takes a staggering 13 minutes to generate a single image — 70 times slower than the RTX 6000 Ada.

Of course, AI image generation technology is moving at an incredible pace. Tools including Flux, Runway and Sora can even be used to generate video, which demands even more from the GPU. When considering what GPU to buy now, it’s essential to plan for the future.

AI diffusion models – a guide for AEC professionals

From the rapid generation of high-quality visualisations to process optimisation, diffusion models are having a huge impact in AEC. In this AEC Magazine article, Nvidia’s Sama Bali explains how this powerful generative AI technology works, how it can be applied to different workflows, and how AEC firms can get on board.

Main image: Image generated with ModelMakerXL, a custom trained LoRA by Ismail Seleit. All Stable Diffusion
architectural images courtesy of James Gray.

This article is part of AEC Magazine’s 2025 Workstation Special report

Subscribe here

• Features
  • AMD Ryzen AI Max Pro: the integrated GPU comes of age
  • AMD Ryzen 9000 vs Intel Core Ultra 200S
  • Workstations for arch viz
  • The AI enigma – challenges for workstations
  • Nvidia RTX GPUs for Stable Diffusion
  • Z by HP Boost: GPUs on demand
  • Workstations for reality modelling

  Reviews

  • HP ZBook Firefly 14 G11 A mobile workstation
  • Lenovo ThinkPad P14s (AMD) mobile workstation
  • Scan 3XS GWP-A1-C24 and GWP-A1-R32 desktop workstations
  • Boxx Apexx A3 desktop workstation
  • Armari Magnetar MM16R9 desktop workstation
  • Comino Grando workstation RM

  News

  • Reshuffle spells end for Dell Precision workstation brand
  • Lenovo powers IMSCAD workstation as a service (WaaS)
  • Nvidia unveils ‘Blackwell’ RTX GPUs
  • HP to launch 18-inch mobile workstation
  • Nvidia reveals AI workstation
  • HP bets big on AMD Ryzen AI Max PRO processor
  • Intel Core Ultra 200HX and 200H processors launch
  • AMD Ryzen AI Max PRO ‘Strix Halo’ processor launches

  ---

The post Nvidia RTX GPUs for Stable Diffusion appeared first on AEC Magazine.

Since the introduction of generative AI, large language models (LLMs) like GPT-4 have been at the forefront, renowned for their versatility in natural language processing, machine translation, and content creation. Alongside these, image generators such as OpenAI’s DALL-E, Google’s Imagen, Midjourney and Stability AI’s Stable Diffusion, are changing the way architects, engineers, and construction professionals visualise and design projects, enabling rapid prototyping, enhanced creativity, and more efficient workflows.

At their core, diffusion models possess a distinctive capability. They can generate high-quality data from prompts by progressively adding and removing noise from a dataset.

Training diffusion models is done by adding noise to millions of images over many iterations and rewarding the model when it recreates the image in the reverse process. Once trained, the model is ready for inference whereby a user is able to generate realistic data, such as images, text, video, audio or 3D models.

Why noise? It helps diffusion models mimic random changes, understand the data, prevent overfitting, and ensure smooth transformations.

Imagine you have a sketch of a building design. You start adding random noise to it, making it look more and more like a messy scribble. This is the forward process. The reverse process is like cleaning up that messy scribble step by step until you get back to a detailed and clear architectural rendering.

The model learns how to do this cleaning process so well that it can start with random noise and end up generating a completely new, realistic building design. With this innovative approach diffusion models can produce remarkably accurate and detailed outputs, making them a powerful tool.

Diffusion models have a reputation for being difficult to control due to the way they learn, interpret, and produce visuals. However, ControlNets, a group of neural networks trained on specific tasks, can enhance the base model’s capabilities. Architects can exert precise structural and visual control over the generation process by providing references.

For example, Sketch ControlNet can transform an architectural drawing into a fully realised render.

Multiple ControlNets can be combined together for additional control. For instance, a Sketch ControlNet can be paired with an adaptor, which can incorporate a reference image to apply specific colours and styles to the design.

ControlNets are highly effective as they can process various types of information, empowering architects and designers with new ways to manage their designs and communicate ideas with clients.

Leveraging Nvidia accelerated compute capabilities further enhances the performance of diffusion models. Nvidia-optimised models, such as the SDXL Turbo and LCM-LoRA, offer state-of-the-art performance with real-time image generation capabilities. These models significantly improve inference speed and reduce latency, enabling the production of up to four images per second–drastically reducing the time required for high-resolution image generation.

Diffusion models offer several specific benefits to the AEC sector, enhancing various aspects of design, visualisation, and project management:

High-quality visualisations

Diffusion models can generate photorealistic images and videos from simple sketches, textual descriptions, or a combination. This capability is invaluable for creating detailed architectural renderings and visualisations, helping decision-makers understand and visualise proposed projects.

Daylighting and energy efficiency

Diffusion models can generate daylighting maps and analyse the impact of natural light on building designs. This helps optimise window placements and other design elements to enhance indoor daylighting and energy efficiency, ensuring that buildings are comfortable and sustainable.

Rapid prototyping

By automating the generation of design alternatives and visualisations, including materials, or object positioning, diffusion models can significantly speed up the design process. Architects and engineers can explore more design options faster, leading to more innovative and optimised solutions.

Cost savings and process optimisation

Diffusion models enable the customisation of BIM policies to suit the needs of specific regions and projects. By ensuring that resources are directed to the areas of greatest need, resource allocation is improved. This flexibility makes sure that ‌policies are tailored to the unique requirements of different regions and projects, leading to reduced project costs and improved overall efficiency.

Use, customise, or build your diffusion models

Organisations can leverage diffusion models in multiple ways. They can use pretrained models as-is, customise them for specific needs, or build new models from scratch and harness their full potential by tailoring them to a user’s unique requirements.

Pretrained models are deployable immediately, reducing the time to market and minimising initial investment. Customising pretrained models enables the integration of domain-specific data, improving accuracy and relevance for particular applications. Developing models from scratch, although resource-intensive, enables the creation of highly specialised solutions that can address unique challenges and provide a competitive edge.

Consider diffusion models in the AEC industry like architecting a house. Using pretrained models is similar to using standard prefabricated homes—they’re ready to use, saving time and initial costs. Customising pretrained models is like modifying standard off-the-shelf house plans to fit specific requirements, making sure the design meets particular needs and preferences. Building models from scratch is similar to creating entirely new blueprints from the ground up. This approach offers the most flexibility and customisation but requires significant expertise, time, and resources.

Each method has advantages and disadvantages, enabling organisations to select the most suitable approach according to their project objectives and available resources.

Pretrained models for quick deployment

For many organisations, the quickest way to benefit from diffusion models is to use pretrained models. Available through the Nvidia API catalog, these models are optimised for high performance and can be deployed directly into applications.

Nvidia NIM offers a streamlined and efficient way for organisations to deploy diffusion models, enabling the generation of high-resolution, realistic images from text prompts. With prebuilt containers, organisations can quickly set up and run diffusion models on Nvidia accelerated infrastructure (available from Nvidia workstations, data centres, cloud services partners, and private on-prem servers).

This approach simplifies the deployment process and maximises performance, enabling businesses to focus on building innovative generative AI workflows without the complexities of model development and optimisation.

Developers can experience and experiment with Nvidia-hosted NIMs at no charge.

Members of the Nvidia Developer Program can access NIM for free for research, development, and testing on their preferred infrastructure.

Enterprises can deploy AI applications in production with NIM through the Nvidia AI Enterprise software platform.

Customising diffusion models

Customising diffusion models can improve the relevance, accuracy, and performance of diffusion models for AEC organisations. It also enables organisations to include their own knowledge and industry-specific terms, and to address specific challenges.

Fine-tuning involves taking a pretrained model and adjusting its parameters using a smaller, domain-specific dataset to better align with the specific needs and nuances of the organisation. This tailored approach improves the quality and utility of the generated content and offers scalability and flexibility. Organisations can adapt the models as their needs evolve.

For firms wanting a user-friendly path to start customising diffusion models, Nvidia AI Workbench offers a streamlined environment that lets data scientists and developers get up and running quickly with generative AI. With AI Workbench users can get started with pre-configured projects that are adaptable to different data and use cases. It’s ideal for quick, iterative development and local testing.

Example projects, such as fine-tuning diffusion models, can be modified to support things like generating architectural renderings. Furthermore, this flexibility extends to ‌ supported infrastructure. Users can start locally on Nvidia RTX-powered AI Workstations and scale to virtually anywhere—data centre or cloud—in just a few clicks. For more details on how to customise diffusion models, explore the GitHub project.

Another lightweight training technique used for fine-tuning diffusion models is Low-Rank Adaptation or LoRA. LoRA models are ideal for architectural firms due to their small size. They can be managed and trained on local workstations without extensive cloud resources.

Check out how you can seamlessly deploy and scale multiple LoRA adapters with Nvidia NIM.

For advanced customisation and high-performance training, Nvidia NeMo offers a comprehensive, scalable, and cloud-native platform. NeMo offers a choice of customisation techniques and is optimised for at-scale inference of diffusion models, with multi-GPU and multi-node configurations.

The DRaFT+ algorithm, integrated into the NeMo framework, enhances the fine-tuning of diffusion models and makes sure that the model produces diverse and high-quality outputs aligned with specific project requirements. For more technical details and to access the DRaFT+ algorithm, visit the NeMo-Aligner library on GitHub.

Nvidia Launchpad provides a free hands-on lab environment where AEC professionals can learn to fine-tune diffusion models with custom images and optimise them for specific tasks, such as generating high-quality architectural renderings or visualising construction projects.

Building diffusion models that match your style

Now that we’ve covered pretrained and customised models, let’s build diffusion models from scratch. Investing in custom diffusion models allows AEC organisations to harness the full potential of AI, leading to more efficient, accurate, and innovative project outcomes.

For instance, an architectural firm might build their own diffusion model to generate design concepts that align with their specific architectural style and client preferences, while a construction company could develop a model to optimise resource allocation and project scheduling.

One example of this approach is the work of Heatherwick Studio, a design firm based in London. They’ve been using AI in their design process. The studio is known for its innovative projects around the world, including Google’s headquarters in London and California, Africa’s first museum of contemporary African art in Cape Town, and a new district in Tokyo. Heatherwick Studio has been developing tools that use their data to streamline design processes, rendering, and data access.

“At the studio, we not only believe in the transformational power of AI to improve the industry but are actively developing and deploying in-house custom diffusion models in our everyday work,” said Pablo Zamorano, head of Geometry and Computational Design at Heatherwick studio.

“We have developed a web-based tool that enables quick design provocations, fast rendering, and image editing as well as a tool that allows for tailored knowledge search from within our BIM tools. These tools empower the work of our designers and visualisers and are now well established.”

Creating custom diffusion models with Nvidia

NeMo provides a powerful framework that provides components for building and training custom diffusion models on-premises, across all leading cloud service providers, or in Nvidia DGX Cloud. It includes a suite of customisation techniques from prompt learning to parameter-efficient fine-tuning (PEFT), making it ideal for AEC customers who need to generate high-quality architectural renderings and optimise construction visualisations efficiently.

Alternatively, Nvidia Picasso is an AI foundry leveraged by asset marketplace companies to build and deploy cutting-edge generative AI models with APIs for commercially safe visual content.

Built on Picasso, generative AI services by Getty Images for image generation and Shutterstock for 3D generation, create commercially safe visual media from text or image. AEC organisations can fine-tune their choice of Picasso-powered models to create custom diffusion models that generate images from text prompts or sketches in different styles. Picasso supports end-to-end AI model development, from data preparation and model training to model fine-tuning and deployment, making it an ideal solution for developing custom generative AI services.

Responsible innovation with diffusion models

Using AI models involves several critical steps, including data collection, preprocessing, algorithm selection, training, and evaluation. Each of these steps requires careful consideration to make sure the model performs well and meets the specific needs of the project.

However, it’s equally important to integrate responsible AI practices throughout this process. Generative AI models, despite their impressive capabilities, are susceptible to biases, security vulnerabilities, and unintended consequences. Without proper safeguards, these models can produce outputs that reinforce harmful stereotypes, discriminate against certain demographics, or contain security flaws.

Additionally, protecting the security of diffusion models is crucial for generative AI-powered applications. Nvidia introduced accelerated Confidential Computing, a groundbreaking security feature that mitigates threats while providing access to the unprecedented acceleration of Nvidia H100 Tensor Core GPUs for AI workloads. This feature makes sure that sensitive data remains secure and protected, even during processing.

Get started

Generative AI, particularly diffusion models, is revolutionising the AEC industry by enabling the creation of photorealistic renderings and innovative designs from simple sketches or textual descriptions.

To get started, AEC firms should prioritise data collection and management, identify processes that can benefit from automation, and adopt a phased approach to implementation. The Nvidia training program helps organisations train their workforce on the latest technology and bridge the skills gap by offering comprehensive technical hands-on workshops and courses.

The post AI diffusion models – a guide for AEC professionals appeared first on AEC Magazine.

In the past, individual brands under the Nemetschek umbrella (of which there are 13 in total, including Graphisoft, Allplan, Bluebeam, Vectorworks and Solibri) were largely left to themselves when it came to developing new solutions and exploring new technologies.

That situation has changed over the past five years, with an increasing focus on sharing technologies between brands, combining those brands and developing workflows that intersect them.

In response to growing interest in artificial intelligence (AI), Nemetschek created an AI Innovation Hub in May 2024, with the goal of driving AI initiatives across its entire portfolio and involving partners and customers in that work.

The stated intention is not just to accelerate product development, but also to test and explore the deployment of AI tools such as AI Visualizer (in Archicad, Allplan and Vectorworks) 3D drawings (part of Bluebeam Cloud), and the dTwin platform. Nemetshek has presented this strategy as AI-as-a-service (AIaaS) to customers and partners.

Explaining Graphisoft’s approach to AI, CEO Daniel Csillag tells us, “There will be new services where we charge a subscription rate and not a perpetual license. Of course, inside of Graphisoft, we’re also trying to use more and more AI – for example, in customer service, such as tools that alert us to a certain user behaviour. Since this might give us an indication that the customer is unhappy, we can reach out proactively to them,” he says.

“Currently there’s a lot of brainstorming going on. I encourage the team to not talk about step-by-step improvements, but also to consider a jump forwards. I’m telling my team: be creative and sometimes disruptive.”

With those words probably ringing in their ears, I also spoke with Márton Kiss, vice president for product success, and Sylwester Pawluk, director of product management. Both executives have active roles to play in the shaping of Graphisoft’s core technologies.

The first Graphisoft AI product was AI Visualizer, a rendering tool that uses Stable Diffusion to produce design alternatives from simple 3D concepts. Graphisoft’s original hypothesis was that users were afraid of putting their IP on the cloud, so the company created a desktop, on-device version, which came with limitations.

“From actually talking to users, nobody is afraid of the cloud now, so with Archicad 28 (currently in technology preview), we have a cloud back-end for AI Visualizer,” says Kiss.

“We have also been experimenting with AI Visualizer Live, which is a camera looking at physical models, simple building blocks, rendered by an AI text prompt. We are thinking about how we productise this, as it’s a very practical way of using existing generative AI for concept iteration,” he says.

Find this article plus many more in the Sept / Oct 2024 Edition of AEC Magazine

Subscribe FREE here

From conversations with Pawluk, it’s clear that Graphisoft is looking to tackle this problem. The great thing about specification books is that they have a rigid formula and, through optical character recognition, are machine readable. They might include 2D and isometric drawings, too. So Graphisoft is experimenting with applying AI to read these catalogues and conjure up GDL-based 3D BIM objects that contain all the manufacturer’s metadata (N.B. GDL objects are parametric objects used within Archicad).

We have been experimenting with AI Visualizer Live, which is a camera looking at physical models, simple building blocks, rendered by an AI text prompt. We are thinking about how we productise this, as it’s a very practical way of using existing generative AI for concept iteration – Márton Kiss, vice president for product success, Graphisoft

As Pawluk explained: “With all generative models, what you need is a large amount of data and good-quality data. What we need first is a large number of good-quality GDL components to train the AI. In our development, the models, the speed of creation and complexity is rapidly increasing. We aren’t working with any specific vendors yet. It’s still in the exploratory stages at the moment.”

On the subject of autodrawings, we knew last year that Graphisoft was exploring this area, as well as its reverse – generating 3D BIM models from 2D drawings. Kiss confirms that building a bridge here is a very important part of the workflow.

Executives at Graphisoft also like the idea of assembly-based modelling, where pre-configured spaces are created, instead of modelling with walls, doors, windows.  “If you have someone creating those assemblies for you, or generated in the right way, a BIM model, from assembly, is actually going to be a very well-structured BIM model,” says Pawluk.

“Then later in the design, when producing the drawing, it’s already got a structure to work from. AI would be brilliant then.”

Catalogues and autodrawings

In this recent article we discuss how the quality of downloadable manufacturing content can be a hit-or-miss affair. It’s very rare for manufacturers such as Velux or Hilti to create full libraries of BIM objects that designers can add to their models.

Generic or personal AI

From our discussion, it was clear that Graphisoft is looking at developing AI capabilities that would benefit all designers, like AI Visualizer.

However, the company also understands the need for individual customers to train AI on their own models and past projects, which are not shared to train the generic AI.

Kiss comments: “This is where you start, from building your own ‘assemblies’ – rectangles with pre-loaded details essentially – and then use them as building blocks for hospitals, schools, whatever, based on what you have created in the past.”

While Kiss and Pawluk could not go into specifics of what such an app would do, they concede that the group has already trialled collaborative working and that this work ended up being presented to the Nemetschek board.

“The whole concept around this was to make sure that we can deliver those proof of concepts and productise initiatives very quickly. I think that has been proven on this occasion,” says Pawluk.

The post Graphisoft in the AI era appeared first on AEC Magazine.

Graphisoft has introduced Archicad AI Visualizer, an AI-driven image generation tool powered by Stable Diffusion. The software is designed to create detailed 3D visualisations during the early design stages via a simple user interface optimised for architecture and interior design.

To get started with Archicad AI Visualizer, users need a valid Archicad 27 license and Nvidia GPUs or Apple Silicon chips.

Users create a simple concept model in Archicad, then, using text prompts or a few descriptive words, like ‘a modern office with wood surfaces,’ generate any number of refined design variations — without creating detailed models for each. The tool produces design alternatives in the early design phase by adding details, context, and ideas to the original concept.

Prompts and results are optimized for architectural and interior design workflows. According to Graphisoft, Intellectual Property rights are fully protected thanks to the local storage of source images on users’ computers. Users can specify image sizes, vary the number of iterations to speed up image generation, edit the prompt strength for more precise results, and much more.

“The buzz around AI breakthroughs has shaken up the tech industry as a whole, with the promise of enabling increased creativity,” said Márton Kiss, vice president of product success at Graphisoft. “We want this tool to be tested in the real world by real users where they need it most — early in the design process when exploring designs and communicating with clients.”

The AI Visualizer will also be available for other Nemetschek Group brands Allplan and Vectorworks in the coming months.

The post Graphisoft launches Archicad AI Visualiser appeared first on AEC Magazine.