Large language models and “do‑everything” image generators have set a new baseline expectation: anyone can type a prompt and receive a plausible image or paragraph in seconds. For architects, engineers and visualisation teams, that shift is impossible to ignore – clients now arrive at meetings with AI‑generated concepts on their phones, and internal teams quietly test public tools on sensitive projects.
Yet in practice, these general‑purpose systems rarely satisfy the demands of design and AEC workflows. They do not understand building performance, construction constraints or the nuance of physically based materials; they are usually trained on uncontrolled web data; and they tend to treat geometry as a suggestion rather than a contract. For teams working with confidential models, demanding clients and tightly governed approval processes, this combination is risky.
That reality has pushed companies like Chaos – which builds tools such as V‑Ray, Enscape and Veras – towards a different pattern: pairing domain‑specific models with strong UX, robust data pipelines and explicit governance, rather than hoping a generic model can do it all. What follows are some of the key lessons from that journey that other ML practitioners can apply in their own design‑adjacent domains.
Pattern 1: Put UX and Domain Logic Around Off‑the‑shelf Models
A recurring theme in design‑native ML is that the model is rarely the product. In many cases, the core model can be an off‑the‑shelf system such as Stable Diffusion – what differentiates a usable tool is the domain‑specific pre‑processing, controls and workflow it sits inside.
Chaos’s AI Material Generator is one example. At first glance it behaves like a straightforward image‑to‑image tool: upload a photograph of a surface and it produces a ready‑to‑use PBR material with albedo, normal and roughness maps, accessible directly via the Cosmos library. Under the hood, however, the pipeline normalises disparate input images, tiles non‑repeating patterns, and enforces constraints that make the output usable in production – for instance, converting the result into a material compatible with V‑Ray and Enscape so it can drop straight into existing scenes.
The same philosophy appears in Chaos AI Enhancer. The underlying enhancement is driven by a diffusion‑based model, but the product value lies in the way it wraps that model in an intuitive review workflow: artists can send EXR renders to Chaos Cloud, apply AI Enhancer, and compare before/after images via sliders and versions without altering core geometry. The tool is tuned specifically to improve people, vegetation and fine details – the areas where noisy or low‑res real‑time renders most obviously break immersion.
Lesson for practitioners: if you are working in a specialised domain, start by deciding what “production‑ready” means for your users – then design pre‑ and post‑processing plus UX flows that enforce those constraints, even if you begin with a relatively generic model. The model choice matters, but the domain logic and interface around it are often where the real differentiation lies.
ALSO READ: From Renders to Data Layers: How AI Is Reshaping Architecture’s Visualisation Stack
Pattern 2: Treat Data Pipelines and Reproducibility as First‑class Citizens
As generative features move from experiments to shipping products, the main bottleneck often shifts from model architecture to data engineering and reproducibility. Chaos’s work on material generation is a good case in point.
Architectural materials are not ordinary 2D images. Many require high‑dynamic‑range or floating‑point textures to render correctly under different lighting conditions, which rules out a lot of standard consumer image pipelines and file formats. To support this, Chaos invested in cloud‑based ETL pipelines that ingest material images from multiple sources, conform them into compatible types, and transform them into formats that can be streamed efficiently into training and inference workloads. In some cases, the team had to hand‑tune data loaders to avoid bottlenecks when reading large floating‑point textures at scale.
On the experiment side, they combined Git, DVC and MLflow to guarantee that any given model could be reproduced one‑to‑one from the same code, parameters and data snapshot. DVC tracks datasets and file‑level lineage alongside code, while MLflow records metrics, hyper‑parameters and artefacts, and acts as a registry for models handed off to deployment teams. The result is an ecosystem where changes to data and code are both visible, and where you can answer questions like “which dataset and commit produced this material generator currently running in Cosmos?”
Lesson for practitioners: once you move beyond toy examples, assume that you will outgrow ad‑hoc folders and spreadsheets. If your domain relies on unusual data types – whether that is floating‑point textures, CAD files or sensor streams – plan for custom ETL and loaders, and invest early in reproducible pipelines that tie code, data and experiments together. It will pay off every time you iterate on a feature or defend a model decision to a demanding customer or regulator.
Pattern 3: Build Explicit Guardrails for IP‑sensitive Domains
Architecture and design projects often involve sensitive client information: unannounced developments, luxury interiors, or public‑sector infrastructure that cannot be leaked via external tools. For that reason, data governance and responsible AI are not optional extras – they are a prerequisite for adoption.
Chaos has formalised this via a Responsible AI initiative and public‑facing tracker, which documents which models are used in which products, what datasets they rely on, and how customer data is handled. For example, the company states that certain tools only collect anonymous usage data on an opt‑in basis, and that Chaos does not claim ownership over customer outputs produced with its AI features. The policy also clarifies where third‑party models are hosted geographically and how that aligns with customer privacy requirements.
In practice, this transparency has two effects. First, it gives risk‑averse firms a concrete document they can share with internal legal and IT teams when evaluating tools. Secondly, it opens conversations with brands and public organisations that might otherwise avoid AI entirely, because they can see exactly what is happening to their models and renders rather than treating the system as a black box.
Lesson for practitioners: if you are deploying generative or agentic features into regulated spaces, plan for governance from day one. At minimum, that means publishing clear information about training data, data retention, customer control over telemetry, and any third‑party models or services in your stack – and making sure the product gives users realistic options to disable or constrain AI features where needed.
ALSO READ: Dr Lee Schlenker’s Playbook For Boards Before EU AI Act’s Enforcement
Pattern 4: Design for Collaboration, not just Output
As generative tools become more capable, the question shifts from “can we make an image?” to “how does this slot into our design and approval processes?”. For many firms, the real value lies not in a single hero render, but in how visual output connects stakeholders and decisions over time.cgconnect.
Chaos’s cloud services reflect this shift. Chaos Cloud and related collaboration tools allow teams to stream 3D scenes, share panorama tours, and review renders in the browser, with comments, mark‑ups and version history captured alongside the imagery. Recent updates add more structured review and approval workflows, turning the platform into a single source of truth for what has been seen, discussed and signed off – effectively moving visualisation closer to a governed digital‑twin environment than a folder of static images.
When generative features such as Veras are layered on top of this stack, the pattern becomes even more interesting: scene‑aware AI can propose stylistic variations, seasonal changes or façade edits directly against the same underlying 3D model, while the collaboration layer records which options were explored and approved. Over time, that combination of geometry, imagery and meta‑data starts to look a lot like a living twin of the design process itself, not just its outputs.
Lesson for practitioners: consider how your ML features will plug into collaboration and approvals. If you already have a review platform or digital‑twin environment, treat generative tools as a layer on top of that shared context, with clear audit trails and the ability to relate AI‑driven variations back to an agreed “source of truth”.
Towards Living Digital Twins and Agentic Assistants
The long‑term direction hinted at by these patterns is a world where the 3D scene or model is the primary source of truth, and generative AI operates as a controlled layer on top. Instead of exporting images for an external tool to manipulate, systems can use scene metadata, camera positions and environment settings to apply targeted changes such as “switch to a winter scenario”, “remove nearby buildings for this view”, or “update materials to match the latest specification package” – all without losing the link to geometry and performance constraints.chaos.
From there, it is not a huge leap to imagine agentic assistants that live inside design and twin platforms: proposing alternative layouts that satisfy code, flagging renders where privacy policies may be breached, or pre‑configuring review sets for different stakeholder groups. The critical point is that such agents must operate within the same governance, data and UX frameworks described above, not as unchecked sidecars.
For ML teams working in adjacent domains – from industrial design to urban planning or manufacturing – the message is clear. Generic models may be the spark, but production‑ready systems will be defined by domain‑specific constraints, data engineering discipline, clear guardrails and careful integration into collaborative workflows. The route from “clever demo” to “living twin” runs through that territory.
ALSO READ: From Siemens Energy to Bank of America: What “Quietly Advanced” Enterprises are Doing Differently
