Section C.2 — Generating Images

Most of the techniques described here are based on diffusion models — a method that produces images from random noise through repeated refinement. Beyond the intuitive, semantic systems, you meet this diffusion technique wherever precise control of an image matters: in ControlNet workflows, in workflow rendering systems and in specialised tools for architectural visualisation.

AI-generated with Claude, January 2026 – editorially revised March 2026 – translation: Claude, August 2026.

Text to image was the starting point. Written text is turned into a picture by artificial intelligence. The prompt is the main lever — the more precisely it is written, the more control you have over the result. This basic technique underlies everything that followed.

The first encounter with it was usually sobering: the system produces images, but not the ones you had in mind. That prompts have a language of their own — with particular terms for style, light and perspective — only becomes clear through trying.

AI-generated with Claude, January 2026 – editorially revised March 2026 – translation: Claude, August 2026.

ControlNet added geometric control. Alongside the text prompt you supply a line drawing that defines the spatial structure. A photograph of our office is reduced to its essential geometric lines; the system uses these as the framework for generating the image. The room geometry is kept, while materiality and atmosphere are reinterpreted.

For architectural practice that was a breakthrough. For the first time you could work from an existing design instead of generating something entirely new out of a prompt. The geometry of the design stayed the basis; the AI filled it with materiality and atmosphere.

ControlNet is still the basis for precise image control in professional workflows. The technique builds on Stable Diffusion, an open-source diffusion model implemented in many specialised tools. Chapter C.2.5 covers ControlNet workflows and practical applications in detail.

AI-generated with Claude, January 2026 – editorially revised March 2026 – translation: Claude, August 2026.

Style Reference added the stylistic dimension. Reference images define the style you want independently of the geometry — materiality, colour and atmosphere are steered through pictures rather than through descriptions in words.

Combining ControlNet (geometry) with Style Reference (aesthetics) gives you control in two dimensions: “this is how the room geometry should be” and “this is how it should look” are steered separately. That matches how architects actually think: the plan and the atmosphere are two different decisions.

AI-generated with Claude, January 2026 – editorially revised March 2026 – translation: Claude, August 2026.

InPainting allows selective editing. A marked area is recalculated; the rest stays untouched. The method resembles retouching in Photoshop, but it is done by AI.

In practice, exactly what should change and how far the change extends is often interpreted unpredictably. We saw this in tests with horizontal louvres on glass façades — the system understood the task, but the edge between the edited and unedited area was hard to control. InPainting works well for large-area changes. With small, precise interventions the limits are noticeable.

Chapter C.2.4 covers practical applications, in particular the integration into Adobe Firefly and Photoshop.

AI-generated with Claude, January 2026 – editorially revised March 2026 – translation: Claude, August 2026.

Image to image was the real leap: semantic image understanding. Since then systems have been able to interpret images and modify them deliberately. “Tidy up the office” produces a newly calculated picture with the furniture in order. The system understands the room geometry, identifies objects and works out how they should sensibly be arranged.

The difference from everything before it: no line drawing is needed any more, no reference image, no precise description of the geometry. You show the system an existing picture and describe what should change. This ability is what fundamentally separates the new leaders from classic diffusion models, which respond only to prompts and geometric input. Chapters C.2.2 and C.2.3 deal with these systems individually.

AI-generated with Claude, January 2026 – editorially revised March 2026 – translation: Claude, August 2026.

Image to prompt runs the other way: the system analyses an existing photograph and writes a detailed text prompt describing it. That prompt can then be used in different image generators.

At first that sounds like a technical aid. In practice it is a useful tool for comparing systems: which generator reconstructs a given reference image most accurately? And it helps you improve your own prompt language — you see how the system describes a picture that you would only have put into words imprecisely.

AI-generated with Claude, January 2026 – editorially revised March 2026 – translation: Claude, August 2026.

Style transfer transforms existing spaces while keeping their geometry. An office becomes a candy shop or a green oasis. Windows and proportions stay; materiality and use change radically.

One thing stands out: with repeated transformations the image drifts further from the original — small deviations add up at every step. What is left after three iterations sometimes shares only the rough spatial structure with the picture you started from. That is not a fault but a property of the method, and one you need to know about.

The line of development shows that each new technique overcame a particular loss of control. Text to image was steering chance through language. ControlNet added geometric anchoring. Style Reference separated form from aesthetics. Image to image finally made the picture itself the thing you steer with.

Two worlds now stand side by side: semantic systems such as ChatGPT, Gemini and Flux for fast, dialogue-based work — and diffusion-based workflows for maximum control and reproducibility. Which approach suits which task is what the following chapters describe.

AI-generated with Claude, January 2026 – editorially revised March 2026 – translation: Claude, August 2026.

Google Gemini 2.0 Flash achieved a breakthrough in spatial image understanding. For the first time you could show a system a photograph of an office and give it a specific instruction such as “Tidy up our office”. What was remarkable: the system genuinely understood what an office is, identified the furniture and interpreted the request correctly. That made Gemini 2.0 Flash the pioneer in this field and opened up complex spatial manipulations that had been unthinkable before.

Gemini 2.0 Flash did not stay in place for long, though. It was withdrawn again fairly quickly and sent back for revision. No official reasons were given, but the withdrawal suggests the kind of teething troubles or unexpected behaviour that often accompany a technical first.

Only two months after the first launch came “Nano Banana” — the unofficial name for Gemini 2.5 Flash — and it was a decisive step forward. It showed how fast the progress is. Where the first attempt had counted as impressive, the same task in the follow-up version looked practically revolutionary: the image quality made the earlier results look almost primitive by comparison.

Concrete examples show best what this development means. The AI could not only tidy an office on command; it could now simulate a change of viewpoint as though the camera were actually moving through the room. Details such as glazing bars, door handles and even hinges came out correctly. Set beside real photographs, they were barely distinguishable.

“Nano Banana” stands for a new quality of image generation: the system moves away from simple image transformation towards genuinely understanding and interpreting space.

AI-generated with Claude, January 2026 – editorially revised March 2026 – translation: Claude, August 2026.

ChatGPT Omni lifted image generation to a new level. It presented comparable functionality only a week after the launch of Gemini 2.0. That speed says a great deal about the innovation cycle in this industry — a race in which technical leaps happen within days.

The particular strength of ChatGPT Omni is its precise spatial understanding and a deliberately conservative reading of image content. It can not only recognise complex scenes but grasp their spatial structure. One outstanding technical feature is that it produces depth masks: for any image, ChatGPT Omni can generate a depth map showing how far each pixel is from the camera. That function is essential for moving from 2D to 3D.

The visual language of ChatGPT Omni is often described as “androgynous” — neutral, smooth and technically perfect, but with a certain uniformity. The aesthetic is professional and flawless, which makes it ideal for technical visualisations, presentations or option studies.

What is genuinely new is that the system no longer only recognises what is in an image but understands how it is arranged in space. The AI can derive a 3D model from a single photograph and render it from different viewpoints. Combining semantic and spatial understanding makes ChatGPT Omni a key tool for the next generation of AI-based image editing.

AI-generated with Claude, January 2026 – editorially revised March 2026 – translation: Claude, August 2026.

With the launch of Flux Playground, Black Forest Labs — based, as the name says, in the Black Forest — set a decisive milestone in AI image generation. While Gemini and ChatGPT were already reshaping the market with semantic image understanding, Black Forest Labs deliberately took another route, bringing its own model into the competition only once it was mature and deliberately positioned.

Flux was originally conceived as an open-source model for professional users. Installing it and building it into complex workflows took technical know-how, but it gave exceptional control over the generation. The real breakthrough came with Flux Playground: moving from a local, hard-to-reach solution to a web application opened the door to a wide audience.

The strength of Flux shows most clearly in direct comparison. Where models such as Midjourney generate a completely new image with every transformation, Flux respects the original and allows targeted, semantically grounded changes. The system does not only understand what is in the picture; it can modify individual elements precisely.

The business model underlines how accessible it is: instead of a fixed subscription, Flux Playground works with a flexible credit system. You pay only for the images you actually generate — and there is a generous free tier that allows extensive testing without any commitment.

AI-generated with Claude, January 2026 – editorially revised March 2026 – translation: Claude, August 2026.

The three new leaders differ from older systems in one basic ability: semantic image understanding. They do not only generate images; they interpret existing ones and can modify them deliberately.

These systems understand what is in the picture. Objects, spaces and materials are recognised and categorised. They understand where things are. Spatial structure, perspective and depth are grasped. They understand how the picture works. Physical plausibility and lighting conditions are taken into account.

Semantic understanding makes targeted manipulation possible. “Empty the office” works. “Change the façade” works. “Move closer to the window” works. The picture is not reinterpreted but modified deliberately.

The contrast with Midjourney is clear. Midjourney reads the prompt as an instruction for a new image. It interprets and creates afresh. The three new leaders read the picture as a basis to be changed. Midjourney is a generative artist; the new systems are digital image editors.

The practical difference shows in how they handle a request to change something. If the building is to get a timber façade, the new systems respect the original and change only the façade. Midjourney treats the picture as a template and produces an independent new image. Control over the result is considerably greater with the new systems.

AI-generated with Claude, January 2026 – editorially revised March 2026 – translation: Claude, August 2026.

The technology is developing at a remarkable pace. Within a few months, what these systems can do has changed fundamentally.

The first versions with mature image understanding produced usable results but still showed clear weaknesses in spatial interpretation. Two months later the improvement in quality was dramatic. Photorealistic representations became possible. Spatial understanding improved considerably. Even complex situations such as perimeter block development with courtyards were interpreted correctly.

The Koblenz examples make the progress clear. The first result was usable for early studies. The next reached presentation quality. The remaining errors became smaller and more specific — individual details where the system overstepped, rather than fundamental misunderstandings.

Within three months the technology went from an interesting experiment to a production-ready tool. That pace continues.

AI-generated with Claude, January 2026 – editorially revised March 2026 – translation: Claude, August 2026.

Semantic image understanding opens up new ways of working. The systems can be combined deliberately and brought into different stages of producing an image.

A typical workflow might start with a hand drawing or a screenshot from the BIM model. The template is uploaded to Flux Playground. From it the system generates a first working model with the spatial interpretation correct.

The working model goes to Midjourney. There it becomes an atmospherically richer image, often an aerial view. Midjourney adds dramatic light and design quality.

The aerial view is worked on further in Nano Banana. Specific changes are made in natural language: summer evening light, more greenery, snow cover, a night-time mood. The system understands the requirements and modifies accordingly.

Natural language as the interface for image editing changes the work fundamentally. Instead of technical parameters you state intentions. “More greenery” instead of colour values. “Evening light” instead of an angle of illumination. You communicate in professional language, not in software commands.

Combining the systems plays to their respective strengths. Flux for spatial precision, Midjourney for design quality, Nano Banana for targeted adjustments. The workflow is adapted flexibly to the task.

AI-generated with Claude, January 2026 – editorially revised March 2026 – translation: Claude, August 2026.

Set beside the models with semantic image understanding, Midjourney shows the fundamental shift of paradigm. Midjourney reads the prompt as a fixed description of an image; the other systems work with a developed semantic understanding. They recognise and interpret the picture, can address particular parts of it, produce variants and respond to change requests in open dialogue. You no longer steer through a single initial prompt but through iterative exchange.

All the same, Midjourney remains indispensable in our work wherever the final aesthetic refinement is at stake. Its strengths — dramatic lighting, atmospheric quality, a creative way with instructions — make it the tool of choice for artistic enhancement. But precisely where technical precision, targeted changes or data protection are decisive, classic prompt systems reach their limits.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

The copyright position on AI-generated images is complex and largely unsettled. As things currently stand, no copyright can be claimed for images produced with ChatGPT, Gemini, Flux or Midjourney. AI systems create images by processing training data statistically. The threshold of originality that copyright protection would require is generally not credited to the person writing the prompt.

The legal debate turns on two questions. First: can AI-generated work be protected by copyright at all? Most legal systems require human creation. Second: who would the author be — the person prompting, the operator of the system, or nobody? These questions are assessed differently from country to country and settled conclusively in none.

Midjourney adds further complexity through its open structure. Every image generated is publicly accessible unless you use private mode. Every generation adds to the open image database and is used to train future models. Confidentiality for sensitive projects is not possible.

For commercial or confidential projects, closed systems with clear terms of use are advisable. Local installations such as ComfyUI or Invoke give full control over images and data. Use stays with you, and the images cannot be fed into someone else’s model training.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

Adobe Firefly is integrated into Photoshop as a plug-in and allows InPainting, generative fill and outpainting directly within the layer structure. Since it opened up to external models, Flux and Nano Banana Pro can be used alongside Firefly. That gives you different visual languages for different tasks — all within a familiar interface.

We tested Firefly on three tasks: removing an excavator bucket from a site photograph, adding a lawn to a photograph, and visualising sketches through Flux and Nano Banana. The system handled the first two reliably. Visualising sketches through the external models gave usable preliminary studies, not print-ready results.

Firefly works with clearly licensed training data. That matters for commercial use and sets it apart from many other systems.

Suitable for: cleaning up photographs, adding natural elements, transforming sketches with external models. Limitations: precise technical detail, regular patterns, structural façade elements. Licensing: Adobe Commercial License — usable for commercial projects.

AI-generated with Claude, January 2026 – editorially revised March 2026 – translation: Claude, August 2026.

Stable Diffusion differs fundamentally from commercial systems. The code is openly available. Anyone can download the model, run it locally and adapt it for their own purposes. That openness led to an explosion of specialised applications.

The technical basis is a diffusion model. Images emerge from random noise through repeated refinement. The process is computationally heavy but precisely controllable. Every parameter can be defined exactly. The result is reproducible.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

Many specialised tools use Stable Diffusion as their image generator. Visoid for architectural visualisation, Krea for real-time generation, various ControlNet implementations — all of them built on Stable Diffusion. The model supplies the computing; the applications supply the interface.

This division of labour makes innovation possible. Developers concentrate on specific use cases. The proven Stable Diffusion model handles the image generation. Architects get tools such as Visoid, tailored precisely to their needs, without anyone having to develop a new image generation model.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

Stable Diffusion needs powerful hardware. A dedicated graphics card with enough video memory is necessary. For many people, installing it locally remains a technical hurdle. Cloud services offer an alternative: virtual machines with Stable Diffusion preinstalled, rented by the hour.

Operating it is fundamentally different from intuitive systems such as ChatGPT. Not a simple dialogue but technical parameters. Not semantic understanding but precise control. It takes considerable time to learn. The gain is control.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

ControlNet added geometric control to Stable Diffusion. Combining an open model with geometric control created new possibilities. Hand sketches become detailed visualisations. Photographs become stylised representations. The spatial structure is kept.

ControlNet works only with Stable Diffusion. Commercial systems such as Midjourney offer no comparable interface. Only the technical openness of Stable Diffusion made this development possible.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

Stable Diffusion allows specialised variants to be trained. LoRAs (Low-Rank Adaptations) adapt the model to particular styles. Form-Finder by Ismail Seleit is optimised for architectural representation. Other LoRAs cover specific materials, periods or ways of drawing.

This kind of specialisation would be impossible with closed systems. The open-source nature allows development driven by the community. Architects, designers and visualisers build tools for their own particular needs.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

Complex image generation workflows are built on Stable Diffusion. ComfyUI, Invoke, X-Figura — all of them use it as their processing engine. The workflow systems organise several generation steps, control parameters and chain processes together. Stable Diffusion does the actual image calculation (see Chapter C.2.7).

These workflows reach a precision that simple prompt systems cannot. Change materials without affecting the geometry. Adjust the lighting without moving objects. Refine step by step with exact control at every stage.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

Stable Diffusion runs on your own hardware. No data goes to an outside server. Full control over the images you generate. Nothing is used to train someone else’s model. For confidential projects that is decisive.

Closed systems such as Midjourney process every image on their servers. Prompts are publicly visible. Images train future models. For commercial or sensitive projects that is a problem. Stable Diffusion is the alternative.

Stable Diffusion keeps developing. Newer versions improve quality and speed. The community contributes optimisations, specialisations and new applications. While semantic systems such as ChatGPT and Flux dominate public perception, Stable Diffusion remains the foundation for specialised, controlled image generation that meets data protection requirements.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

Veras is an AI-supported visualisation tool from Chaos that runs as a plug-in inside common BIM and CAD programs (Revit, SketchUp, Rhino, Vectorworks, ArchiCAD, Forma). From 3D models, 2D images or hand sketches it produces photorealistic renderings and short animations within seconds. You steer it through text prompts and sliders — for fidelity to the geometry, material intensity and the influence of the prompt. The underlying BIM geometry stays unchanged.

Resolutions up to 4K are possible, and image-to-video conversion is built in. Processing happens in the cloud. Veras suits the early design phase, developing ideas and quick client presentations — not as a replacement for final renderings.

AI-generated with Claude, January 2026 – supplemented with AI-assisted web research – translation: Claude, August 2026.

Typus.AI is a browser-based SaaS platform from Cologne for AI-supported architectural visualisation. Plug-ins for Revit, ArchiCAD, Rhino and SketchUp send 3D models, viewports, floor plans, photographs or sketches to the web app. From these it generates photorealistic visualisations in under 60 seconds, keeping the original geometry and texture assignments. Materials can be adjusted from a built-in catalogue. One particular feature is creative upscaling to 13K resolution.

Typus.AI is aimed at architects and planners who need presentable visualisations quickly without an external rendering service — above all for façades and exterior views.

AI-generated with Claude, January 2026 – supplemented with AI-assisted web research – translation: Claude, August 2026.

PeliCAD is a cloud-based visualisation platform, a spin-off from RWTH Aachen, which processes 3D CAD and BIM data in IFC format and turns it into renderings that respond to their context. Its distinguishing feature is that it generates a textured model of the surroundings automatically: enter an address and the platform produces the real setting, including neighbouring buildings and terrain. Design intent is given through text prompts or moodboard images; the geometry of the model is preserved.

Through the Speckle plug-in there is a live connection to Revit, ArchiCAD, Rhino and other planning software. According to the provider, high-quality visualisations take about 30 minutes — without a local GPU or any rendering expertise. The tool suits feasibility studies, public consultation and early project phases particularly well.

Reference: Yannic Schwarz.

AI-generated with Claude, January 2026 – supplemented with AI-assisted web research – translation: Claude, August 2026.

Snaptrude is a browser-based concept design platform that aims to cover the route from the schedule of accommodation to an exportable BIM model in a single tool. Chapter C.3.3.4 describes how it works in more detail. Part of it is a built-in AI visualisation, so first concepts can be visualised quickly.

AI-generated with Claude, January 2026 – supplemented with AI-assisted web research – translation: Claude, August 2026.

ComfyUI shows workflows as graphical diagrams. Individual nodes represent steps. Connections show the flow of data. The interface is reminiscent of visual programming.

Different image generation models can be brought in — Stable Diffusion in various versions, Flux, specialised LoRAs. Every node offers a long list of settings. The learning curve is steep.

ComfyUI could originally only be installed locally. Powerful hardware with a dedicated graphics card and enough video memory was required. There is now a cloud-based version. The system runs in the browser. The hardware hurdle disappears.

It takes several weeks to learn. ComfyUI does not suit occasional use. It needs continuous engagement and technical understanding. The investment pays off for complex, recurring rendering tasks.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

Invoke translates workflow principles into a Photoshop-like interface. You work in layers. Individual areas of the image are marked in colour. Each area gets its own prompt and parameters.

A typical workflow: the background gets a prompt for landscape. The middle layer a prompt for buildings. The foreground a prompt for vegetation. The composition builds up step by step. Changes to one layer do not affect the others.

Invoke has been acquired by Adobe. Integration into the Adobe Creative Cloud is on the horizon. The technology will presumably find its way into Photoshop. Development as a standalone product may be discontinued.

It takes several weeks to learn. Less complex than ComfyUI, but more demanding than simple prompt systems. The layer metaphor makes it easier for Photoshop users to get in.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

X-Figura translates the workflow approach into a more intuitive graphical interface. It is more accessible than ComfyUI but offers similar possibilities. Different image generation models can be combined. Workflow steps are shown in an environment you can follow.

X-Figura is in beta. The system keeps developing. Ease of use comes first — it is meant for people who want to use complex workflows without having to master an interface that resembles programming.

The balance between accessibility and range of functions is what makes X-Figura interesting. Not as powerful as ComfyUI, but considerably easier to learn. Not as intuitive as Midjourney, but considerably more controllable.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

Low-Rank Adaptations, or LoRAs, are small packets of weights laid over an existing image generation model that reprogramme it for a particular purpose, without retraining it completely. You train a LoRA on your own set of images and then attach it to a base model. The model thereby learns a specific style, a material language, a building typology. LoRAs work across models: they run on Stable Diffusion, SDXL and Flux alike, and several can be combined to steer the generation precisely.

Ismail Seleit has taken this approach furthest for architectural image generation. Under the name Form-Finder he has produced — alongside his regular office work, in his own time, over a few months — a body of architectural imagery remarkable in both range and quality. We saw him at the AI in Practice Summit in London: a young architect who, through deliberate manipulation of image models, had generated more unusual architectural visions in a short time than many established practices manage in years. That raises a serious question: who still sets the visual standards — and how, in this flood of images, do we keep the ability to select and to criticise?

At the Architects Network we were able to follow, over a two-day course, how Ismail Seleit has developed an entirely new approach to architectural form-finding using ComfyUI and deliberate image manipulation through task-specific LoRAs. Starting from a project narrative and your own reference images, a workflow emerges that is developed further for each project and can lead all the way from image to model.

Reference: Ismail Seleit.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

Workflow rendering systems have decisive advantages for confidential projects. ComfyUI runs locally, or on rented servers with guaranteed ownership. The images stay yours. Nothing is used to train someone else’s model. Nothing is publicly visible.

The contrast with Midjourney is clear. Every Midjourney image is public, except in the expensive private mode. Every generation trains future models. Prompts can be read by anyone. For sensitive projects it is unsuitable.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

Workflow rendering requires specialisation. Not everyone in a practice can or wants to master these systems. Our recommendation: train one person as the AI specialist.

That specialist tests new tools, develops workflows, trains colleagues and is the point of contact for complex visualisation tasks. The time it takes: regular, but not full time. Around one day a week for continuous learning and workflow development.

Workflow rendering does not suit every task. Recurring rendering tasks with the same structure benefit from predefined workflows. Maximum control for presentations to demanding clients justifies the effort. Confidential projects need local control rather than cloud services.

For quick moodboards, early design studies or simple presentations, semantic systems such as ChatGPT or Flux are the better choice. For final, controlled renderings with reproducible results there is no way round workflow systems.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

Krea is a browser-based AI creative suite that brings image, video and 3D generation together with editing and upscaling on one platform. What is special about Krea is not a model of its own but the integration of the leading third-party models — among them Flux, Google’s Veo 3, Runway Gen-4, Ideogram, Imagen 4 and Krea’s own Krea-1 — under one simplified interface. You choose the model that suits the task without switching between platforms.

Two functions are particularly relevant for architectural practice: real-time rendering and the Enhancer. The real-time canvas produces previews in under 50 milliseconds — sketches, geometric forms or keywords turn into photorealistic output as you type. That allows an exploratory way of working, close to sketching, which is foreign to other AI image generators. The Enhancer upscales images to as much as 22K resolution and includes seven different upscaling models, among them Topaz Photo AI and Topaz Gigapixel.

Krea also offers LoRA fine-tuning: from a handful of your own reference images you can train adapted models that reproduce a particular architectural style, a material or a practice’s project consistently. The free version allows first tests; productive use needs a paid subscription.

AI-generated with Claude, January 2026 – supplemented with AI-assisted web research – translation: Claude, August 2026.

There is no curriculum for this technology and how to use it, and no textbooks. Development is so fast that any account is out of date by the time it has been written down and published.

At its core it is about acquiring skills by learning from people who already have them. Put formally: you learn a language by speaking it. In many respects, though, the approach needed is far simpler and more direct — like the day someone idly taught you to whistle loudly through two fingers.

Translated into our professional context, the recommendation to management is: give your staff the time and the room to move freely on virtual playgrounds — and to learn there how the new craft works.

Ismail Seleit’s course at Architect Network was exactly that. A strong recommendation: make room for play and experiment. Let the people in your practice attend events like it, even when no measurable benefit is apparent at first sight.

AI-generated with Claude, January 2026 – supplemented with AI-assisted web research – translation: Claude, August 2026.

Runway is a cloud-based platform for AI video generation from New York. Its core product is the Gen family of models. They support text-to-video, image-to-video and video-to-video workflows. Clips are produced at up to 4K and can be extended to 26 seconds. A particular feature is precise camera control: you can set the direction and intensity of movement and add handheld effects. Keyframe control lets you specify the first, middle and last frame of a sequence.

Runway demonstrably works with architectural practices, among them KPF, who use the system for internal rendering workflows and for animating design models. What matters most for architectural practice are the camera controls and the ability to turn rendered images into moving sequences.

AI-generated with Claude, January 2026 – supplemented with AI-assisted web research – translation: Claude, August 2026.

Kling is an AI video generator from the Chinese technology company Kuaishou. The current version supports text-to-video, image-to-video and structured multi-scene generation with up to six shots per clip. Videos are produced at up to 1080p at 48 frames per second; individual clips can be extended to three minutes — considerably longer than most competitors.

From Kling 2.6 onwards, audio and video are generated in one pass; speech, ambient sound and music are created in sync with the picture. For architectural use the interesting parts are the camera control, the motion brush for steering movement in particular areas of the image, and the ability to develop massing models and renderings into animated sequences. There are limits on availability in Europe, and on data protection, since processing happens on Chinese servers.

AI-generated with Claude, January 2026 – supplemented with AI-assisted web research – translation: Claude, August 2026.

Veo is Google’s model for AI video generation, developed at Google DeepMind. It is built on a latent diffusion transformer architecture that processes video and audio data in a shared latent space. As a result, audio — dialogue, sound effects, ambient sound — is produced in sync with the picture in a single generation step.

Videos are produced at up to 1080p; by putting clips one after another you can assemble sequences of up to 60 seconds. Every generated video is marked with SynthID, Google’s AI watermarking standard. In independent comparisons Veo currently gives the strongest results among the available models for visual quality, physically based motion and fidelity to the prompt. For architectural use, the physically correct rendering of light, materiality and movement matters most.

AI-generated with Claude, January 2026 – supplemented with AI-assisted web research – translation: Claude, August 2026.

Midjourney released its first video model as a pure image-to-video system: you upload an existing image — or use one generated earlier in Midjourney — and animate it into a five-second clip. Repeated extension allows a maximum of 21 seconds. There are two animation modes: automatic, where the system works out the logic of the movement itself, and manual, where camera movement and subject movement are described in a text prompt.

The model puts stylistic consistency with Midjourney’s characteristic look above technical control; the output feels artistic and dreamlike rather than photorealistic. Compared with Runway, Kling or Veo it is considerably more limited in resolution, length and control. As it stands, the tool suits stylistic animation of rendered images rather than precise design visualisation.

AI-generated with Claude, January 2026 – supplemented with AI-assisted web research – translation: Claude, August 2026.

For quick design studies and moodboards: Midjourney delivers attractive images with dramatic light. Usable immediately, with nothing to learn first. Limitation: no precise geometric control.

For targeted image manipulation: ChatGPT or Flux understand the content of an image and allow deliberate changes through dialogue. “Empty the office”, “Change the façade”, “Make it evening” work reliably. Flux offers better value; many people already have ChatGPT.

For maximum control: workflow rendering with ComfyUI or X-Figura gives precise control over every parameter. Time to learn: several weeks. Worth it for recurring complex rendering tasks and confidential projects.

For confidential projects: local systems based on Stable Diffusion guarantee data protection and ownership. Commercial cloud systems train their models on user data.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

The minimum kit has three components.

Midjourney: the proven basic tool for attractive visualisation. A free account is not enough for serious work.

Nano Banana or ChatGPT: semantic image understanding for targeted manipulation. Many people already have ChatGPT; Flux offers a generous free tier. The credit system keeps costs down.

Specialisation optional: workflow rendering when a concrete, deeper need arises. Train one person as the specialist rather than teaching everyone in the practice the same technique.

Recommended credit for AI-generated images: “Produced by [practice name] with [AI system]”. Example: “Produced by seg Architekten with Midjourney.” Transparency builds trust and documents the process.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.

Development is rapid. In the short term: further improvement in spatial interpretation, more precise rendering of materials, simpler interfaces. In the medium term: integration into CAD workflows, automated option studies, real-time visualisation during design.

The central challenge is shifting from “How do I generate images?” to “How do I deal with the flood of images in a structured way?” Clear criteria and limits on the number of variants become decisive. It is time to start. The technology is production-ready.

AI-generated with Claude, January 2026 – editorially revised February 2026 – translation: Claude, August 2026.