Architecture & interior design visualisation

Classic rendering means building detailed 3D scenes, assigning materials and waiting for a render engine, which is slow and expensive in early design phases. Image-generation models can take a rough viewport, sketch or photo and return a photorealistic image in seconds, so a designer can show a client ten material and lighting options in an afternoon. The images are illustrations of intent, not geometry-accurate documentation, and must be labelled and checked accordingly.

Problem & value

• Today: early-stage presentations rely on hand sketches, white-model screenshots or conventional rendering (Enscape, Lumion, V-Ray, Twinmotion, D5 Render) that needs a finished, textured 3D model. Rendering studios charge per image and turnaround takes days. Interior designers assemble mood boards by hand.
• AI value: fast design exploration (many variants per hour), cheaper client communication early on, easier "what if" questions for non-expert clients, and lower barriers for small practices and private renovators.
• Evidence: reviews of 2026 tools consistently describe the sweet spot as concept and schematic design, not final construction visuals; vendor and blog claims on time savings are mostly anecdotal, so measure your own baseline [1][2].

Variants

• Sketch/massing to render: SketchUp or Rhino viewport plus prompt gives a styled image that respects the modelled geometry.
• Interior restyling: photograph of an existing room, then new furniture, colours and materials (private renovation, showrooms). Staging for sale listings is covered in GEN-018.
• Facade and context studies: variants of cladding, glazing and landscaping on a fixed volume.
• Walkthrough clips: image-to-video for short camera moves (see GEN-002), or photo-to-3D capture for explorable scenes (GEN-025).
• Public participation: municipalities and developers visualise planning options for residents; here labelling as illustration is essential.
• Private use: homeowners test paint, floor and furniture options before buying.

How it works

1. sketch / photo / BIM viewport export (colour + depth + edges)
2. conditioning (geometry control: depth map, edge map, segmentation, reference image)
3. prompt + style/material references (text, mood board, IP-adapter image)
4. image generation (diffusion / flow model)
5. variants
6. selective editing (inpainting of a wall, floor, furniture; upscaling)
7. designer review
8. labelled export (AI marker)
9. client board / presentation

• Geometry control: the key technique. ControlNet-style conditioning on depth, line or segmentation maps keeps walls, openings and proportions of the source view while the model invents materials, light and atmosphere. Plugin tools expose this as a "geometry fidelity" slider.
• Instruction-based editing: newer models (FLUX Kontext-class, Gemini image models, GPT image models) edit a given picture by instruction ("replace the oak floor with terrazzo") and keep the rest stable.
• Human-in-the-loop: the designer chooses and corrects variants; anything shown to a client, authority or buyer must be marked as an AI-assisted illustration, and dimensions, materials and code compliance are taken from the model, never from the picture.

Data preparation

Needed inputs are viewport exports (colour, depth, edge maps) from SketchUp, Revit, Rhino or Archicad, site and room photos, and material or mood-board references. These typically live on project file shares, in the CAD/BIM project folders or in the architect's mail; client photos may sit on phones and in messengers. Typical gaps are inconsistent file naming, low-resolution or badly lit photos, missing project metadata, scattered material libraries and unclear usage rights for client images and third-party reference pictures.

Preparation steps: agree on a project folder structure and naming scheme (PREP-009), define a standard export preset from the CAD tool (view, resolution, depth and line pass), collect an approved material and style library with licence notes, and keep a few gold-standard renderings as references (PREP-016). Remove personal data such as faces or addresses from site photos where they are not needed. Tools are simple: batch renaming, the export presets of the CAD software and a shared asset folder with access rights per project (PREP-023).

Data preparation is a small share of the total effort, roughly 10 to 20 percent, because the model needs no training data. The effort mostly goes into repeatable exports and a clean material library.

Implementation paths

Which path for whom

• SMB: A, because mature rendering tools with AI features exist at SMB prices and the use is standard design work.
• Enterprise: B, because a firm with many projects can connect image APIs to its BIM/CAD pipeline and keep project data under its own control.
• Developer: C, because ComfyUI with open diffusion and ControlNet models gives full control over viewport-to-render workflows.
• Private: A, because consumer apps for room restyling are cheap and need no setup.

Path A — Buy (off-the-shelf)

• In-CAD AI render plugins: Veras (EvolveLAB, now part of Chaos) for SketchUp, Revit, Rhino, Archicad, Vectorworks and Autodesk Forma; Chaos tools with AI enhancement; Enscape and Twinmotion/Lumion/D5 Render for classic real-time rendering, increasingly with AI features (check current feature lists).
• Web tools: Gendo, Rendair, PromeAI, Mnml, Midjourney, Krea; consumer interior apps such as Reroom or RoomGPT-style tools for private use.
• Concept-phase analysis, not rendering: Autodesk Forma (sun, wind, noise, carbon studies) pairs with a renderer; see GEN-024.
• Check: where images and uploaded models are processed and stored, whether inputs are used for training, commercial licence terms of outputs, and EU hosting or a DPA. Most listed tools are US-based; EU options are rarer, so verify per vendor (mark none as EU unless confirmed).
• Choose A when: a practice wants results this week without IT work.

Path B — Build on cloud APIs

• Image APIs: Black Forest Labs FLUX models (German provider, EU-based company; API with regional options), Stability AI, Google Gemini/Imagen, OpenAI image models, Adobe Firefly (commercially safe training claim) via API; Azure OpenAI or Vertex AI in EU regions for data-residency needs.
• Pipeline: export viewport passes (colour, depth, normals) from the CAD tool via script or add-in, call a conditioned model, return results into a shared board (Miro, Figma, Notion).
• Orchestration: n8n or a small service; a VLM checks results against the brief (for example, window count, room type) and flags drift.
• Choose B when: a firm wants a branded in-house tool or integration with its own project data.

Path C — Self-host (open source)

• Models: Stable Diffusion XL or SD 3.5 with ControlNet (depth, canny, segmentation) and IP-Adapter; FLUX.1 [dev] and FLUX Kontext [dev] (non-commercial licence for the dev weights, check terms) or FLUX Schnell (Apache 2.0); Qwen-Image and Qwen-Image-Edit, HiDream (open weights, verify licence).
• Stack: ComfyUI or InvokeAI with workflow templates; Krita/Photoshop plugins; upscaling with open super-resolution models.
• Hardware: HW2 (one 24 to 32 GB GPU) runs SDXL and quantised FLUX with ControlNet at a few seconds to under a minute per image; HW3 for larger edit models and batch overnight work.
• Effort: a technically minded team member must maintain ComfyUI workflows and model updates.
• Choose C when: client designs are confidential (competitions, private residences, defence or critical-infrastructure buildings) or volume is high.

Path D — Custom training

LoRA fine-tuning on a practice's own project images or a regional style (for example, local timber-frame vernacular, a developer's corporate look) is feasible with 20 to 200 curated, rights-cleared images on one HW2 GPU. Full model training is not worthwhile. Only train on images you own or have licensed; client project photos may carry copyright and confidentiality constraints.

State of the art (2026)

• Strengths: convincing materials, daylight and mood; fast variant generation; instruction-based editing that preserves layout; plugin integration that reads geometry directly from the CAD viewport.
• Limits and failure modes: inconsistent scale and proportions; invented or missing doors, stairs and structural logic; garbled text on signs; furniture that is physically implausible; drift between successive views of the same building (no true 3D consistency); lighting that is pretty but physically wrong, so unsuitable for daylight or energy claims.
• Not a replacement for BIM or physically based rendering: for dimensionally accurate construction visuals, planning applications and daylight studies use model-based renderers; AI is the front-end for ideation and styling.
• Trend: video and 3D-aware generation (image-to-video, Gaussian-splat scenes, GEN-025) to provide consistent walkthroughs; agentic tools that propose layouts from briefs (GEN-024).
• Provenance: the AI Act and the EU code of practice on marking and labelling AI-generated content push tools towards watermarking and C2PA-style metadata [3][4].

Shared resources & combinations

• Same stack (image-gen, ControlNet, ComfyUI on HW2): GEN-001 marketing imagery, GEN-011 retouching and inpainting, GEN-018 virtual staging, GEN-016 illustration, GEN-014 product design ideation: one GPU server and one workflow library serve all.
• Upstream: GEN-024 (site planning, massing, floor plans), VIS-040 (scan-to-BIM as-built models giving real geometry), DOC-030 (quantities and cost from plans).
• Downstream: GEN-002 and GEN-025 for walkthroughs and tours, GEN-005 for presentations.

Effort, cost & KPIs

• Path A: hours to days; plugin subscriptions typically tens of euros per user and month (prices vary by source, verify) [1].
• Path B: 2 to 6 weeks for a pipeline and board integration; image API cost per render is small relative to staff time.
• Path C: 2 to 6 weeks to tune workflows; one GPU workstation plus part-time upkeep.
• KPIs: time from brief to first client-ready concept image, variants per hour, share of AI images accepted by the client, revision rounds saved, cost per visual versus external studio, share of images correctly labelled.

People & change

Involve management first for budget and a decision on which tools are allowed, then the head of design as process owner, who defines where AI renderings are acceptable (concept phase) and where they are not (approval drawings). IT clarifies accounts and hardware; customers should be told when AI visuals are shown, because renderings can promise details that the plan does not deliver. Run a pilot with two or three designers on one real project and let them present results to colleagues, as in GUIDE-009 and GUIDE-008.

Works council: usually no, because the tool processes images, not employee behaviour. It becomes relevant (depends) if the vendor logs per-user activity centrally, since the BAG treats objective suitability for monitoring under §87(1) no. 6 BetrVG as enough; inform the council early under §90(1) no. 3 in any case if one exists. A DPO is needed only when client photos with people or addresses are uploaded; then check the AVV (Art. 28 GDPR). AI literacy: support staff with a short training on prompts, limits, copyright and labelling (Art. 4 AI Act, GUIDE-006). Typical concern: designers fear being replaced; position AI as a faster sketching tool and keep the designer responsible for every image sent to clients.

Risks, regulation & compliance

• EU AI Act: typical use is not high-risk. Art. 50 transparency applies to synthetic image content: providers must make outputs machine-detectable as AI-generated, and deployers of realistic "deep fake" content must disclose it. Classified as limited; Art. 4 AI literacy applies to staff. Not an Art. 5 case. Do not present renders as official planning documents [3].
• Misleading depiction: consumer-protection and unfair-commercial-practice law (UWG in Germany) apply to sales visuals of unbuilt property; show "illustration, not binding" and keep binding specs in the building description.
• Copyright: outputs of purely AI-generated images may lack copyright protection; styles imitating a named architect or living artist raise legal risk; check training-data and output indemnities (for example, Adobe Firefly). Check the licence of open weights for commercial use.
• GDPR: photos of private homes can reveal occupants, neighbours and vehicles; faces and number plates in site photos need blurring before upload to cloud tools. Client project data may include personal data and requires a DPA and an EU or adequacy-covered processing location.
• Confidentiality: competition entries and sensitive buildings should use Path C or contractual no-training terms.
• Mitigations: label images, keep the BIM model as source of truth, review each image against the plan, store prompts and seeds in the project file.

Getting started

1. Pick one live early-phase project and collect the viewport exports and a short brief.
2. Trial one plugin (for example Veras, with a free trial) and one open workflow (ComfyUI with a depth ControlNet) on the same views; compare fidelity and effort.
3. Write a house rule: label AI visuals, state "illustration", and never take dimensions from them.
4. Set up a rights-cleared reference library and, for Path C, a shared GPU workstation.
5. Measure variants per hour and client feedback after one month, then decide on scaling.

Sources

1. Gendo — Top 10 AI rendering software for architects in 2026 — https://gendo.ai/blog/top-10-ai-rendering-software-architects-2026
2. Visioplan — AI architecture rendering: 10 tools compared — https://visioplan.io/en/blog/ai-rendering-tools-architects-comparison-2026
3. Regulation (EU) 2024/1689 (AI Act), Art. 50 — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
4. European Commission — AI Act policy page (code of practice on AI-generated content) — https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
5. ControlNet (Zhang et al.) — Adding Conditional Control to Text-to-Image Diffusion Models — https://arxiv.org/abs/2302.05543
6. Black Forest Labs — FLUX documentation — https://docs.bfl.ai
7. ComfyUI — https://github.com/comfyanonymous/ComfyUI
8. Veras by EvolveLAB — https://www.evolvelab.io/veras