Published research on AI for interior design covers two kinds of systems: models that arrange furniture in a room from a floor plan, such as ATISS (2021), and language-guided generators such as Holodeck (2023) that build 3D rooms from a text prompt. Neither is a licensed design service, and both were built mainly for research on 3D scenes. A separate rule applies to marketing: since January 1, 2026, California real estate licensees who use AI-altered images to advertise a property for sale must disclose the change and point to the original image.
Key takeaways
- ATISS generates a furniture layout given only a room type and floor plan, and was evaluated on four room types in the 3D-FRONT dataset.
- 3D-FRONT holds 18,968 furnished synthetic rooms and 13,151 textured furniture objects, with layouts drawn from professional designs.
- Holodeck asks a large language model for spatial constraints, then optimizes object placement; its stated goal is training embodied AI agents.
- California Business and Professions Code section 10140.8 covers images altered with photo software or AI, including furniture, flooring, and views.
- Plausible layouts in research datasets are not a check against building codes, accessibility needs, or a client's budget.
Layout generation: what ATISS does
ATISS, short for Autoregressive Transformers for Indoor Scene Synthesis, was described in an October 2021 paper. Given a room type and a floor plan, it produces a furnished layout. It treats a room as an unordered set of objects instead of a sequence, which lets the same trained model complete a partly furnished room, rearrange chosen items, or suggest objects. Training used only labeled 3D bounding boxes as supervision. The authors report that it produced more plausible layouts than earlier methods on four room types and ran up to eight times faster.
Plausible, here, is a statistical statement. The model learns which objects tend to appear together and where, based on example rooms. The abstract does not claim that layouts meet clearance distances, door swings, or a household's habits.
The dataset behind the layouts
ATISS was evaluated on 3D-FRONT, a repository released in 2020 and revised in 2021. According to its authors, it contains 18,968 rooms furnished with 3D objects, and 13,151 furniture objects with high-quality textures. Floor plans and layout designs come from professional creations, while furniture style, color, and texture were curated with a recommender system to look consistent. The scenes are synthetic, so they show how rooms are drawn in design software, not how finished homes are used.
Language-guided generation: Holodeck
Holodeck, posted in December 2023 and revised in April 2024, takes another route. A user-supplied prompt, for example an apartment for a researcher with a cat, goes to GPT-4, which supplies knowledge about what the scene might contain and generates spatial relations between objects. The system fills the scene with assets from the Objaverse collection, then optimizes the layout to satisfy the constraints. In a large human evaluation, annotators preferred Holodeck to manually designed procedural baselines for residential scenes. Its stated application is training embodied AI agents to navigate new spaces, not advising homeowners.
Comparing the two research approaches
The table summarizes what each paper describes. It is a simplified orientation drawn from the abstracts, not a ranking, and the systems answer different questions.
| System | Input | Output | Stated purpose | Not claimed |
|---|---|---|---|---|
| ATISS (2021) | Room type and floor plan | Set of furniture objects with positions | Interactive 3D tools, data synthesis for training and simulation | Code compliance, cost, or client fit |
| 3D-FRONT (2020, revised 2021) | Not a model; a dataset | 18,968 furnished synthetic rooms | Benchmark for scene and texture synthesis | Real homes in use |
| Holodeck (2023) | Text prompt | 3D environment from large model constraints | Training embodied AI agents | Advice to homeowners |
When AI edits a listing photo: California's rule
Assembly Bill 723 was approved by the governor on October 10, 2025, as Chapter 497, and adds section 10140.8 to the Business and Professions Code. It applies to a real estate broker or salesperson, or a person acting on their behalf, who includes a digitally altered image in an advertisement or promotional material for the sale of real property. The advertisement must carry a statement that the image was altered, plus a link, URL, or QR code to the original, unaltered image. The statement must be reasonably conspicuous and on or next to the image.
The statute defines a digitally altered image as one changed by photo editing software or artificial intelligence to add, remove, or change elements such as fixtures, furniture, appliances, flooring, walls, paint color, landscape, floor plans, or views through windows. Adjustments to lighting, sharpening, white balance, color correction, angle, straightening, cropping, or exposure are excluded if they do not change how the property is represented. When a website is under the licensee's control, the unaltered versions must be included in the posting, which may be done by link. The bill's digest notes that willful violations of the Real Estate Law are a crime.
The rule covers California licensees. Rules elsewhere, including in Canadian provinces, were not verified for this page, and this is general information, not legal advice.
What these sources do not show
No source reviewed here measures whether AI-generated room plans save money, shorten projects, or satisfy clients, so no such figures appear. The research papers also stop short of work that licensed designers do, such as reading building codes and coordinating with contractors.
FAQ
Can AI design a room layout from a floor plan?
Research systems can. ATISS generates object arrangements from a room type and floor plan, but its authors evaluate plausibility against a dataset, not safety rules or personal preferences.
Does AB 723 ban virtual staging?
No. It requires a disclosure and a way to see the original image when a licensee advertises a property for sale with an altered image.
Does ordinary photo cleanup need a disclosure?
Not under the statute's text. Lighting, white balance, cropping, straightening, and similar edits are excluded when they do not change the representation of the property.
Is Holodeck a tool for homeowners?
Its paper presents it as a system for generating 3D environments, with an application in training embodied AI agents.
Related reading
- How artificial neural networks are applied
- AI design and what it implies
- AI in conceptual design work
- AI in the hospitality industry
More in AI Tools and Industry.
Last verified: October 2026
Sources
- ATISS: Autoregressive Transformers for Indoor Scene Synthesis, Paschalidou et al., arXiv, October 2021
- 3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics, Fu et al., arXiv, 2020, revised 2021
- Holodeck: Language Guided Generation of 3D Embodied AI Environments, arXiv, December 2023, revised April 2024
- Assembly Bill 723 (Pellerin), Chapter 497, Statutes of 2025, chaptered text, approved October 10, 2025