SKY-001 plans a design through a hierarchy and a sequence of visual states. In Skyfall’s October 19 research article, cited in the source notes, the Hierarchical Composer defines the target and its parts; the Construction World Model predicts how individual elements are built. This explains the published approach—not an independently tested implementation or a guarantee of production-ready files.
Two technical stages, not four separate models.
Skyfall’s product overview presents four workflow steps: explore designs, decompose a selected design, build its parts and assemble them. The research instead names two model stages. These are different levels of explanation, not competing architectures; both descriptions are linked in the source notes.
- Explore and decompose: the Hierarchical Composer generates a target, splits it into a tree and refines the leaves. The research says it can generate candidate targets for a user to choose between.
- Build: the Construction World Model predicts ordered construction states for the elements, using their parent context.
- Assemble: the overview’s final workflow step should be read alongside the research’s stated output: the design tree plus its per-element sequences form a visual design plan. It is not evidence of a separate assembly model or a downloadable native CAD assembly.
Stage 1: decide what the design contains.
The Composer follows three operations: generate the whole, split it level by level, then refine each leaf after splitting stops. A leaf is an element that is not subdivided further in that tree.
CAD: image to part tree
The described CAD example starts with a single image. A rough 3D whole is divided into assemblies and then individual parts. The initial shape is incomplete, especially inside the object; leaf refinement fills missing geometry and adds detail. Generated hidden geometry is a prediction, not a measurement of the photographed object.
BIM: envelope to layout hierarchy
The described BIM input includes a storey envelope, wall thicknesses and open edges, plus a programme and room sizes and positions. A simple layout sketch carries the room arrangement. The Composer generates and checks a whole plan; large storeys are split, replanned and joined.
Why use hierarchy? A parent gives its children an explicit subgoal. Once a parent is fixed, children can be developed in parallel instead of extending one enormous sequential chain. Skyfall argues that this makes long rollouts tractable; it does not imply that interactions between components disappear or that every fit is verified.
Stage 2: predict the states that build each element.
For CAD, the Construction World Model first plans intermediate 3D states: the solid after each step. Each consecutive pair supplies the context for a 2D sketch describing the transition. The research illustrates extrusions and cuts; a complex part becomes a sequence of smaller geometric subgoals.
For BIM, the construction rollout is already two-dimensional. It predicts successive floor-plan drawing states, progressing through walls, circulation, rooms, doors, windows and fixtures. Here, “construction sequence” means the order of drawing the plan—not a validated schedule for constructing the physical building.
Our CAD workflow guide turns this into questions about dimensions and editability. The BIM workflow guide applies it to room relationships and building-model handoff.
Plan coarsely; restore detail afterward.
Both domains use a diffusion-based 2D rollout in two passes. First, the model predicts a rough sequence in a compressed latent space. Then a refiner redraws the states at full resolution.
The article describes a Matryoshka-style autoencoder: a state becomes a sequence of codes, with training progressively dropping the trailing codes while still requiring reconstruction. Earlier codes must carry the overall layout; later codes supply finer detail. The coarse rollout uses the leading codes rather than every detail of every canvas.
This separates sequence planning from high-resolution drawing. The source attributes its efficiency to hierarchy and compression, but neither is a proof of engineering correctness. Sharp corners restored by a refiner do not establish dimensional tolerances. Read the benchmark methodology and evidence limits before translating reported task speed into production throughput.
A visual plan is not an executable model.
| Artifact | What it represents | What it does not establish |
|---|---|---|
| Design tree | Parent–child decomposition into assemblies, parts or layout elements. | A verified bill of materials or valid assembly constraints. |
| Visual state | A predicted intermediate solid, sketch or floor-plan canvas. | A command that executes correctly in your engineering tool. |
| Visual design plan | The hierarchy together with the elements’ ordered build sequences. | An editable parametric history or structured BIM export. |
| Executable / editable model | Operations or building elements that a design system can execute and edit. | A present deliverable merely because the visual plan looks plausible. |
Read the demonstrations and future work separately.
An explicit demo limitation: the research notes that animated transitions are illustrative and that saved sketches and 3D geometry can disagree. A smooth transition therefore cannot demonstrate that the sketch executes into the displayed solid.
Use our turbofan, sports-car and welding-robot example guide to connect the published visual demonstrations to questions about part hierarchy, interfaces and the evidence a reviewer would need.
Under “What Comes Next,” Skyfall describes turning CAD sketch rollouts into executable operations with an editable feature history. Text-driven edits that update neighboring parts are also future work. For BIM, the roadmap turns plans and drawing sequences into editable walls, doors, windows and rooms.
Those statements should not be rewritten as shipped parametric CAD, Revit integration or IFC export. IFC appears as a source of building-model benchmark data; that provenance does not establish an output format. Confirm actual deliverables through the access and evaluation checklist.
Architecture questions.
Is SKY-001 simply generating a finished image?
No. The described output includes a hierarchy and intermediate construction states, rather than only a final image. That richer representation still must be distinguished from an executable engineering file.
Does the BIM rollout show how to build a building on site?
No. The published example predicts floor-plan drawing states. It does not establish site sequencing, structural approval, building-code compliance or construction safety.
Does refinement guarantee the generated part is accurate?
No. Refinement adds detail to predicted geometry or canvases. Accuracy against required dimensions, interfaces and constraints needs separate evaluation of the actual artifact.
Sources and scope.
Introducing SKY-001 is the primary source for the two-stage architecture, compression method, demonstration limitation and roadmap. The official product page supplies the four-step overview. This guide interprets those public descriptions; it does not claim access to model internals or independent reproduction.