Featured project · architectural AI systems
A controllable visual-computing studio for architectural spaces: edit a room, retrieve the furniture that inspired it, and carry the result into 3D. I built the AI work presented here—ComfyUI workflows, the recommendation engine, Kaggle operations, documentation, and presentations.
Edited roomOriginal room
Drag the frame. This is the visible surface of a larger system: task routing, conditioning, sampling, retrieval, and export are separated into inspectable paths.
Interior visualization asks for different kinds of control: furnishing an empty shell, replacing one object, applying a reference style, changing time of day, cleaning a room, or preparing an asset for 3D. The system turns those intentions into explicit routes instead of asking an operator to rebuild a graph from scratch.
A large shared graph sits above the focused exports. Rail-Switch mode selection, controller branches, fast bypasses, scene understanding, conditioning, reference injection, and a shared KSampler/output route make the pipeline understandable as a system. The graph is wide because the work is wide; the reader can pan across it and then enter the five guided regions below.
Mode selection and rail switches decide which work happens. The switch routes the prompt, image, and controls down one branch and mutes the rest, so each mode pays only for its own path.
Click to inspect
Controller branches and fast muters skip inactive sections. A focused task travels a short path instead of dragging the whole studio behind it.
Click to inspect
Florence-2 and SAM2 turn the input into an edit surface: detection finds what is in the room, segmentation carves out exactly where an edit may land.
Click to inspect
Canny, depth, inpaint, and M-LSD conditioning are selectable per task. Structure, geometry, or masked repair each get the guidance they need — and nothing they do not.
Click to inspect
Reference inputs flow through controlled injection into a shared KSampler and output route, so every branch exits through one calibrated delivery path.
Click to inspectEvery job below rides the same universal graph, yet each enters with different inputs, obeys different constraints, and leaves as a different deliverable. That is what makes them capability families rather than interchangeable style presets — so each one gets its own record here.
Populate a bare or badly styled shell with coherent, high-end furnishings while the architecture itself stays untouched.
Inject specific commercial products into a shell or a photograph so the render sells a real catalogue item, not an approximation.
Edited roomOriginal room
Regenerate only a masked sub-region — a bed, a carpet — while the rest of the room keeps its perspective, lighting, and clutter.
Routed through the universal graph — branch switches and rails in §01.
Re-render one space in a different visual language: full colourisation of a grey-box blockout, or coastal daylight, anime, and block-world presets over a photograph.
Remove scattered clutter, then re-grade time of day, lamps, curtains, or weather — while every piece of furniture stays exactly where it was.
Turn hand-drawn line art straight into a photorealistic render, without assembling a multi-ControlNet stack.
Resolve the micro-noise and texture seams that diffusion leaves on large smooth surfaces before anything ships.
Move a matched product from pixels to geometry: mesh, texture, and a GLB that survives a realtime scene. The full route is documented in §05.
The older SD1.5 path made structure explicit through Canny, M-LSD, depth, ControlNet, masks, and IP-Adapters. The newer Flux2Klein path simplifies the edit route with native image-to-image conditioning. Both belong in the story, but they answer different engineering problems.
The recommendation engine connects visual generation to a catalog. A selected crop becomes both a vector and a structured description, then a local orchestrator makes the model’s vocabulary searchable and stable.
The crop travels in parallel: SigLIP2 creates a normalized 1,152-dimensional visual embedding while Qwen3-VL produces attributes and a description. The orchestrator snaps free-form tags to a canonical taxonomy, writes vectors plus metadata to ChromaDB, filters by color/material/style, and falls back to visual similarity when filters become too restrictive.
Running a visual system is part of the engineering. The Kaggle work turns a fragile setup into a recoverable operating surface: restore snapshots, redirect caches, install only what is needed, repair custom nodes, and isolate failures by stage.
Once the engine can identify the right product, the workflow continues into a Hunyuan3D path. Background removal, normalization, SDF inference, Marching Cubes, cleanup, multiview rendering, and texture repair turn a 2D reference into a textured GLB that belongs back inside a spatial design tool.
The output is judged by more than geometry. It needs a clean surface, coherent views, useful texture, and a format the product can load.
IntelliRoom is not a model demo floating outside a user workflow. The visual systems connect to a marketplace, design workspace, synchronized 2D/3D planner, PBR materials, and camera tools. My contribution was the attached AI and workflow layer; the surrounding product was built as a team system.
The planner’s engineering details matter to the AI handoff: synchronized 2D/3D views, 150 mm wall snapping, a 100 mm grid, collision checks, and viewport-aware prompting make generated content usable in a real spatial interface.
The most valuable parts of this project are the boundaries: where a model ends, where a deterministic rule begins, and which system behavior deserves an operator-visible control.
Free-form model tags are useful for language, but catalog search needs a fixed vocabulary. Known aliases map first; unknown values snap to the nearest allowed taxonomy value.
The universal graph makes bypasses visible. An operator can see whether a task used segmentation, structural conditioning, reference injection, or a shorter route.
Snapshots, cache routing, dependency repair, staged launch, and recovery controls turn notebook research into something another person can actually run.
The narrative keeps the main path visual. The full record remains one click away for reviewers who want to inspect the original system, the complete presentation, or the thesis-level product context.
Generative workflows, recommendation, 3D assets, and the complete project story.
Focused editing paths, reference behavior, and capability examples.
Product architecture, planner, backend, AI workflows, and implementation context.
