Intent01
Start from a sketch or design direction
The process began with a human design idea: a sketch, a program direction, a facade ambition, or a massing concept that needed fast visual exploration.
Local AI visualization for architectural concept work
At B+H, I supported concept and competition workflows where teams needed fast AI-assisted visual iteration, but confidential project material could not be casually pushed through public cloud tools.
Core Move
The useful move was connecting image generation to architectural inputs: sketch, massing, linework, depth, context, and post-production judgment.
Local First
Confidential project material stayed out of public cloud image tools.
Controlled Inputs
Sketches, Rhino massing, line drawings, depth maps, masks, and site context anchored the output.
2024 Toolchain
Automatic1111, ComfyUI, Stable Diffusion, ControlNet, local models, LoRAs, and Photoshop.
Design Review
The goal was early visual exploration, not final rendering or replacing architectural judgment.
Why It Mattered Then
The difficulty was not the image. It was making the image useful.
In 2024, the workflow still had to be assembled by hand: local setup, model testing, control images, inpainting, and compositing all had to be made usable inside real design constraints.
01
The tools were not turnkey
This was before polished architecture AI workflows were common. Getting useful images meant wiring together local models, control images, prompts, masks, and manual cleanup.
02
Privacy changed the method
Because the work lived inside a company context, public cloud tools were not the right default. The workflow had to run locally and respect project confidentiality.
03
Architecture needed control
Pure prompting was too loose for design work. The useful move was to make AI follow architectural evidence: massing, perspective, linework, depth, and site context.
Workflow
A design-controlled image pipeline
The sequence matters. Each step adds constraint before the AI gets to improvise, then human judgment brings the output back into architectural communication.
Geometry02
Translate the idea into massing
Rhino gave the workflow a simple but crucial architectural base: proportion, viewpoint, site relationship, and the broad geometry the image should respect.
Controls03
Manufacture the constraints the model needs
Line drawings and depth maps were exported as ControlNet inputs. This was the technical hinge: the AI could explore atmosphere without losing the architecture.
Depth04
Add spatial hierarchy
Depth maps gave the model another reading of foreground, background, height, and mass. The more the image was constrained, the less it behaved like a random generator.
Local AI05
Iterate locally with model and prompt tests
Automatic1111 and ComfyUI made it possible to test models, LoRAs, ControlNet settings, prompts, and inpainting without sending project material to external services.
Judgment06
Composite and clean the result
The final step was not automatic. Outputs were selected, repaired, layered, and placed into context through Photoshop so the image could support a design conversation.
Tooling
The complexity makes the workflow visible.
The screenshots show the historical limitation: the workflow was not packaged yet. It had to be constructed from local interfaces, node graphs, settings, prompts, control inputs, and manual finishing.
ComfyUI
Node-based local workflow
This was not a single prompt box. It was a local graph of models, conditioning, image inputs, and output handling.
Automatic1111
Prompt, ControlNet, and generation tuning
Automatic1111 was useful for fast tests, img2img, ControlNet settings, inpainting, model selection, and keeping the workflow on a local machine.
Photoshop
Selection, repair, and compositing
The final artifact still needed design judgment: selecting useful outputs, repairing artifacts, balancing context, and composing the image for review.
Iteration Range
The value was comparison, not one perfect image.
The workflow let the team hold massing logic relatively steady while testing material, atmosphere, facade expression, lighting, and proportion across multiple outputs.
Project Range
Two design situations, one repeatable method
The value was not a single lucky image. The same pattern could move across different project types because the workflow was based on controls, not only prompts.
Project 01
Entertainment and retail concept exploration
A concept image used to test mood, scale, lighting, and public-facing atmosphere during early design iteration.
Project 02
High-rise hotel context composite
A tower concept moved from sketch and massing through controlled generation and Photoshop placement into an urban context.
Scope
Real constraints made the workflow useful.
The workflow mattered because the constraints were real: local execution, confidentiality, controlled inputs, and design judgment.
Scope
Not AI designing architecture by itself
The work is not autonomous design or final client rendering. It is a controlled visualization workflow for early design review.
Workflow
A controlled visualization layer for early design
The useful work was translating immature AI tools into a local workflow for architects: private, constrained, iterative, and tied to design intent.
Operating Pattern
Emerging tools become valuable when the workflow respects the work.
The specific 2024 methods will keep aging, but the pattern still matters: understand the design problem, respect confidentiality, build the missing workflow, test it against real project constraints, and communicate the result clearly enough for a team to use.