* This blog post is a summary of this video.

Using AI to Recreate and Control Architectural Renderings

Author: Agile BIMTime: 2023-12-29 00:30:00

Table of Contents

Recreating an Old Architectural Perspective with AI

I recently experimented with using AI to recreate an old architectural perspective rendering that I had previously created manually. As an architect, I used to create these types of renderings by exporting files from AutoCAD into Photoshop. This was a time-consuming process that also didn't allow for quick iterations or changes late in the design process.

With recent advances in AI image generation, I was curious to see if I could use text prompts to recreate one of my old renderings. I decided to test out the DALL-E integration in ChatGPT to assess the accuracy and level of control possible with current AI tools. Overall, the rendering was decent but lacked precise control, indicating limitations but also potential useful applications if combined properly with 3D modeling tools.

Describing the Reference Image to ChatGPT

I started by showing ChatGPT the original architectural perspective image I wanted to recreate. Without even enabling the DALL-E plugin, ChatGPT was able to describe key elements of the photo including the wood slab ceiling, pendant lighting, furnishings, and activity of people within the space. This demonstrates ChatGPT's strong visual recognition capabilities, as it picked up on both high-level semantic aspects as well as some finer architectural details. While not perfect, having ChatGPT describe the image provided a useful starting point for generating the AI rendering prompt.

Generating Initial AI Renderings

I then enabled the DALL-E 3 plugin in my ChatGPT conversation to start generating AI images. My initial prompt described wanting to recreate the school cafeteria image but from the perspective style of an architectural competition rendering. I also included approximate dimensions and key details like the wood slab ceiling and pendant lighting. The first set of four images were decent approximations but lacked precision in elements like the lighting, furniture layout, and ceiling structure. However, being able to generate multiple rendering variations quickly shows the potential efficiency benefits of leveraging AI in early concept visualization.

Iterating to Improve the Renderings

To try refining the rendering, I provided additional feedback and modification requests on the third image I liked best. I asked for changes like making the space dimensions smaller, tweaking the ceiling to be more waveform-like with depth variation, and reducing the number of light fixtures. ChatGPT took this feedback and automatically incorporated it with the original prompt to generate another set of four renderings. The latest images are improved, particularly the setting, layout and lighting. However, fine details like the ceiling structure still lack accuracy and it's impossible to precisely control every element with just text prompts.

Assessing Accuracy and Control of AI-Generated Architecture

While being able to quickly generate architectural perspective renderings with AI represents an exciting efficiency improvement, assessing the accuracy and level of control is critical.

Upon evaluating multiple iterative image generations aimed at recreating my reference rendering, a few key limitations emerged:

First, while broad semantic aspects like furniture and figures can be described to a fair level of alignment, finer architectural details lack precision no matter how descriptive the text prompt. For example, accurately conveying ceiling connections and lighting fixtures remains challenging.

Second, dimensions and spatial relationships of elements can be suggested but not pinned down to exact measurements or alignments critical for professional renderings.

Finally, having full control to pivot or tweak any element at will to create iterative design variations is not yet possible with AI image generation alone. The rendered outputs remain somewhat randomized based on the AI model’s training distribution.

Potential Applications and Limitations of AI in Architecture

Based on these assessments, AI-generated architectural renderings show both promising applications if used judiciously as well as inherent limitations to overcome:

Using AI to Quickly Generate Design Ideas

During early concept design phases focused on testing overall spatial adjacencies, ambiance and loose furnishings layouts, AI renderings can provide quick visualizations to evaluate high-level options. The efficiency of generating multiple rendered variations from text descriptions alone frees up designer time to focus on more strategic creative exploration rather than manual production time.

Lack of Precise Control Over AI Renderings

However, fundamental imprecision with spatial relationships and architectural detailing inherent to AI generation methods hinders rigorous professional application for presentations or construction. Assessing fidelity across iterative renders reveals little improvement in accuracy despite additional prompting feedback. Precision ultimately requires complementary 3D modeling tools.

The Future of AI in Architectural Workflow and Rendering

In the near term, the largest value of AI in architecture will likely come from hybrid workflows pairing generated images with underlying 3D coordinates and parametric modeling tools.

If spatial constraints from a BIM model could connect with an AI engine capable of interpreting and visualizing geometric constraints, rapid design iterations become possible while maintaining needed precision.

Fully replacing traditional techniques remains challenging but complementing existing pipelines shows promise and could significantly enhance conceptual architect and client conversations through quick, mass options exploration.

Conclusion and Next Steps for Exploring Architectural AI

In the test case of recreating one of my old renderings with the latest AI, impressive capability was demonstrated but limitations around accuracy and control reinforce the need for pairing with underlying precision 3D models.

As architectural AI continues advancing rapidly, I’m looking forward to running more experiments assessing integrations with various 3D design tools to find the right solutions that enhance architect productivity without sacrificing quality.


Q: How was an old architectural perspective recreated with AI?
A: An old hand-drawn perspective was described to ChatGPT in natural language. ChatGPT then generated a prompt to create similar renderings using the DALL-E AI image generator.

Q: What was the accuracy of the AI architectural renderings?
A: The AI renderings approximated the original perspective but lacked precision in elements like the ceiling design and lighting placement.

Q: What are some applications of AI in architecture?
A: AI shows promise for quickly generating design ideas, material studies, and photorealistic renderings to enhance architectural workflow.

Q: What limitations exist with current AI architectural capabilities?
A: It remains challenging to exert precise control over AI renderings through text prompts alone.