AI Architectural Rendering

The structure and composition of prompts are critical in achieving precise and consistent AI-generated architectural images. By breaking down prompts into specific components—View, Subject, Landscape, and Environment—users can better control the output, ensuring the generated visuals align with the intended design goals. The order in which these elements are specified directly impacts the image generation process, with certain elements taking precedence in shaping the final output. A clear and structured prompt not only refines the result but also fosters a habit for users to think systematically about their input, leading to more efficient and accurate iterations.

Query Builder

I helped develop the Query Builder plugin, a tool that provides a structured approach to crafting prompts for AI-generated architectural visuals. It guides users through the above four key categoriest—offering detailed suggestions to inspire and streamline their creative process.

ControlNet Integration

I helped develop the ControlNet input system, enabling more detailed and precise control over image generation. This significantly improved the accuracy and quality of architectural visualizations by allowing specific design elements to be more accurately represented.

Checkpoint & LoRA Comparison

I organized our collection of Checkpoint and LoRA models, testing them with consistent prompts to compare their outputs. From this, I compiled a user guide summarizing key differences and optimal parameter settings to help users choose the right model for their needs.

Art Style LoRA Training

Minimum training sets were identified for each specific art style, with prompts tailored to ensure style consistency. LoRA performance was tested across various checkpoint models and LoRA weights to find the optimal model for generating stable and distinct artistic styles.

Architectural Function Training

Models were trained for specific architectural functions, such as patient rooms, by evaluating their ability to generate detailed features. Training parameters were refined based on learning results, and example prompts and styles were created for users to implement.