Project Objective
The project objective is to develop an interface where users can batch upload images of an established design, usually in the Design Development phase when materials and small tweaks are being considered. The goal is for a user in a traditional architecture office to batch upload images of their 3d model to quickly iterate options in a consistent fashion.
The project also incorporates pretrained LORAs. The intention is that in an office, they would train LORAs of images based on their preferences, and they can build and refine a database over time.
Interface
Hosted in Gradio with Google Collab, in the interface the user can upload multiple images, select a pre-trained LORA, and write a prompt for intended to be generated. As well, the user can control with sliders inputs such as the Lora Guidance Scale, Interference, and the depth scale.
Following that, images can be generated, including depth maps for it. The user can browse through the images in the interface, preview them, download the images individually, or download in a batch.
For this project, we trained 2 LORAs. One, consisting of red brick buildings and the second one, consisting of a more organic, bohemian, architecture in the jungle.
Example of an image developed during Design Development and rendered through a prompt in the generator
Batch Input Images
Generated Images Iterations
This is an example of an output iteration. Some of the advantages of the workflow include scalability in terms of having this workflow apply to many images. Graphical consistency, and the speed of the process. One aspect that was not successful and would be one of the next steps in refining, would be the image consistency in terms of the final output, as the given inputs where interpreted differently across, contradicting the intention of total consistency and speed.
Conclusion & thoughts
Include more fine-tuning options to explore image generation in a more sophisticated way
Incorporate canny into the workflow as depending just on depth limits the workflow of image generation.
Standardize the initial input process to achieve image consistency and uniformity
Realize that batch uploading and generating images using the same prompt and setting might, as an idea, needs technical refinement as these initial variables apply different to each image.
