Project idea
I wanted to learn how to neuralize the Standard Diffusion to generate interior concepts in my style (II and visualization of projects) while preserving characteristic colors, textures and combinations. In the process of creating the first generation, I thought it would be interesting to try to collect some successful images into a whole project. The basic idea is to create a hotel interior concept based only on the collected images of the trained model. The main objective is to acquire applied knowledge that can be applied in the initial stages of project concepts.
Several images from the database (my generation and visualization)
Total generators
The model was able to convey characteristic lighting from my generators, project materials and a general impression of ease and tranquillity. When I created the hotel’s interiors, I tried to change the original prompt as little as possible in order to preserve the style, and I also divided the main areas by color.
In yellow, I marked the hotel lobby space, which is the most welcome, warm, and brightest. /prompt: photo of a hotel lobby interor in CROSBY style, reception desk, White and Light yellow colors
Reception and lobby area
The blue color chose a gallery and a restaurant for the public spaces. /prompt: photo of a interor in CROSBY style, white Wall, light blue floor, photographed by Hiroshi Sugimoto
Gallery space
Pink color for hotel rooms, private spaces. /prompt: photo of a cafe interior in CROSBY style, light blue colors / light pink colors
Restaurant and dining area in room
In the room and in the gallery space, I wanted to achieve a «other angle» effect, so I added to the prompt an interior photographer known for front-end filming. /prompt: photo of a hotel bedroom interor in CROSBY style, light pink colors, photographed by Hiroshi Sugimoto / prompt: photo of a interior in CROSBY style, sofa, journal table, light pink colors, photographed by Hiroshi Sugimoto
Hotel room
Learning process
First, I downloaded everything I needed for the library: Diffusers and DreamBooth. She then downloaded the images from the folder on Google Drive on Google Colab. It’s done through a colab-based drive function. Additional formatting of images (cut to square ratio) is also done by code. Then I check that the images are actually uploading, and I put a few of them on the screen using image_grid.
Next to all the images, the code automatically generates signatures (captions) using the BLIP model from Salesforce. This is also the production stage of the dataset: each image receives a descriptive text so that the model understands what is on it.
Another training preparation, I’m putting access token on the Hugging Face account.
This part of the code already indicates model training. The 512 resolution was selected, and the maximum training step was set at 500, and the checkpoint was set at 250. This reduces the time for the training process (for example, it took 30 to 35 minutes).
Then I save a trained model on Hugging Face and create the first generators.
After creating the first trial generators, she tried to create 20 images of different spaces («gallery», «hotel lobby», «bar», «cafe», «restaurant», «office», «shop», «art centre», «museum») using a code that framed the word by listing them each other. The rest of the prop didn’t change. In general, there are already the dominant colors of my work, smooth shapes, paste shades, soft lighting. /prompt: photo of a {place}interior in CROSBY style
Promptu generation with space listing
And then I decided to do a little more detail on space and color gamma, and I also tried to identify the necessary objects. That’s the approach I left for the final generators. Image patterns are as follows: /photo of a interor in CROSBY style, White Walls, Light Blue Accents, Captured by Hiroshi Sugimoto / sofa in CROSBY style, sofa, purple fabric, Polish metal details, glass, pink colors, White backrund furniture division / photo of a home interor in CROSBY style, recaption desk, Polish metal details / photo of a interor in CROSBY style
Prompt generators with supplements