Lost in data (2024)
This experiment started when I got my hands on some data (duh). Data in question is the catalogue from the DesignMuseumGent. Plowing through carefully crafted descriptions and professional photos that come with it, I started by finetuning a flux.dev diffusion model, based on a subset (chairs) of the data.
I used a script to create a dataset for finetuning, matching a picture with its description, using 'llava:34b' for translating the Dutch to English, better suiting the flux model. After handpicking about 33 good image/text pairs I started the finetuning.
Data in Diffusion
I also used the image to text capabilities of the same model to generate a description of the image from the databases.. and then I started to use these descriptions to generate an image again, with the LoRa to make it look lkike it could be part of the collection.
Lost in Diffusion
some cherry picked examples from the first batch...I'm not (yet) elaborating on the aspects that are (not) intersting, but it's a worthy experiment...what info is lost when you let a human describe an object from a picture and use that description to let an AI model generate an image using a text2img model. The Flow is Text > Translated Text > Image
Schaalmodel van de zetel 'Capitello'
Losing my Diffusion
for the next iteration, I used an image2text model to describe the original image and generated a new image from that description. so the flow is Image > Text > Image
Zetel uit de eigen woning van de ontwerper