Skip to content

text to vector to text to image

training data

  • lot's of data
    • laion-5b dataset
      • 5.85 billion image-text pairs.
    • train the entire model?
      • a hard drive of 240TB
      • 32 x 8 x A100 GPUs
      • cost: approx $ 600,000
      • carbon cost: 11,250 kg CO2
        • = ± 2.5 cars driving 15,000 km/year

note: if you rent a special PC with massive specs it will still take over a week to just download the data

training on data

the latent space, a gut-feeling approach:

model ==> encoding (a vector/coordinate):

src

latent space: cats and dogs

note: dichtheid op landkaart = vectoren die dicht liggen bij elkaar ==> afbeeldingen die veel gelijkenis vertonen enter LEXICA.art ==> zoeken naar LAIKA in space

latent space: cats and dogs

latent space: cats and dogs

note: our AI encoder gave us a sequence of numbers, so we can position the dog on the landmap

latent space: cats and dogs

latent space: cats and dogs

latent space: cats and dogs

latent space: cats and dogs

latent space: cats and dogs

latent space: cats and dogs

latent space: cats and dogs

latent space: cats and dogs

latent space: cats and dogs

latent space: IN BETWEEN cats and dogs

Different diffusion AI models

DALL-E2 MidJourney Stable Diffusion ...