Qanvas (PAUSED TO FOCUS ON QED-BASE-V3)

Purpose of Research

Development of a text-to-image model for mobile devices that provides good quality of the resulting image while reducing memory consumption and computing power required.

The aim is to design the solution that can comprehend the meaning of prompts and generate images using knowledge-based representations instead of memorizing templates.

Current Status

• QED-Base-V1 Tokenizer (REUSED): Completed

• QLIP: Completed

  • Lightweight CLIP-style text encoder trained from scratch
  • The purpose of the model is to provide semantic alignment between texts and images in an resource-efficient way

• Q-VAE: Completed

  • Custom VAE trained from scratch to enable image latents compression/reconstruction

• Qanvas Generator: Failed-Planning something new.

  • Latent image generator conditioned on QLIP embeddings