viiika/Meissonic

[ICLR 2025] Official Implementation of Meissonic: Revitalizing Masked Generative Transformers for Efficient High-Resolution Text-to-Image Synthesis

What it solves

Meissonic addresses the challenge of generating high-resolution images efficiently. It provides a way to synthesize images up to 1024x1024 pixels while remaining lightweight enough to run on consumer-grade graphics cards (GPUs).

How it works

Meissonic is a non-autoregressive masked image modeling transformer. Unlike traditional autoregressive models that generate pixels or tokens one by one, it uses a masking approach to synthesize images. It supports text-to-image and image-to-image tasks, and can be further optimized using FP8 quantization to reduce memory usage and maintain performance.

Who it’s for

This project is for AI researchers, developers, and digital artists who want to generate high-quality, high-resolution images on their own hardware without needing industrial-scale compute resources.

Highlights

  • High Resolution: Supports image generation up to 1024x1024.
  • Consumer GPU Friendly: Optimized for accessibility on home hardware.
  • Versatile Capabilities: Supports text-to-image, inpainting, and outpainting.
  • Performance Optimization: Includes FP8 quantization support to lower VRAM requirements to approximately 8.7GB.