viiika/Meissonic

What it solves

Meissonic addresses the challenge of generating high-resolution images efficiently on consumer-grade hardware. It revitalizes the Masked Generative Transformer (MGT) approach to provide a non-autoregressive alternative to traditional text-to-image synthesis, reducing the computational burden typically associated with high-resolution output.

How it works

Meissonic uses a non-autoregressive mask image modeling approach. Unlike autoregressive models that generate pixels or tokens sequentially, Meissonic predicts multiple parts of the image simultaneously through a masking process. It supports high-resolution synthesis (up to 1024x1024) and can be optimized further using FP8 quantization to reduce memory usage and maintain performance on consumer GPUs.

Who it’s for

This project is designed for researchers and developers working with generative AI, as well as artists and creators who want to run high-resolution text-to-image generation and image-to-image tasks (like inpainting and outpainting) on their own hardware.

Highlights

• High Resolution: Supports image generation up to 1024x1024 pixels.
• Consumer GPU Friendly: Optimized for consumer graphics cards, with memory usage as low as 8.7GB using FP8 quantization.
• Versatile Applications: Supports text-to-image, image-to-image, inpainting, and outpainting.
• Training Support: Includes scripts and guidelines for training the model on custom datasets.