LlamaGen
The LlamaGen project provides innovative image generation capabilities using autoregressive models for both text and class-conditional scenarios, demonstrating significant advancements over traditional diffusion techniques. The project delivers image tokenizers, models scaling from 100M to 3B parameters, and utilizes vLLM for substantial speed improvements during serving. Models are available via online demos and a serving framework maturing image creation efficiency by 300% to 400%, with continuous updates and resources supporting modern AI image generation exploration.
