✦
Summary
Lumina-mGPT 2.0 is a stand-alone autoregressive model that unifies a broad spectrum of image generation tasks into a single framework, built from scratch with a decoder-only architecture that applies the sequential prediction principles of large language models to image pixels and tokens.
Key Features
Stand-alone decoder-only autoregressive model for unified image generation tasks
Supports text-to-image, image pair generation, multi-turn editing and dense prediction
Unambiguous Image Representation (Uni-Rep) for flexible aspect ratios and styles
Multi-turn conversational image refinement and editing
Dense prediction including segmentation, depth estimation and pose estimation
Open-source with inference code, checkpoints, fine-tuning scripts and technical report
Acceleration techniques including speculative decoding and model quantization
About Lumina-mGPT 2.0
It handles text-to-image generation, image pair generation, subject-driven creation, multi-turn image editing, controllable generation and dense prediction including segmentation, depth estimation and pose estimation, using Unambiguous Image Representation (Uni-Rep) to generate images of varying aspect ratios and styles with high fidelity, and it supports multi-turn interactions so users can iteratively refine images through conversational prompts. The open-source release includes inference code, pretrained checkpoints, fine-tuning scripts and a technical report under an Apache 2.0 license, with acceleration strategies such as speculative decoding and model quantization.
Compared with diffusion-based models it offers advantages in world knowledge comprehension, controllability and task unification, though it may be slower and requires substantial computational resources with high-VRAM GPUs for efficient inference, and it is aimed at AI researchers and developers working on creative and research applications.