Lumina-Image 2.0

Efficiency is the selling point of Lumina-Image 2.0, a 2.6B parameter flow-based diffusion transformer from the Alpha-VLLM group at Shanghai AI Laboratory that generates 1024x1024 images with far less compute than the much larger models it competes against. The architecture pairs a flow-matching DiT backbone with Gemma-2-2B as the text encoder and the 16-channel FLUX VAE, and the accompanying paper, accepted to ICCV 2025, details the captioning and training strategies behind its benchmark results. Practical integration is unusually broad for a research release: the model ships natively in Hugging Face diffusers as Lumina2Pipeline, has official ComfyUI support, offers multiple solvers including midpoint, Euler, and DPM, and extends to controllable generation, editing, and identity preservation through the companion Lumina-Accessory framework. Inference fits consumer NVIDIA GPUs, with CPU offload available to trim VRAM use further, and fine-tuning code is included in the repository. Code and weights are Apache-2.0 licensed, free for commercial use, and the repo holds about 1,000 stars.