SeFi-Image: A Text-to-Image Foundation Model with Semantic-First Diffusion

A research team has introduced SeFi-Image, a text-to-image foundation model that uses a semantic-first diffusion approach to match or exceed the performance of larger-scale competitors while using a fraction of the training compute [1][2]. The model, detailed in a paper submitted to arXiv on June 21, 2026, is built on a novel latent diffusion modeling paradigm the authors call semantic-first diffusion [1][2]. The researchers instantiated SeFi-Image at three scales: 1 billion, 2 billion, and 5 billion parameters, allowing for a systematic study of scaling behavior and flexible deployment under different compute budgets [2]. The largest 5-billion-parameter variant was trained with 125,000 A800 GPU hours, which the authors state is roughly 10-20% of the training compute used by the competing Z-Image model [1][2]. Despite this modest training footprint, the model achieves results comparable to or even superior to Qwen-Image and Z-Image [1][2]. The paper reports strong performance across a broad suite of benchmarks, including GenEval, DPG, LongTextBench, OneIG, and CVTG-2K [1][2]. The work contrasts with prior efforts to use semantic guidance for training acceleration, which the authors note were limited to simple datasets like ImageNet, low resolutions, and small-scale models [1][2]. To address diverse hardware and latency requirements, the team also provides DMD2-distilled few-step turbo variants for each model scale [1][2]. The project’s code and model weights have been publicly released, with the authors expressing hope that the work will offer useful insights into semantic-guided diffusion modeling for text-to-image generation [1][2]. The submission, led by Jinming Liu, was updated on June 23, 2026, with a revised manuscript size of 35,169 KB [1].