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

What changed

SeFi-Team has introduced SeFi-Image, a new text-to-image foundation model that employs a "semantic-first diffusion" approach. This novel latent diffusion modeling paradigm is designed to enhance both the training efficiency and the generation quality of text-to-image models. The researchers have instantiated SeFi-Image across three distinct parameter scales: 1 billion, 2 billion, and 5 billion parameters. This scaling allows for a systematic study of how model size impacts performance and offers flexibility for deployment based on available computational resources.

A key claim is the significant reduction in training compute. The largest 5 billion parameter model was trained using only 125,000 A800 GPU hours. This represents approximately 10-20% of the compute resources reportedly used for training Z-Image, a comparable model. Despite this more modest training budget, SeFi-Image is reported to achieve performance on par with, or even superior to, existing models like Qwen-Image and Z-Image across a variety of benchmarks. These benchmarks include GenEval, DPG, LongTextBench, OneIG, and CVTG-2K.

Furthermore, the team has developed "DMD2-distilled few-step turbo variants" for each model scale. These variants are intended to address diverse hardware constraints and latency requirements, providing users with options for faster inference.

The training data pipeline for SeFi-Image involves 450 million internal image-text samples and 28 million synthetic text-rendered image-text pairs. The internal data was re-annotated using Qwen3.5-2B, focusing on accuracy, objectivity, and selective thoroughness in captions to provide cleaner supervision and bridge the gap between training and inference prompts. The synthetic data generation includes plain text rendering on solid backgrounds and structured layout rendering with diverse templates, aiming to improve the model's ability to handle text within images.

Why it matters for builders

SeFi-Image's release provides developers with a new set of tools for building text-to-image applications. The availability of models at different scales (1B, 2B, 5B) allows for choices based on deployment environments, from resource-constrained devices to high-performance servers. The reported efficiency in training compute suggests that future advancements in T2I models might be achievable with more accessible resources, potentially lowering the barrier to entry for developing and fine-tuning such models.

The inclusion of few-step turbo variants is particularly relevant for applications requiring low latency, such as real-time image generation or interactive tools. The focus on semantic guidance in the diffusion process could lead to improved prompt adherence and image quality, enabling builders to create more sophisticated and reliable visual content generation systems.

Practical impact

Developers can explore the publicly released code and weights of SeFi-Image to integrate its capabilities into their projects. Experimentation with the different model scales (1B, 2B, 5B) is recommended to determine the optimal balance between performance and resource usage for specific applications. Builders working on applications that require precise text rendering within images may find the synthetic data generation strategies and the model's performance in this area particularly beneficial.

Testing the few-step turbo variants is advised for use cases where inference speed is critical. The comprehensive benchmark results provided in the paper can guide developers in understanding the model's strengths and weaknesses across various evaluation metrics, aiding in the selection of the most suitable model variant.

Caveats and source limits

The source material is a research paper, and claims regarding performance and efficiency are based on the authors' experiments. Independent verification of benchmark results and direct comparisons with other models on diverse real-world tasks are not yet available. The exact pricing for cloud-based deployment or specific hardware requirements for the turbo variants are not detailed in the provided excerpt. While code and weights are released, the full extent of community adoption and practical deployment challenges will emerge over time. The paper focuses on the technical aspects of the model and its training, with less emphasis on user-facing application development frameworks or detailed API documentation.