Paper 2604.18168
Extending One-Step Image Generation from Class Labels to Text via Discriminative Text Representation
- Published
- Apr 2026
- Research lab
- Independent
- Citations
- 2
- GitHub
- 109 stars
01 In brief
Summary
This paper extends MeanFlow, a one-step generation framework originally designed for class-label conditioning, to flexible text-to-image (T2I) generation.
The authors find that directly integrating LLM-based text encoders into MeanFlow with standard training fails, and they identify that high-quality text representations must possess strong semantic discriminability and disentanglement to work under limited denoising steps.
They validate these properties through retrieval and subsequence similarity experiments, showing the BLIP3o-NEXT text encoder excels.
They then adapt MeanFlow to the BLIP3o-NEXT model, introducing a modified temporal embedding that encodes interval length and end time, and train it on 170,000 samples.
The resulting model, EMF, achieves a GenEval score of 0.90 with 4 steps, nearly matching the 30-step baseline of 0.91, and outperforms other distilled models.
It also shows stable convergence and benefits from additional steps, unlike consistency-distilled models.
The work provides the first successful extension of MeanFlow to text conditioning, offering a practical reference for future research.
02 From the paper
Abstract
Few-step generation has been a long-standing goal, with recent one-step generation methods exemplified by MeanFlow achieving remarkable results. Existing research on MeanFlow primarily focuses on class-to-image generation. However, an intuitive yet unexplored direction is to extend the condition from fixed class labels to flexible text inputs, enabling richer content creation. Compared to the limited class labels, text conditions pose greater challenges to the model's understanding capability, necessitating the effective integration of powerful text encoders into the MeanFlow framework. Surprisingly, although incorporating text conditions appears straightforward, we find that integrating powerful LLM-based text encoders using conventional training strategies results in unsatisfactory performance. To uncover the underlying cause, we conduct detailed analyses and reveal that, due to the extremely limited number of refinement steps in the MeanFlow generation, such as only one step, the text feature representations are required to possess sufficiently high discriminability. This also explains why discrete and easily distinguishable class features perform well within the MeanFlow framework. Guided by these insights, we leverage a powerful LLM-based text encoder validated to possess the required semantic properties and adapt the MeanFlow generation process to this framework, resulting in efficient text-conditioned synthesis for the first time. Furthermore, we validate our approach on the widely used diffusion model, demonstrating significant generation performance improvements. We hope this work provides a general and practical reference for future research on text-conditioned MeanFlow generation. The code is available at https://github.com/AMAP-ML/EMF.