Temporal Equilibrium MeanFlow: Bridging the Scale Gap for One-Step Generation

Kong 2 ByteDance, Seed https://temf.github.io Figure 1. One-Step Generation of 256×256 ImageNet Images from Scratch. Our model delivers significantly superior generation quality compared to prior state-of-the-art one-step diffusion and flow methods. Specifically, iCT [30], Shortcut [8], MeanFlow [10], and our method all achieve generation with 1 network forward evaluation (1-NFE), whereas the 1-step result of IMM [41] requires 2-NFE for guidance. Detailed quantitative results are provided in Table 2, and all images presented herein are generated by our 1-NFE model.