Abstract
Flow-based transformer models for image generation have achieved state-of-the-art performance with larger model parameters, but their inference deployment cost remains high. To enhance inference performance while maintaining generation quality, we propose progressive rectified flow transformers. We divide the rectified flow into different stages according to resolution, using fewer transformer layers at the low-resolution stages to generate image layouts and concept contours, and progressively adding more layers as the resolution increases. Experiments demonstrate that our approach achieves fast convergence and reduces inference time while ensuring generation quality. The main contributions of this paper are summarized as follows: (1) We introduce progressive rectified flow transformers that enable multi-resolution training, accelerating model convergence; (2) NAMI leverages piecewise flow and spatial cascading of Diffusion Transformer (DiT) to rapidly generate images, reducing inference time by 40% to generate a 1024 resolution image; (3) We propose NAMI-1K benchmark to evaluate human preference performance, aiming to mitigate distributional bias and prevent data leakage from open-source benchmarks. The results show that our model is competitive with state-of-the-art models.
Introduction
High-resolution image generation has reached incredible heights with models like SD3 and FLUX, but these "Flow-based Transformers" come with a heavy price tag: massive computational costs and high inference latency. When every single pixel-token has to pass through dozens of transformer layers at every step of the generation process, hardware requirements become a significant bottleneck.
A new framework, NAMI (Bridged Progressive Rectified Flow Transformers), addresses this by decomposing the generation process into stages. It uses fewer transformer layers at low resolutions to establish image layouts and progressively adds more layers as the resolution increases, slashing inference time significantly without sacrificing aesthetic quality.
Methodology
NAMI introduces three core innovations to optimize the generation flow:
- Spatiotemporal Separation: The generation process is divided into $K$ stages (e.g., 256px, 512px, and 1024px).
- Architectural Partitioning: Instead of using the full model at every stage, NAMI uses a subset of layers for low-res stages. For a 2B parameter model, the authors utilize a 9:7:6 layer ratio across three stages, saving 11% in inference time on top of resolution cascading gains.
- BridgeFlow Module: Transitioning between resolutions requires aligning noise probability distributions. NAMI introduces BridgeFlow, a learnable linear transformation ($x = W \cdot Up(x) + B$) that ensures a smooth, data-driven transition between resolutions.
- Multi-Resolution Training: To prevent the model from "forgetting" how to handle different resolutions, a joint training strategy samples images of various resolutions, progressively shifting focus from low to high resolution as training proceeds.
Evaluation and Results
The model was trained on approximately 100 million samples, including LAION, GRIT-20M, and internal high-quality datasets.
Speed and Quality
NAMI-2B generates 1024×1024 images in 2.98 seconds on an A100 GPU, compared to 8.47 seconds for a standard baseline—a 64.82% reduction in latency. On the GenEval benchmark, NAMI-2B outperformed SDXL (2.6B) and remained highly competitive with SD3-Medium (2B).
NAMI-1K Benchmark
The authors introduced NAMI-1K, a new benchmark consisting of 1,000 prompts across 14 categories (e.g., "Human activities," "Natural scenery"). This benchmark aims to better reflect real-world usage and mitigate the distributional bias found in existing open-source benchmarks. In blind human preference tests, NAMI-2B was rated higher in "Realism" and "Overall Quality" than other 2B-class models like Infinity and SANA.
Conclusion
NAMI demonstrates that using full model depth at low resolutions is redundant. By partitioning the model and learning the "bridge" between resolutions, NAMI achieves significant efficiency gains. While it currently trails massive 12B parameter models in pure aesthetic scores, its framework is naturally extendable to video generation, where the computational burden of high-resolution tokens is even more severe.
Summary updated