NAMI: Efficient Image Generation via Bridged Progressive Rectified Flow Transformers

Conference: CVPR 2026
Paper: CVF Open Access
Code: Not mentioned
Area: Diffusion Models / Image Generation
Keywords: Rectified Flow, Progressive Generation, Multi-resolution Training, Inference Acceleration, DiT

TL;DR

NAMI partitions the rectified flow of text-to-image generation into multiple time windows based on resolution. Low-resolution stages utilize fewer Transformer layers to rapidly construct layouts, while high-resolution stages gradually stack layers for detail refinement. A learnable BridgeFlow module aligns distributions between adjacent stages. At a 2B parameter scale, it reduces inference time for \(1024 \times 1024\) images by 64% while maintaining quality comparable to state-of-the-art models.

Background & Motivation

Background: Text-to-image models represented by SD3 and FLUX achieve leading generation quality using rectified flow and MM-DiT architectures. Rectified flow connects noise and data via linear trajectories, offering higher efficiency in training and inference than traditional diffusion. MM-DiT concatenates text and image tokens for joint attention.

Limitations of Prior Work: While quality has improved, parameter counts and computational overhead have surged. For instance, FLUX reaches 12B parameters, requiring full-load computation across all denoising steps, tokens, and model layers to generate a high-resolution image, leading to prohibitive inference latency and cost. Existing acceleration strategies (e.g., high-ratio VAE downsampling, token reduction, linear attention) often sacrifice generation quality.

Key Challenge: Current methods apply uniform denoising across all sampling stages, failing to exploit the inherent "coarse-to-fine" structure of image generation. The authors observe that early diffusion stages primarily establish rough conceptual layouts and object contours, which can be accomplished quickly at low resolutions using only a subset of model parameters. Computationally intensive detail enhancement occurs later. Forcing maximum resolution and model capacity throughout the process creates significant redundancy.

Goal: To decompose the generation process across temporal, spatial, and architectural dimensions simultaneously, making early layout stages faster and more efficient without compromising quality.

Key Insight: Segment the rectified flow by resolution (pyramid-style), assigning fewer layers to low-resolution segments and more layers to high-resolution ones, forming a "time-segmented + spatial-cascaded" progressive structure. The difficulty lies in smoothing the probability distribution transitions between adjacent segments to avoid quality loss from jumps.

Core Idea: Replace the non-learnable re-noising transition used in pyramid flows with a learnable BridgeFlow module, combining "Progressive Rectified Flow Transformers + Multi-resolution Joint Training" into an end-to-end efficient framework.

Method

Overall Architecture

NAMI splits a text-to-image denoising trajectory into \(K\) resolution stages (typically \(K=3\), with resolutions 256→512→1024). The rectified flow is partitioned into \(K\) corresponding time windows \([t_{k-1}, t_k]\). Each stage is assigned a module \(m_k\) composed of MM-DiT blocks, where the network for stage \(k\) is nested and cumulative: \(\theta_k = \{m_1 \oplus \cdots \oplus m_k\}\). As resolution increases, more layers participate in the computation. Inference starts from low-resolution noise; between stages, upsampling and the BridgeFlow module align the endpoint of one segment with the starting point of the next. The training phase employs a multi-resolution joint strategy to accelerate convergence.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Low-res Gaussian Noise (256)"] --> B["Progressive Rectified Flow Transformer<br/>Stage 1 · Few layers m1: Layout & Contours"]
    B --> C["BridgeFlow Transition<br/>Upsampling + Learnable Linear Transform"]
    C --> D["Progressive Rectified Flow Transformer<br/>Stage 2 · m1⊕m2 (512)"]
    D --> E["BridgeFlow Transition"]
    E --> F["Progressive Rectified Flow Transformer<br/>Stage 3 · Full Layers (1024): Detail Refinement"]
    F --> G["Output 1024×1024 Image"]

Key Designs

1. Progressive Rectified Flow Transformer: Aligning "Coarse-to-Fine" across Temporal, Spatial, and Capacity Axes

This design eliminates uniform denoising redundancy. The flow is split into \(K\) windows via a pyramid approach. Within the \(k\)-th window, the starting point \(\hat{x}_{s_k} = \text{BridgeFlow}(\text{Up}(\text{Down}(x_{t_{k-1}}, 2^{k+1})))\) is derived from the previous stage's endpoint, while the endpoint \(\hat{x}_{e_k} = \text{Down}(x_{t_k}, 2^{k})\) is obtained by downsampling data. Intrawindow interpolation follows \(\hat{x}_t = t'\hat{x}_{e_k} + (1-t')\hat{x}_{s_k}\), where \(t' = (t - t_{k-1})/(t_k - t_{k-1})\). "Spatial cascading" is reflected in the model: stage \(k\) only uses blocks \(\theta_k = m_1 \oplus \cdots \oplus m_k\). Low-resolution stages use fewer layers and tokens, resulting in high speed. The total optimization objective is summed across time windows:

\[\min_{\theta_k} \sum_{k=1}^{K} \mathbb{E}\Big[\int_{t_{k-1}}^{t_k} \big\|(\hat{x}_{s_k} - \hat{x}_{e_k}) - v_{\theta_k}(\hat{x}_t, t)\big\|^2 \, dt\Big]\]

2. BridgeFlow: Learnable Linear Transformation for Smooth Transitions

When stages are segmented, probability paths can break at jump points. Unlike Pyramid Flow, which uses non-learnable rescaling and re-noising for Gaussian distribution matching—a process that lacks robustness and scales poorly with token length—BridgeFlow uses data-driven alignment. Each stage's endpoint is upsampled and passed through a linear transformation \(\hat{x}_{s_k} = W \cdot \text{Up}(\hat{x}_{e_{k-1}}) + B\) to align it with the next stage's distribution. These modules are pre-trained independently using MSE loss.

3. Multi-resolution Joint Training: Concurrent Processing to Prevent Forgetting

Instead of sequential fine-tuning (low to high), NAMI trains on multiple resolutions simultaneously within a single batch. Data for each time window is downsampled via \(\text{Down}(\cdot)\) to calculate losses. Loss weights for different stages are dynamically adjusted during training. This encourages knowledge sharing and prevents catastrophic forgetting of semantic information learned at lower resolutions.

4. NAMI-1K Benchmark: A Diverse Human Preference Evaluation Set

To address bias in existing benchmarks like GenEval or DPG, the authors constructed NAMI-1K: 1000 prompts including 360 short prompts from GenEval/LumiereSet, 320 community-sourced prompts, and 320 long prompts (up to 120 words) generated by Cogvlm2. Evaluations cover relevance, coherence, aesthetics, and authenticity.

Loss & Training

The core training objective is the windowed rectified flow loss mentioned above. NAMI-2B consists of 22 layers (2048 width, 16 heads) with a 3-stage layer ratio of 9:7:6. NAMI-0.6B has 12 layers with a ratio of 5:4:3. The training set includes approximately 100 million images (LAION + GRIT-20M + high-quality internal data).

Key Experimental Results

Main Results

On GenEval (short prompt text-to-image alignment), NAMI-2B leads its parameter class:

Model (Params)	Overall	Single	Two	Count	Color	Pos	Color Attr
SD3-medium (2B)	0.62	0.98	0.74	0.63	0.67	0.34	0.36
Sana (1.6B)	0.66	0.99	0.77	0.62	0.88	0.21	0.47
NAMI-2B (2B)	0.65	0.99	0.78	0.64	0.82	0.20	0.45
FLUX-dev (12B)	0.67	0.99	0.81	0.79	0.74	0.20	0.47

In human preference testing (NAMI-1K), NAMI-2B achieves the highest total score among similar-sized models:

Model (Params)	Relevance	Coherence	Aesthetics	Authenticity	Total
SD3-medium (2B)	75.74	65.90	61.64	75.74	69.97
NAMI-2B (2B)	76.07	66.89	62.30	76.72	70.69

Efficiency at 1024 resolution: NAMI-2B takes 2.98s compared to 8.47s for a FLUX-based baseline (30-step uniform sampling), a 64.82% reduction.

Ablation Study

Configuration	Key Metrics	Note
Full NAMI	FID 8.93 / CLIP 25.57	Flow segments + model blocks enabled
BridgeFlow vs Pyramid Flow	0.05s / FID 8.93 vs 0.12s / FID 9.82	BridgeFlow is faster and better

Key Findings

• Source of Speedup: At 1024 resolution, flow segmentation alone saves 53.01% of inference time. Layer nesting adds another 11.81%. The primary gain comes from reducing token processing in low-resolution stages.
• Simplicity in Transitions: A single learnable linear layer in BridgeFlow provides the best trade-off. More complex structures (MLP, CNN) did not yield further quality improvements.
• Layer Allocation: Performance saturates once a sufficient number of layers is assigned to low-resolution stages. Uniform time window partitioning (1:1:1) is generally optimal.

Highlights & Insights

• Triple-Axis Decomposition: Decomposing generation across time, space, and capacity axes is highly effective. These dimensions are orthogonal and can be combined with other techniques like linear attention.
• Linear Distribution Alignment: BridgeFlow replaces complex re-noising with a simple linear transformation, demonstrating that distribution alignment at stage boundaries does not require heavy architectures.
• Joint Training for Multi-Scale: Training on multiple resolutions in the same batch is a practical trick to share knowledge across modules and mitigate forgetting during high-resolution tuning.

Limitations & Future Work

• Gap with Massive Models: NAMI-2B still trails the 12B FLUX-dev in absolute quality. Gains are most visible when comparing models of the same parameter scale.
• Manual Hyperparameters: Layer ratios and time window boundaries are manually tuned based on experience; automated search is currently lacking.
• BridgeFlow Pre-training: The requirement for separate MSE pre-training for BridgeFlow adds complexity to the training pipeline.
• Evaluation Scale: Human preference scoring on NAMI-1K remains limited by the scale and consistency of manual annotators.

Related Work & Insights

• vs Pyramid Flow: Both use temporal pyramids, but NAMI's learnable BridgeFlow is more robust than non-learnable re-noising. NAMI further introduces architectural nesting to save more computation.
• vs MatryoshkaDM: Matryoshka uses nested UNets in pixel space but does not address DiT parameter redundancy. NAMI addresses both spatial and capacity redundancy in latent space.
• vs SANA: While SANA uses high-ratio VAE downsampling, NAMI decomposes the generation process, a strategy that is orthogonal and potentially complementary.

Rating

• Novelty: ⭐⭐⭐⭐
• Experimental Thoroughness: ⭐⭐⭐⭐
• Writing Quality: ⭐⭐⭐⭐
• Value: ⭐⭐⭐⭐

Related Papers

• [CVPR 2026] Probabilistic Precipitation Nowcasting with Rectified Flow Transformers
• [CVPR 2026] DDiT: Dynamic Patch Scheduling for Efficient Diffusion Transformers
• [CVPR 2026] From Sketch to Fresco: Efficient Diffusion Transformer with Progressive Resolution
• [CVPR 2026] RecTok: Reconstruction Distillation along Rectified Flow
• [NeurIPS 2025] Efficient Rectified Flow for Image Fusion