# NFIG: Multi-Scale Autoregressive Image Generation via Frequency Ordering Zhihao Huang1,2 Xi Qiu2 Yukuo Ma2,4 Yifu Zhou1,2 Junjie Chen2 Hongyuan Zhang2,3,∗ Chi Zhang2,∗ Xuelong Li2,∗ 1 Northwest Polytechnical University 2 TeleAI, China Telecom 3 University of Hong Kong 4 Beihang University huangzhihao@mail.nwpu.edu.cn, hyzhang98@gmail.com, zhangc120@chinatelecom.cn, xuelong_li@ieee.org # Abstract Autoregressive models have achieved significant success in image generation. However, unlike the inherent hierarchical structure of image information in the spectral domain, standard autoregressive methods typically generate pixels sequentially in a fixed spatial order. To better leverage this spectral hierarchy, we introduce Next-Frequency Image Generation (NFIG). NFIG is a novel framework that decomposes the image generation process into multiple frequency-guided stages. NFIG aligns the generation process with the natural image structure. It does this by first generating low-frequency components, which efficiently capture global structure with significantly fewer tokens, and then progressively adding higher-frequency details. This frequency-aware paradigm offers substantial advantages: it not only improves the quality of generated images but crucially reduces inference cost by efficiently establishing global structure early on. Extensive experiments on the ImageNet-256 benchmark validate NFIG’s effectiveness, demonstrating superior performance (FID: 2.81) and a notable $1 . 2 5 \times$ speedup compared to the strong baseline VAR-d20. The source code is available at https://github.com/Pride-Huang/NFIG. # 1 Introduction 2 The synthesis of images has emerged as a fundamental challenge in computer vision [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]. Rapid progress in this field has been propelled by deep generative models, such as autoregressive models (AR) [13], Generative Adversarial Networks (GANs), and diffusion models (SD) [14]. Despite remarkable advances in existing methods, AR models for image generation still face several fundamental challenges. On the one hand, due to their inherently local and sequential nature, most current AR models struggle to effectively capture long-range dependencies and global structure [15]. For example, PixelCNN [9] generates an image by predicting each pixel in a raster scanning sequence, which neglects the global image structure and relationships with distant elements. On the other hand, the generation process is computationally intensive and time-consuming, as AR models always generate pixels or patches sequentially in a predetermined order, with each new element requiring the computation of conditional probabilities based on all previously generated content [16, 17]. For instance, ViTVQ [18] requires more than 6 seconds to generate a $2 5 6 \times 2 5 6$ image over ![](images/figures/nfig-fig-0001.jpg) Figure 1: Illustration of three autoregressive image generation frameworks. The figure demonstrates three prediction approaches: Next-Patch Prediction (patch-based progression), Next-Scale Prediction (coarse-to-fine resolution generation), and Next-Frequency Prediction (NFIG), which performs image generation by progressively predicting and synthesizing frequency components from low to high, resulting in a coarse-to-fine spatial reconstruction. 1024 steps, making it impractical for real-time applications. Most importantly, AR models face a fundamental challenge in defining a meaningful autoregressive sequence. Traditional AR models using raster scanning or predefined arbitrary orders fail to reflect the natural hierarchical structure and dependencies in images[19]. This improper sequence design makes it difficult for models to capture the true causal relationships between different image components, ultimately affecting the coherence and visual quality of generated outputs. To address these limitations, recent works have explored incorporating various improvements into the generation process. For example, Taming Transformer [20] partially addresses long-range dependency challenges through its discrete latent space and transformer architecture, but still suffers from computational inefficiency. Fast PixelCNN $^ { + + }$ [21] speeds up generation in convolutional autoregressive models by caching hidden states to avoid redundant computation, achieving up to $1 8 3 \times$ speedups, yet it doesn’t fundamentally change the autoregressive sequence design. VAR [19] leverages the Laplacian Pyramid as a prior to guide autoregressive image generation across different resolutions, achieving improved generation quality with reduced computational overhead. However, these methods do not fully exploit the potential of natural priors inherent in raw images to guide the generation process and improve the efficiency of AR models. In fact, the natural structure of images follows a hierarchical frequency distribution—low frequencies encode global structures while high frequencies contain local details. This organization suggests an efficient autoregressive generation sequence from low to high frequencies, aligning with visual information’s natural structure. Since low-frequency components require fewer tokens to represent, this approach enhances computational efficiency. Similar frequency-progressive principles have proven effective in diffusion models, which build from low-frequency foundations before adding higher-frequency details [22]. Motivated by this insight, we propose a Next-Frequency Image Generation (NFIG) framework for AR models that: (1) first generates a low-frequency image with few tokens to capture global structure; (2) then progressively adds higher-frequency components conditioned on the low-frequency foundation. This process has been shown in Figure 1. Grounded in information theory, this approach efficiently represents information across the frequency spectrum using our Frequency-guided Residual-quantized VAE (FR-VAE). Key contributions of the NFIG framework include: • We introduce a Next-Frequency Image Generation (NFIG) framework that incorporates frequency analysis into AR image generation. To our knowledge, this work is the first to guide autoregressive generation using the image’s frequency spectrum, associating low frequencies with lower resolutions and high frequencies with higher resolutions; • To demonstrate the feasibility of the NFIG paradigm, we design a Frequency-guided Residualquantized VAE as our image tokenizer. FR-VAE separates low and high-frequency components in the representation learning process, with low frequencies encoding global structure and high frequencies preserving local details. Experiments show FR-VAE achieves a reconstruction FID of 0.85, validating its image content preservation capability; • Through extensive experimentation, we show that our approach achieves state-of-the-art image generation quality, evidenced by an FID of 2.81 from a relatively small model. This improvement paves the way for more effective and efficient AR image generation models, making them more practical for real-world applications. Table 1: Main Notation Table
SymbolMeaningDimension
xInput imageH × W × 3
xReconstructed imageH × W × 3
fFeature map from VAE encoderH ′ × W ′ × C
MiThe i-th frequency selection maskH′ × W ′ × C
$f$Set of i-th frequency component feature mapsH′ × W ′ × C
viScaled feature map for i-th frequency componenthi × wi × C
Quantized representation for i-th frequency componenthi × wi × C
Cumulative signal residual through level iH′ × W ′ × C
ZThe learnable codebook of FR-VAEK×C
FiThe i-th frequency bandN/A
# 2 Related Work Autoregressive Image Generation Autoregressive image generation has demonstrated remarkable capabilities in producing high-quality images by modeling the joint distribution of image tokens as a product of conditionals[23]. PixelCNN [9] generates images sequentially, processing pixels one by one (typically top-left to bottom-right). It employs masked convolutions so that the generation of each pixel depends solely on the pixels already generated. Taming Transformer [20] introduces an autoregressive approach that generates high-resolution images by predicting the next latent patch token in a discrete compressed space learned through vector quantization. Emu3 [24] patchifies an image into a series of tokens and generates images by predicting tokens in a raster-scan manner. VAR [19] incorporates the prior knowledge of Laplacian Pyramid into transformer architecture and generates images in a next-resolution manner. MAR [25] improves the quality of generated images by replacing discrete tokens with continuous features, recognizing that autoregressive models primarily need per-token probability distributions. FAR [26] attempts to enhance MAR performance through frequency-based approaches, yet lacks critical insights into the distinctive information characteristics across different frequency bands. Infinity [25] combines autoregressive modeling with Bit-wise Modeling to enhance visual details in high-resolution image synthesis. ImageFolder [27] utilizes folded image tokens to generate high-quality images in a next-scale prediction manner, achieving superior performance. These approaches collectively demonstrate the evolution of autoregressive image generation techniques, progressing from pixel-level prediction to more sophisticated methods involving latent spaces, hierarchical structures, and physical priors. Despite their differences in implementation, all these methods share the fundamental autoregressive principle of sequentially generating image elements conditioned on previously generated content. Image Tokenizer Image tokenizers, which transform continuous image data into discrete representations, have become a critical component in modern image generation systems. VQ-VAE [28] introduces vector quantization into VAE, reducing the pressure on the downstream generative model by transforming the continuous latent space into a discrete one. VQ-VAE-2 [29] extends this idea with a hierarchical framework and multi-scale codebooks, where top-level codes capture global structure, and bottom-level codes model local details, enabling higher-quality reconstruction and generation at increased resolutions. However, VQ-VAE-based methods often suffer from codebook collapse, where only a few codebook entries are effectively used. FSQ [30] attempts to address this issue by utilizing finite scalar quantization to learn the codebook, but it does not learn a meaningful feature of images. RQ-VAE [31] employs residual quantization with a shared codebook to enhance reconstructed image quality. XQGAN [32] introduces feature product decomposition and a residual quantizer to enhance VQ-VAE’s performance, leading to improved image generation results. While these approaches have made significant progress in image tokenization, they often neglect the inherent multi-scale structure of natural images, which is crucial for efficient and effective representation learning. ![](images/figures/nfig-fig-0002.jpg) Figure 2: Overview of the Next-Frequency Image Generation (NFIG) Framework: (a) The Frequencyguided Residual-Quantization VAE encodes images into and decodes from frequency-guided residual quantized representations; (b) The image is decomposed into frequency components (low to high) and reconstructed progressively by merging these components for a coarse-to-fine process; (c) Next-Frequency Prediction model employs a frequency-aware Transformer to auto-regressively generate token sequences, with each block with same color representing a specific frequency band, enabling sequential image synthesis from low to high frequencies. $\tilde { N _ { i } ^ { t o k } } \doteq h _ { i } w _ { i }$ is the number of image tokens used for the $i _ { t h }$ frequency band. # 3 Next-Frequency Image Generation To provide a comprehensive understanding of our NFIG methodology, this section delves into its intricate architectural structure. The essential operational sequence, illustrating the flow and interaction of the system’s key components, is clearly visualized in Figure 2. The details of loss function have been listed in Appendix B.1. # 3.1 Frequency-guided Residual-quantized VAE The workflow of Frequency-guided Residual-quantized VAE has been shown in Figure 2 (a). To generate images in a frequency-aware manner, we propose a Frequency-guided Residual-quantized VAE (FR-VAE) with VQ-GAN framework. The key idea is to represent lower-frequency signals with fewer tokens and higher-frequency components with more tokens. # 3.1.1 Frequency-guided Reconstruction As illustrated in Figure 2 (b), raw images can be decomposed into components across different frequency bands: low frequencies encode the global structure, while high frequencies retain fine details. Utilizing the Frequency-guided Decomposer and Composer, these components can be recombined without loss, ensuring a complete and accurate visual representation. Frequency-guided Decomposer. Given a image $x \in R ^ { H \times W \times 3 }$ and a encoder $E ( \cdot )$ , there is image latent feature $f = E ( x )$ and $f \in \mathbb { R } ^ { H ^ { \prime } \times W ^ { \prime } \times C }$ . FR-VAE decomposes $f$ into several component with different frequency by Frequency-guided Decomposer via Fast Fourier Transform (FFT): $$ \hat { f } _ { i } = \mathcal { F } ^ { - 1 } ( \mathcal { F } ( f ) \odot M _ { i } ) , \forall i \in \{ 1 , \cdots , n \} . $$ Here, $\odot$ represents the element-wise product, $\mathcal { F }$ signifies the FFT operation, ${ \mathcal { F } } ^ { - 1 }$ indicates the inverse FFT, and $M _ { i }$ is the $i$ -th frequency mask used to select the desired frequency range, $n$ is the total number of frequency masks, $\hat { f } _ { i }$ is the component corresponding to $M _ { i }$ . Frequency-guided Composer. Frequency-guided Composer reconstructs the raw image by interpolating different frequency components to a uniform size and merging them into a single image, as illustrated in Figure 2 (b). $$ \tilde { f } = \sum _ { i = 1 } ^ { n } \mathcal { T } ( \hat { f } _ { i } , H ^ { \prime } , W ^ { \prime } ) , $$ where $\mathcal { T } ( \cdot , H ^ { \prime } , W ^ { \prime } )$ is the interpolation function, which enables the Frequency-guided Composer to process images of different frequencies at varying resolutions. # 3.1.2 Frequency-guided Residual-quantization To efficiently represent images with minimal tokens, we implement a frequency-guided residual quantization approach that addresses information loss during downsampling. Our method progressively captures different frequency components of an image through a residual learning scheme. Residual Token Extraction. Given a sequence of feature maps with different dimensions $\{ ( h _ { 1 } , w _ { 1 } ) , \cdot \cdot \cdot , ( h _ { n } , w _ { n } ) \}$ , where $h _ { i } \geq h _ { j }$ and $w _ { i } \geq w _ { j }$ if $i \geq j$ , and $h _ { n } = H ^ { \prime }$ and $w _ { n } = W ^ { \prime }$ , we supervise the learning process using accumulated signals from the lowest frequency to the current frequency band. The residual $R _ { i } \in \mathbb { R } ^ { H ^ { \prime } \times W ^ { \prime } \times C }$ and representation $v _ { i } \in \mathbb { R } ^ { h _ { i } \times w _ { i } \times C }$ of the $i$ -th frequency component can be computed as follows: $$ \begin{array}{c} \begin{array} { r l r } & { } & { R _ { i } = \left\{ \hat { f } _ { i } - \mathcal { Z } ( v _ { i } , H ^ { \prime } , W ^ { \prime } ) ) , \right. \ } \\ & { } & { R _ { i } = \left\{ \begin{array} { l l } { \displaystyle { \hat { R } _ { i } - 1 + ( \hat { f } _ { i } - \mathcal { Z } ( v _ { i } , H ^ { \prime } , W ^ { \prime } ) ) } ) , } & { i \ge 1 } \end{array} \right. , } \\ & { } & { v _ { i } = \left\{ \arg \operatorname* { m i n } _ { v _ { i } } \| \hat { f } _ { i } - \mathcal { Z } ( v _ { i } , H ^ { \prime } , W ^ { \prime } ) \| ^ { 2 } , \right. \ ~ \left. ~ i = 0 \right. } \\ & { } & { \left. \operatorname { a r g m i n } _ { v _ { i } } \| ( R _ { i - 1 } + \hat { f } _ { i } ) - \mathcal { Z } ( v _ { i } , H ^ { \prime } , W ^ { \prime } ) \| ^ { 2 } , \quad i \ge 1 \right.} \end{array} , \end{array} $$ where $\mathcal { T } ( v _ { i } , H ^ { \prime } , W ^ { \prime } )$ is the interpolation function that upsamples $v _ { i }$ to the original feature map size, and $R _ { i }$ represents the difference between the accumulated frequency components up to the $i$ -th level and the learnable features. Vector Quantization. In general, autoregressive models utilize the discrete tokens to generate a image. To achieve this goal, we take a simple vector quantization to transform the continuous token into discrete tokens. We define a quantizer $Q$ with a learnable codebook $Z \in \mathbb { R } ^ { K \times C }$ containing $K$ code vectors. Using this codebook, the quantizer $Q$ transforms a continuous feature map $v _ { i } \in \mathbb { R } ^ { h _ { i } \times w _ { i } \times C }$ into a set of ns {t(1,1)i , t(1,2)i , · $\{ t _ { i } ^ { ( 1 , 1 ) } , t _ { i } ^ { ( 1 , 2 ) } , \cdot \cdot \cdot , t _ { i } ^ { ( h _ { i } , w _ { i } ) } \}$ , where each token timal code represen $t _ { i } ^ { ( j , k ) }$ has the corresponds to a vector involves: $z _ { i } ^ { ( j , k ) } \in \mathbb { R } ^ { C }$ $$ t ^ { ( j , k ) } = \mathrm { l o o k u p } ( Z , \underset { z _ { i } ^ { ( j , k ) } \in Z } { \arg \operatorname* { m i n } } \| z _ { i } ^ { ( j , k ) } - v _ { i } ^ { ( j , k ) } \| _ { 2 } ) . $$ From $v _ { i }$ , the quantized feature map $v _ { i } ^ { q } \in \mathbb { R } ^ { h _ { i } \times w _ { i } \times C }$ and a set of discrete token s {t(j,k)i } a re obtained through quantization using the codebook $Z$ . Here, lookup $( Z , x )$ is a function that finds the index of the closest entry to $x$ in codebook $Z$ . # 3.2 Autoregressive Image Generation To generate images progressively from low to high frequency components, we implement a decoderonly transformer framework and block-wise causal attention [19]. Next-Frequency Image Prediction Unlike conventional autoregressive image generation models that employ a "token-by-token prediction" strategy, which often neglects spatial relationships and inherent image structure. NFIG adopts a “Coarse-to-Fine Generation" approach, first synthesizes the low-frequency components of an image, then iteratively incorporates higher-frequency details, progressively refining the generated output at each step, as shown in Figure 2 (c). The generation process for next-frequency prediction is given by the autoregressive factorization: $$ p ( T _ { 1 } , T _ { 2 } , \cdot \cdot \cdot , T _ { n } ) = \prod _ { i = 1 } ^ { n } p ( T _ { i } | T _ { 1 } , T _ { 2 } , \cdot \cdot \cdot , T _ { i - 1 } ) $$ where $T _ { i } \in [ K ] ^ { h _ { i } \times w _ { i } }$ is the matrix of code indices for the $i$ -th frequency component, and the set $[ K ] = \{ 1 , 2 , \dots , K \}$ represents all available index values. Frequency Band Division Strategy We treat the lower frequency components as the foundation for generating $T _ { i }$ , represented by $\{ \bar { T } _ { 1 } , T _ { 2 } , \cdot \cdot \cdot , T _ { i - 1 } \}$ . According to information theory principles, lower-frequency signals contain less information and require fewer tokens, while higher-frequency components carry more detailed information and need more tokens for accurate representation. Consequently, we establish an increasing scale sequence $\{ ( h _ { 1 } , w _ { 1 } ) , ( h _ { 2 } , w _ { 2 } ) , \colon \colon \colon , ( h _ { n } , w _ { n } ) \}$ for components with increasing frequency bands $\{ F _ { 1 } , F _ { 2 } , \cdot \cdot \cdot , F _ { n } \} = \{ [ 0 , \sigma _ { 1 } ) , [ \sigma _ { 1 } , \sigma _ { 2 } ) , \cdot \cdot \cdot , [ \sigma _ { n - }$ 1, σn]}. Here, $\sigma _ { m a x }$ denotes the maximum frequency of the entire image feature map $f$ , with $\sigma _ { n } = \sigma _ { m a x }$ . We divide the frequency bands based on their corresponding resolution as: $$ \sigma _ { i } = \sigma _ { i - 1 } + \frac { h _ { i } \cdot w _ { i } } { \sum _ { j = 1 } ^ { n } h _ { j } \cdot w _ { j } } \times \sigma _ { m a x } . $$ This frequency-guided progressive approach allows our model to capture and prioritize salient components at each stage. The method improves both computational efficiency and image quality by explicitly modeling the multi-scale frequency structure inherent in natural images. # 4 Experiment This section details our experimental methodology, covering datasets, evaluation metrics, comparison baselines, and implementation specifics. We then evaluate NFIG against state-of-the-art approaches on image generation benchmarks. Subsequently, ablation studies and motivation verification experiments are conducted to analyze the impact of different components and validate design decisions. # 4.1 Experimental Settings Dataset. For the purpose of our experiments, we use the ILSVRC 2012 subset of ImageNet [33], which comprises a total of 1.2 million training images, 50k validation images, and 100k test images. This subset focuses on 1k object categories, with each category having approximately $1 . 2 \mathrm { k }$ training images, 50 validation images, and 100 test images. Evaluation Metrics. We adopt four metrics for quantitative evaluation: Fréchet Inception Distance (FID) which measures distribution similarity between generated and real images, Inception Score (IS) which assesses quality and diversity, and Precision (Pre) and Recall (Rec) which evaluate sample fidelity and diversity coverage respectively. Baselines. Our method is benchmarked against several leading image generation techniques, including generative adversarial networks (GAN), diffusion models (Diff.), mask diffusion (Mask.), and autoregressive models (AR). These approaches have demonstrated strong performance on various image synthesis tasks and serve as robust comparators. Implementation Details. Our model was implemented using the PyTorch framework [44] and trained on NVIDIA H100 graphics cards. To ensure the reproducibility of our experiments, our implementation is built upon open-source research code, while incorporating improvements specific to this study. For the image tokenizer, the FR-VAE incorporates a VQGAN architecture with a DINO discriminator. The image encoder is initialized with pretrained weights from DINOv2-base. Since VAR’s image tokenizer training code is not open-source, we adopted XQGAN’s implementation strategy. The frequency residual quantizer employs multiple scaling factors $[ 1 , 2 , 3 , 4 , 5 , 6 , 8 , 1 0 , 1 3 , 1 6 ]$ across different frequency bands, resulting in a vocabulary size of 680 tokens. The FR-VAE codebook size of 4096 was utilized. The image generator employs a VAR Transformer backbone with a depth of 16, enabling multi-scale image prediction. Optimization was performed using the Adam optimizer, setting the learning rate to $8 \times 1 0 ^ { - 5 }$ and the batch size to 768. Training of the model ran for 350 epochs on the ImageNet dataset. For inference, we configured CFG to 4.5 and top_k to 990. Table 2: Performance on class-conditional ImageNet $2 5 6 \times 2 5 6$ for image generative model. rFID represents reconstruction FID , while gFID indicates generation FID. “↓” or “↑” indicate lower or higher values are better. “#Step”: the number of model runs needed to generate an image. Wall-clock inference time relative to NFIG is reported. Models with the suffix “-re” used rejection sampling. $\dagger$ : taken from MaskGIT [34]. For comprehensive evaluation, we separately compare autoregressive (AR) and non-autoregressive (non-AR) models, with the best metrics highlighted in bold.
TypeModelrFID↓gFID↓IS↑Pre↑Rec↑#Para#StepTime
GANBigGAN [35]-6.95224.50.890.38112M1
GANGigaGAN [36]-3.45225.50.840.61569M1—
GANStyleGan-XL [3]2.30265.10.780.53166M10.75
Diff.ADM [37]-10.94101.00.690.63554M250420
Diff.CDM [38]4.88158.78100
Diff.LDM-4-G [39]-3.60247.7400M25025077.5
Diff.DiT-L/2 [40]0.95.02167.20.750.57458M
Diff.DiT-XL/2 [40]0.92.27278.20.830.57675M250112.5
Diff.L-DiT-3B [41]0.92.10304.40.820.603.0B250>112.5
Diff.L-DiT-7B [41]0.92.28316.20.830.587.0B250>112.5
Mask.MaskGIT [34]2.286.18182.10.800.51227M81.254.75
Mask.RCG [42]-3.49215.5−502M20
ARVQVAE-2† [34]2.031.1145.00.360.5713.5B5120
ARVQGAN† [20]7.9418.6580.40.780.26227M25647.5
ARVQGAN [20]7.9415.7874.31.4B25660
ARViTVQ [18]1.284.17175.11.7B1024>60
ARViTVQ-re [18]1.283.04227.4−1.7B1024>60
ARRQTran. [43]1.837.55134.03.8B6852.5
ARRQTran.-re [43]1.833.80323.7—3.8B6852.5--
ARFAR-B [26]-4.26248.90.790.51208M10
ARFAR-B [26],3.45282.20.800.54427M10
ARFAR-H [26]-3.21300.60.810.55812M10-
ARXQGAN-310M [32]0.782.96---310M101
ARVAR-d16 [19]0.93.55274.40.840.51310M101
ARARVAR-d20 [19]NFIG(Ours)0.90.852.95302.60.830.56600M10101.251
2.81332.420.770.59310M
# 4.2 Main Results Table 2 provides a detailed comparison of our approach against leading image generative models evaluated on ImageNet $2 5 6 \times 2 5 6$ . The findings indicate that NFIG achieves superior performance within the AR model family while establishing itself as a formidable competitor among diverse generative methods across different paradigms. AR Model Comparison. NFIG achieves the best gFID (2.81) and IS (332.42) scores, significantly outperforming other AR models. Compared to VAR-16 (gFID: 3.55, IS: 274.4), our approach reduces FID by 0.74 and improves IS by more than $21 \%$ . XQGAN-310M has a better image tokenizer with rFID 0.78 with gFID 2.96. This indicates that NFIG’s performance improvement is not solely due to a good image tokenizer, but more importantly, to the injection of image frequency priors. Additionally, the proposed approach outperforms VAR-d20, a relatively larger model, while delivering $2 5 \%$ faster inference speed. Cross-family Comparison. NFIG achieves competitive performance with the best models from other families. NFIG outperforms the best mask diffusion model RCG, which has a gFID score of 3.49 and IS score of 215.5. While some GANs like StyleGAN-XL have lower gFID scores (2.30) with moderate IS (265.1), and diffusion models like DiT-L/2 show excellent gFID (2.27) and strong IS (278.2), NFIG uniquely balances both metrics at high levels (gFID: 2.81, IS: 332.42). This establishes NFIG as not only the leading AR model but also a strong competitor across all model types. Qualitative Results. Figure 3 qualitatively shows NFIG’s impressive ability to generate diverse ImageNet $2 5 6 \times 2 5 6$ images across a wide variety of categories. Appendix B.5 show the Failure case of NFIG. Scaling Up. To validate NFIG’s scaling behavior, we train 310M and 600M parameter models for 55 epochs under computational constraints. The results are shown in Table 3. Performance of Different Epochs. VAR sets different training epochs for models of different sizes: 200 epochs for 310M, 250 epochs for 600M, 300 epochs for 1B, and 350 epochs for 2B. Limited by computational resources, we focus on training our 310M model. As Table 4 shows, NFIG outperforms VAR at matched epochs and demonstrates superior parameter efficiency. At 200 epochs, NFIG already outperforms VAR-d16 of the same size. With extended training, NFIG-310M achieves performance comparable to or better than VAR-d20-600M (twice the parameters) while using significantly fewer resources. Table 3: Performance of NFIG with different parameters at 55 epochs.
ModelFID↓IS↑Precision↑Recall↑EpochParastepsTime
NFIG-310M5.47224.200.75690.491455310M101
NFIG-600M5.07225.160.71840.554655600M101.25
![](images/figures/nfig-fig-0003.jpg) Figure 3: Generated $2 5 6 \times 2 5 6$ examples by NFIG trained on Imagenet. Table 4: The performance of NFIG and VAR at different epochs.
ModelEpochsFID↓IS↑Pre↑Rec↑ParamsStepsTime
VAR-d162003.55274.40.840.51310M101
VAR-d202502.95302.60.830.56600M101.25
NFIG(ours)2003.35309.20.790.55310M101
NFIG(ours)2503.16311.90.780.56310M101
NFIG(ours)3002.93325.90.790.56310M101
NFIG(ours)3502.81332.40.770.59310M101
# 4.3 Ablation Study To evaluate the contribution of various components within our proposed NFIG model, we perform a comprehensive ablation analysis on the ImageNet validation set. Table 5 summarizes the results of this study, with performance evaluated using rFID and gFID. We start with the baseline AR model with a sequence length of 256, which achieves an rFID of 1.62 and a gFID of 18.65. Image Tokenizer. We incrementally add components to progressively improve the model’s performance. First, incorporating frequency-guided residual quantization into the VAR framework reduces the rFID to 1.40. Next, we integrate the DINO discriminator from VAR’s tokenizer, which substantially improves the rFID from 1.40 to 0.85. Transformer. We then utilize the image tokenizer (FR-VAE) to train the transformer model. Without Top_k and Classifier Free Guidance (CFG), NFIG achieves a gFID of 9.7. The addition of Top_k sampling strategy further reduces the gFID to 6.83. Finally, incorporating CFG yields the best overall performance, maintaining the rFID at 0.85 while dramatically improving the gFID to 2.81. Our experimental results demonstrate that the combination of FR-VAE with CFG provides optimal generation quality. Moreover, we observe that both DINO discriminator and FR-VAE contributes significantly to improving rFID for the image tokenizer. Additionally, Top_k sampling and CFG prove essential for reducing gFID. These results underscore the significance of discriminator guidance and conditional generation strategies for improving image generation quality. # 4.4 Motivation Verification Frequency Distribution Analysis. The experimental results in Figure 4 demonstrate the progressive refinement of generated images and the effective capture and synthesis of multi-scale visual features by FR-VAE. The frequency spectrum visualizations reveal the model’s ability to hierarchically incorporate information from low to high frequencies, resulting in generated images with rich details and natural appearance. Appendix B.2 compares the frequency keep ability of NFIG and VAR. These results demonstrate that NFIG’s frequency-guided approach enables more effective feature learning, particularly at lower resolutions, by maintaining balanced loss values throughout the hierarchical generation process. ![](images/figures/nfig-fig-0004.jpg) Figure 4: Generated images at different steps 2, 4, 6, 8, 10 of a 10-step process by FR-VAE, with corresponding frequency spectrum. In these spectrograms, brightness (red/yellow) indicates higher frequency energy while darker colors (blue) represent lower energy components. The center of each plot shows low-frequency information, with frequencies increasing radially outward, revealing the evolving distribution during the generation process. Frequency Guidance. Similar to NFIG, VAR follows a "coarse-to-fine" approach but differs significantly in loss computation across resolutions. VAR computes loss between different resolutions and the raw image, causing disproportionately large loss values at lower resolutions. In contrast, NFIG utilizes frequency components to guide feature learning, providing more balanced loss values throughout generation. As Figure 5 shows, this leads to dramatic variations in vector quantization loss—VAR exhibits substantially higher values across all scale factors, while NFIG maintains considerably lower loss values across all resolutions. ![](images/figures/nfig-fig-0005.jpg) Figure 5: Vector quantization loss comparison between NFIG and VAR across image scales. Table 5: Ablation study on the improvement of NFIG. We evaluate rFID and gFID on the ImageNet validation set. “FR-Quantizer” is the quantizer of FR-VAE. “DINO-Disc” means “DINO Discriminator”, which denotes the discriminator used in VAR’s ([19]) tokenizer.
TypeMethodLengthMetric
rFID↓gFID↓
1AR2561.6218.65
Image Tokenizer
2+ FR-Quantizer6801.40
3+ DINO-Disc6800.85
Generation Transformer
4+ AdaLN6800.859.7
5+ Top_k6800.856.83
6+ CFG6800.852.81
# 5 Conclusion This paper introduces Next-Frequency Image Generation, a novel autoregressive framework that decomposes image generation into frequency-guided stages. Our key insight leverages the hierarchical spectral distribution of natural images: low-frequency components encode global structures and long-range dependencies, while high-frequency components contain local details requiring greater information entropy. By progressively generating from low to high frequencies, the proposed method significantly outperforms existing models with comparable parameter counts, demonstrating superior quality metrics while maintaining computational efficiency. Experiments confirm that our frequencyguided approach represents an important advancement in autoregressive image synthesis. # 6 Limitation and Future Work Our frequency-guided autoregressive image generation approach shows promise, but has limitations. Improving frequency decomposition. A primary issue is the simplistic frequency band division by scale, which inadequately captures information in the first band. Implementing a more rigorous division based on statistical analysis and physical principles would likely enhance NFIG’s performance. Recent advances in beneficial noise theory [45, 46] suggest that properly designed noise can reduce task complexity, which could inform better frequency decomposition and data augmentation strategies [47, 48]. Extension to other modalities. Beyond 2D spatial frequency, future work could extend to video generation by incorporating temporal frequency decomposition, or to 3D object generation where frequency analysis is vital for accurate light field representation. For multi-modal generation, techniques that enhance cross-modal alignment through learnable noise [49] may offer insights for frequency-based fusion strategies. Privacy-preserving generation. Adversarial noise techniques [50] could be integrated with our framework to ensure privacy protection in generated content. Due to computational constraints and time limitations, these promising directions remain unexplored and are left for future investigation. # References [1] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. [2] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4401–4410. Computer Vision Foundation / IEEE, 2019. [3] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1–10, 2022. [4] Jihyun Kim, Changjae Oh, Hoseok Do, Soohyun Kim, and Kwanghoon Sohn. Diffusiondriven gan inversion for multi-modal face image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10403–10412, 2024. [5] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. [6] Yufan Zhou, Bingchen Liu, Yizhe Zhu, Xiao Yang, Changyou Chen, and Jinhui Xu. Shifted diffusion for text-to-image generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 10157– 10166. IEEE, 2023. [7] Alexandros Graikos, Srikar Yellapragada, Minh-Quan Le, Saarthak Kapse, Prateek Prasanna, Joel H. Saltz, and Dimitris Samaras. Learned representation-guided diffusion models for largeimage generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 8532–8542. IEEE, 2024. [8] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. [9] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelcnn decoders. Advances in neural information processing systems, 29, 2016. [10] Mengqi Huang, Zhendong Mao, Zhuowei Chen, and Yongdong Zhang. Towards accurate image coding: Improved autoregressive image generation with dynamic vector quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 22596–22605. IEEE, 2023. [11] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 26429–26445. IEEE, 2024. [12] Haohan Weng, Danqing Huang, Yu Qiao, Zheng Hu, Chin-Yew Lin, Tong Zhang, and CL Chen. Desigen: A pipeline for controllable design template generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12721–12732, 2024. [13] Runqi Wang, Caoyuan Ma, Guopeng Li, Hanrui Xu, Yuke Li, and Zheng Wang. You think, you act: The new task of arbitrary text to motion generation, 2025. URL https://arxiv.org/ abs/2404.14745. [14] Runqi Wang, Caoyuan Ma, Jian Zhao, Hanrui Xu, Dongfang Sun, Haoyang Chen, Lin Xiong, Zheng Wang, and Xuelong Li. Leader and follower: Interactive motion generation under trajectory constraints, 2025. URL https://arxiv.org/abs/2502.11563. [15] Rewon Child. Very deep vaes generalize autoregressive models and can outperform them on images. In International Conference on Learning Representations, 2021. [16] Chenlin Meng, Lantao Yu, Yang Song, Jiaming Song, and Stefano Ermon. Autoregressive score matching. Advances in Neural Information Processing Systems, 33:6673–6683, 2020. [17] Jiawei Shao and Xuelong Li. Ai flow at the network edge. IEEE Network, 2025. [18] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022. [19] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. In Advances in Neural Information Processing Systems 38, 2024. URL http://papers.nips.cc/paper_files/paper/2024/ hash/9a24e284b187f662681440ba15c416fb-Abstract-Conference.html. [20] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873–12883, 2021. [21] Prajit Ramachandran, Tom Le Paine, Pooya Khorrami, Mohammad Babaeizadeh, Shiyu Chang, Yang Zhang, Mark Hasegawa-Johnson, Roy Campbell, and Thomas Huang. Fast generation for convolutional autoregressive models. In 5th International Conference on Learning Representations, ICLR 2017, 2017. [22] Yurui Qian, Qi Cai, Yingwei Pan, Yehao Li, Ting Yao, Qibin Sun, and Tao Mei. Boosting diffusion models with moving average sampling in frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8911–8920, 2024. [23] Xiang Li, Kai Qiu, Hao Chen, Jason Kuen, Zhe Lin, Rita Singh, and Bhiksha Raj. Controlvar: Exploring controllable visual autoregressive modeling. arXiv preprint arXiv:2406.09750, 2024. [24] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, Yingli Zhao, Yulong Ao, Xuebin Min, Tao Li, Boya Wu, Bo Zhao, Bowen Zhang, Liangdong Wang, Guang Liu, Zheqi He, Xi Yang, Jingjing Liu, Yonghua Lin, Tiejun Huang, and Zhongyuan Wang. Emu3: Next-token prediction is all you need. CoRR, abs/2409.18869, 2024. [25] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. arXiv preprint arXiv:2406.11838, 2024. [26] Hu Yu, Hao Luo, Hangjie Yuan, Yu Rong, and Feng Zhao. Frequency autoregressive image generation with continuous tokens. arxiv: 2503.05305, 2025. [27] Xiang Li, Kai Qiu, Hao Chen, Jason Kuen, Jiuxiang Gu, Bhiksha Raj, and Zhe Lin. Imagefolder: Autoregressive image generation with folded tokens. CoRR, abs/2410.01756, 2024. [28] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. [29] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. [30] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: VQ-VAE made simple. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024, 2024. [31] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523–11532, 2022. [32] Xiang Li, Kai Qiu, Hao Chen, Jason Kuen, Jiuxiang Gu, Jindong Wang, Zhe Lin, and Bhiksha Raj. Xq-gan: An open-source image tokenization framework for autoregressive generation, 2024. URL https://arxiv.org/abs/2412.01762. [33] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211–252, 2015. [34] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. [35] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=B1xsqj09Fm. [36] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10124–10134, 2023. [37] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780–8794, 2021. [38] Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1–33, 2022. [39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. Highresolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. [40] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023. [41] Alpha-VLLM. Large-dit-imagenet, 2024. URL https://github.com/Alpha-VLLM/ LLaMA2-Accessory/tree/f7fe19834b23e38f333403b91bb0330afe19f79e/ Large-DiT-ImageNet. Accessed: 2, 7, 8. [42] Tianhong Li, Dina Katabi, and Kaiming He. Self-conditioned image generation via generating representations. CoRR, 2023. [43] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523–11532, 2022. [44] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. [45] Xuelong Li. Positive-incentive noise. IEEE Transactions on Neural Networks and Learning Systems, 35(6):8708–8714, 2024. [46] Hongyuan Zhang, Sida Huang, Yubin Guo, and Xuelong Li. Variational positive-incentive noise: How noise benefits models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(9):8313–8320, 2025. [47] Hongyuan Zhang, Yanchen Xu, Sida Huang, and Xuelong Li. Data augmentation of contrastive learning is estimating positive-incentive noise. arXiv preprint arXiv:2408.09929, 2024. [48] Siqi Huang, Yanchen Xu, Hongyuan Zhang, and Xuelong Li. Learn beneficial noise as graph augmentation. In Forty-second International Conference on Machine Learning, 2025. [49] Sida Huang, Hongyuan Zhang, and Xuelong Li. Enhance vision-language alignment with noise. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 17449–17457, 2025. [50] Junying Wang, Hongyuan Zhang, and Yuan Yuan. Adv-cpg: A customized portrait generation framework with facial adversarial attacks. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21001–21010, 2025. # NeurIPS Paper Checklist # 1. Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: None Guidelines: • The answer NA means that the abstract and introduction do not include the claims made in the paper. • The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. • The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. • It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. # 2. Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: None Guidelines: • The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. • The authors are encouraged to create a separate "Limitations" section in their paper. • The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. • The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. • If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren’t acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. # 3. Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [Yes] Justification: None Guidelines: • The answer NA means that the paper does not include theoretical results. • All the theorems, formulas, and proofs in the paper should be numbered and crossreferenced. • All assumptions should be clearly stated or referenced in the statement of any theorems. • The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. • Theorems and Lemmas that the proof relies upon should be properly referenced. # 4. Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? Answer: [Yes] Justification: None Guidelines: • The answer NA means that the paper does not include experiments. • If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. • If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general. releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. • While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example (a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. (b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. (c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). (d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. # 5. Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: Due to security review requirements for the code and data, the code will be open-sourced upon paper acceptance. Guidelines: • The answer NA means that paper does not include experiments requiring code. • Please see the NeurIPS code and data submission guidelines (https://nips.cc/ public/guides/CodeSubmissionPolicy) for more details. • While we encourage the release of code and data, we understand that this might not be possible, so “No” is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). • The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https: //nips.cc/public/guides/CodeSubmissionPolicy) for more details. • The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. • The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. • At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). • Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. # 6. Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: None Guidelines: • The answer NA means that the paper does not include experiments. • The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. • The full details can be provided either with the code, in appendix, or as supplemental material. # 7. Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [No] Justification: Constrained by computational resources, we report results from a single run. Guidelines: • The answer NA means that the paper does not include experiments. • The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. • The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). • The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) • The assumptions made should be given (e.g., Normally distributed errors). • It should be clear whether the error bar is the standard deviation or the standard error of the mean. • It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a $96 \%$ CI, if the hypothesis of Normality of errors is not verified. • For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). • If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. # 8. Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: None Guidelines: • The answer NA means that the paper does not include experiments. • The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. • The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. • The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn’t make it into the paper). # 9. Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? Answer: [Yes] Justification: None Guidelines: • The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics. • If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. • The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). # 10. Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [No] Justification: The scope of this work is limited to academic research in computer science, with no immediate further societal impact. Guidelines: • The answer NA means that there is no societal impact of the work performed. • If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. • Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. • The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. • The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). # 11. Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? Answer: [NA] Justification:The scope of this work is limited to academic research in computer science, with no immediate further safeguards impact. Guidelines: • The answer NA means that the paper poses no such risks. • Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. • Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. • We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. # 12. Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: None Guidelines: • The answer NA means that the paper does not use existing assets. • The authors should cite the original paper that produced the code package or dataset. • The authors should state which version of the asset is used and, if possible, include a URL. • The name of the license (e.g., CC-BY 4.0) should be included for each asset. • For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. • If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. • For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. • If this information is not available online, the authors are encouraged to reach out to the asset’s creators. # 13. New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [NA] Justification: The paper does not release new assets. Guidelines: • The answer NA means that the paper does not release new assets. • Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. • The paper should discuss whether and how consent was obtained from people whose asset is used. • At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. # 14. Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? Answer: [NA] Justification: The paper does not involve crowdsourcing nor research with human subjects. Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. • Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. • According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. # 15. Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? Answer: [NA] Justification: The paper does not involve crowdsourcing nor research with human subjects. Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. • Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. • We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. • For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. # 16. Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. Answer: [Yes] Justification: None Guidelines: • The answer NA means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. • Please refer to our LLM policy (https://neurips.cc/Conferences/2025/LLM) for what should or should not be described. # Appendix A: Fourier Analysis in Natural Images This appendix provides essential mathematical formulations and conceptual insights into the Fourier analysis of natural images, building upon the concepts discussed in the main text. We focus on the Discrete Fourier Transform (DFT) and its implications for image representation and the characteristics of natural scenes. # A.1 2D Discrete Fourier Transform (2D DFT) The 2D DFT transforms an $M \times N$ digital image $f ( x , y )$ from the spatial domain (where $x \in$ $\{ 0 , \ldots , M - 1 \}$ and $y \in \{ 0 , \ldots , N - 1 \}$ are spatial coordinates) to the frequency domain, yielding an $M \times N$ representation $F ( u , v )$ (where $u \in \{ 0 , \ldots , M - 1 \}$ and $v \in \{ 0 , \dotsc , N ^ { \cdot } - 1 \}$ are frequency coordinates). The formula is given by: $$ F ( u , v ) = \sum _ { x = 0 } ^ { M - 1 } \sum _ { y = 0 } ^ { N - 1 } f ( x , y ) e ^ { - j 2 \pi \left( { \frac { u x } { M } } + { \frac { v y } { N } } \right) } $$ Here, $j$ is the imaginary unit $( j ^ { 2 } = - 1 )$ ), and the exponential term represents the basis functions (complex sinusoids) at different frequencies $( u , v )$ . # A.2 Inverse 2D Discrete Fourier Transform (2D IDFT) The 2D IDFT allows us to reconstruct the original spatial domain image $f ( x , y )$ from its frequency domain representation $F ( u , v )$ . The formula is: $$ f ( x , y ) = { \frac { 1 } { M N } } \sum _ { u = 0 } ^ { M - 1 } \sum _ { v = 0 } ^ { N - 1 } F ( u , v ) e ^ { j 2 \pi \left( { \frac { u x } { M } } + { \frac { v y } { N } } \right) } $$ Note the scaling factor $\frac { 1 } { M N }$ and the positive sign in the exponent compared to the forward transform. # A.3 Magnitude Spectrum and Power Spectrum The frequency domain representation $F ( u , v )$ obtained from the DFT is generally a complex number. Its magnitude, $| F ( u , v ) |$ , is known as the Magnitude Spectrum, which quantifies the amplitude of each frequency component present in the image. $$ | F ( u , v ) | = \sqrt { \mathrm { R e } ( F ( u , v ) ) ^ { 2 } + \mathrm { I m } ( F ( u , v ) ) ^ { 2 } } $$ Closely related is the Power Spectrum (or Power Spectral Density), defined as the square of the magnitude spectrum. It represents how the total signal energy is distributed across the different frequencies: $$ P ( u , v ) = | F ( u , v ) | ^ { 2 } $$ A key characteristic of natural images is that their power spectrum typically exhibits a rapid decay as frequency increases. Specifically, the power √ $P ( u , v )$ tends to fall off with increasing radial frequency $f _ { r } = \sqrt { u ^ { 2 } + v ^ { 2 } }$ , often approximated by a $1 / f _ { r } ^ { \alpha }$ law, where $\alpha$ is a constant typically around 2. This 1/f property implies that low spatial frequencies (corresponding to coarse structures and overall variations) contain significantly more energy than high spatial frequencies (corresponding to fine details and sharp transitions). This fundamental statistical feature of natural images is widely utilized and modeled in various image processing and computer vision tasks. # Appendix B: Addition Experiments VAR sets different training epochs for models of different sizes: 200 epochs for 310M, 250 epochs for 600M, 300 epochs for 1B, and 350 epochs for 2B. Limited by computational resources, we focus on training our 310M model. As the table above shows, NFIG outperforms VAR at matched epochs and demonstrates superior parameter efficiency. At 200 epochs, NFIG already outperforms VAR-d16 of the same size. With extended training, NFIG-310M achieves performance comparable to or better than VAR-d20-600M (twice the parameters) while using significantly fewer resources. # B.1 Details of Loss function We will provide detailed mathematical formulations of our loss function components and their respective roles in the training process. The total loss function for the FR-VAE (image tokenizer for NFIG) is defined as: $$ \mathcal { L } = | | I - \hat { I } | | _ { 2 } ^ { 2 } + | | \hat { f } - \hat { f } | | _ { 2 } ^ { 2 } + \mathcal { L } _ { p } ( I ) + 0 . 5 \mathcal { L } _ { g } ( I ) . $$ Here, the first two terms represent the reconstruction loss for the image $I$ vs $\hat { I } _ { { \bf \Pi } _ { \it - \hat { \Pi } } }$ ) and and its frequencyguided quantized loss $( \hat { f } \ \mathbf { v } \mathbf { s } \ \hat { f } )$ , respectively, ensuring fidelity in both pixel and feature. ${ \mathcal { L } } _ { p }$ is LPIPS perceptual loss and $\mathcal { L } _ { g }$ is gan loss. For the NFIG Transformer, which predicts the frequency tokens, we utilize a standard cross-entropy loss: $$ \mathcal { L } ( T , \tilde { T } ) = - \sum _ { i = 1 } ^ { n } t _ { i } \log ( \tilde { t } _ { i } ) $$ This loss is computed between the predicted tokens $\tilde { T }$ and FR-VAE ground truth tokens $T$ , ensuring accurate prediction of the quantized frequency representations across all scales. # B.2 Frequency Keep Ability We are grateful for your encouragement to discuss both successes and challenges. Your suggestion for a frequency analysis was particularly insightful. As you requested, we conducted a frequency-domain comparison between our model (NFIG) and VAR-16. As requested, we provide frequency-domain comparisons between VAR-16 and NFIG: (1) Power Spectral Density (PSD): Overall frequency fidelity; (2) Frequency Keep Score (FKS): Weighted similarity across High/Mid/Low frequency bands (weights: 0.15, 0.28, 0.57, emphasizing structural low-frequency information). Our analysis revealed that while both models effectively preserve low-frequency information, and NFIG preserves middle and high frequency information with higher fidelity.
ModelPSD↓FKS↑Low↑Middle↑High↑
VAR-160.8779.5%98.3%57.6%48.2%
NFIG(ours)0.4787.6%98.9%75.3%66.7%
# B.4 Diverse Image Types To demonstrate broader applicability, we conducted preliminary reconstruction evaluations (FID) of FR-VAE across diverse image types.
ModelDTDQRCODEDiagramsChest-XCelebA-HQCOCOLSUN-Bedroom
FR-VAE6.8611.01210.743.517.516.12
# B.5 Failure Case Despite strong overall performance, NFIG occasionally produces visual artifacts. As shown in Figure 6, these include anatomical errors (extra bird leg), texture abnormalities (goldfish patterns), and fine detail loss (bird claws). Red boxes highlight the anomalies. These failures reflect challenges in maintaining semantic consistency across frequency bands. The issues are particularly pronounced for complex structures and fine details. Such limitations are common to frequency-based generation approaches and present opportunities for future improvement. ![](images/figures/nfig-fig-0006.jpg) Figure 6: Failure case for NFIG.