FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
Sucheng Ren 1Qihang Yu†2Ju He†2Xiaohui Shen 2Alan Yuille 1Liang-Chieh Chen†2
1Johns Hopkins University 2ByteDance
Abstract
Autoregressive (AR) modeling has achieved re-
markable success in natural language processing
by enabling models to generate text with coher-
ence and contextual understanding through next
token prediction. Recently, in image generation,
VAR proposes scale-wise autoregressive model-
ing, which extends the next token prediction to
the next scale prediction, preserving the 2D struc-
ture of images. However, VAR encounters two
primary challenges: (1) its complex and rigid
scale design limits generalization in next scale
prediction, and (2) the generator’s dependence
on a discrete tokenizer with the same complex
scale structure restricts modularity and flexibil-
ity in updating the tokenizer. To address these
limitations, we introduce FlowAR, a general next
scale prediction method featuring a streamlined
scale design, where each subsequent scale is sim-
ply double the previous one. This eliminates
the need for VAR’s intricate multi-scale resid-
ual tokenizer and enables the use of any off-the-
shelf Variational AutoEncoder (VAE). Our sim-
plified design enhances generalization in next
scale prediction and facilitates the integration of
Flow Matching for high-quality image synthesis.
We validate the effectiveness of FlowAR on the
challenging ImageNet-256 benchmark, demon-
strating superior generation performance com-
pared to previous methods. Codes is available
at https://github.com/OliverRensu/FlowAR.
1. Introduction
Autoregressive (AR) models have significantly advanced
natural language processing (NLP) by modeling the proba-
bility distribution of each token given its preceding tokens,
1
Johns Hopkins University
2
ByteDance. Correspondence
to: Sucheng Ren
<
sren19@jh.edu
>
.
†
: Work done while at
ByteDance.
Proceedings of the
42 nd
International Conference on Machine
Learning, Vancouver, Canada. PMLR 267, 2025. Copyright 2025
by the author(s).
~100M ~600M
1.6
3
FID
Parameters
~300M ~2B
lower is better
VAR
FlowAR (Ours)
6
-2.10
-0.65
-1.05
-0.32
Figure 1.
Performance Comparison. The proposed FlowAR, a
general next-scale prediction model enhanced with flow matching,
consistently outperforms state-of-the-art VAR (Tian et al.,2024)
variants across different model sizes.
allowing for coherent and contextually relevant text gener-
ation. Prominent models like GPT-3 (Brown et al.,2020)
and its successors (OpenAI,2022;2023) have demonstrated
remarkable language understanding and generation capabili-
ties, setting new standards across diverse NLP applications.
Building on the success of autoregressive modeling in NLP,
this paradigm has been adapted to computer vision, particu-
larly for generating high-fidelity images through sequential
content prediction (Esser et al.,2021;Yu et al.,2022;Sun
et al.,2024). In these approaches, images are discretely tok-
enized, with tokens flattened into 1D sequences, enabling
autoregressive models to generate images token-by-token.
This approach leverages the sequence modeling strengths of
AR architectures to capture intricate visual details. However,
directly applying 1D token-wise autoregressive methods to
images presents notable challenges. Images are inherently
two-dimensional (2D), with spatial dependencies across
height and width. Flattening image tokens disrupts this 2D
structure, potentially compromising spatial coherence and
causing artifacts in generated images. Recently, VAR (Tian
et al.,2024) addresses these issues by introducing scale-
1
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
wise autoregressive modeling, which progressively gener-
ates images from coarse to fine scales, preserving the spatial
hierarchies and dependencies essential for visual coherence.
This scale-wise approach allows autoregressive models to
retain the 2D structure during image generation, capturing
the layered complexity of visual content more naturally.
Despite its effectiveness, VAR (Tian et al.,2024) faces
two significant limitations: (1) a complex and rigid scale
design, and (2) a dependency between the generator
and a tokenizer that shares this intricate scale structure.
Specifically, VAR employs a non-uniform scale sequence,
{1,2,3,4,5,6,8,10,13,16}
, where the coarsest scale tok-
enizes a
256 ×256
image into a single
1×1
token and the
finest scale into
16 ×16
tokens. This intricate sequence
constrains both the tokenizer and generator to operate ex-
clusively at these predefined scales, limiting adaptability to
other resolutions or granularities. Consequently, the model
struggles to represent or generate features that fall outside
this fixed scale sequence. Additionally, the tight coupling
between VAR’s generator and tokenizer restricts flexibility
in independently updating the tokenizer, as both components
must adhere to the same scale structure.
To address these limitations, we introduce FlowAR, a flexi-
ble and generalized approach to scale-wise autoregressive
modeling for image generation, enhanced with flow match-
ing (Lipman et al.,2022). Unlike VAR (Tian et al.,2024),
which relies on a complex multi-scale VQGAN discrete
tokenizer (Razavi et al.,2019;Lee et al.,2022), we uti-
lize any off-the-shelf VAE continuous tokenizer (Kingma &
Welling,2014) with a simplified scale design, where each
subsequent scale is simply double the previous one (e.g.,
{1,2,4,8,16}
), and coarse scale tokens are obtained by di-
rectly downsampling the finest scale tokens (i.e., the largest
resolution token map). This streamlined design eliminates
the need for a specially designed tokenizer and decouples
the tokenizer from the generator, allowing greater flexibility
to update the tokenizer with any modern VAE (Li et al.,
2024;Chen et al.,2024).
To further enhance image quality, we incorporate the flow
matching model (Lipman et al.,2022) to learn the probabil-
ity distribution at each scale. Specifically, given the class
token and tokens from previous scales, we use an autore-
gressive Transformer (Radford et al.,2018) to generate con-
tinuous semantics that condition the flow matching model,
progressively transforming noise into the target latent rep-
resentation for the current scale. Conditioning is achieved
through the proposed spatially adaptive layer normalization
(Spatial-adaLN), which adaptively adjusts layer normaliza-
tion (Ba et al.,2016) on a position-by-position basis, captur-
ing fine-grained details and improving the model’s ability
to generate high-fidelity images. This process is repeated
across scales, capturing the hierarchical dependencies inher-
ent in natural images. The final image is then produced by
de-tokenizing the predicted latent representation at the finest
scale. As shown in Figure 1, Our FlowAR consistently out-
perform VAR under similar parameters. Specifically, with
about 2B parameters our FlowAR outperform VAR by 0.32
FID.
The seamless integration of the scale-wise autoregressive
Transformer and scale-wise flow matching model enables
FlowAR to capture both the sequential and probabilistic
aspects of images with multi-scale information, resulting in
improved image synthesis performance. We demonstrate
FlowAR’s effectiveness on the challenging ImageNet bench-
mark (Deng et al.,2009), where it achieves state-of-the-art
results.
2. Related Work
Autoregressive Models. Autoregressive modeling (Radford
et al.,2018;2021;Brown et al.,2020;OpenAI,2023;Tou-
vron et al.,2023a;b;Chowdhery et al.,2023;Team et al.,
2023;Ren et al.,2024b;a;2025b;a;Wang et al.,2024;2025;
Dubey et al.,2024;Bai et al.,2023;Yu et al.,2024a) began
in natural language processing, where language Transform-
ers (Vaswani,2017) are pretrained to predict the next word
in a sequence. This concept was first introduced to computer
vision by PixelCNN (Van den Oord et al.,2016), which uti-
lized a CNN-based model to predict raw pixel probabilities.
With the advent of Transformers, iGPT (Chen et al.,2020)
extended this approach by modeling raw pixels using Trans-
former architectures. VQGAN (Esser et al.,2021;Weber
et al.,2024) further advanced the field by applying autore-
gressive learning within the latent space of VQ-VAE (Van
Den Oord et al.,2017), thereby simplifying data represen-
tatiown for more efficient modeling (Yu et al.,2024b;Kim
et al.,2025). Taking a different direction, Parti (Yu et al.,
2022) framed image generation as a sequence-to-sequence
task akin to machine translation, using sequences of image
tokens as targets instead of text tokens, and leveraging sig-
nificant advancements in large language models through
data and model scaling. LlamaGen (Sun et al.,2024) ex-
panded on this by applying the traditional “next token pre-
diction” paradigm of large language models to visual genera-
tion, demonstrating that standard autoregressive models like
Llama (Touvron et al.,2023a) can achieve state-of-the-art
image generation performance when appropriately scaled,
even without specific inductive biases for visual signals. The
work most similar to ours is VAR (Tian et al.,2024), which
transitioned from token-wise to scale-wise autoregressive
modeling by developing a coarse-to-fine next scale predic-
tion. However, VAR faces significant challenges due to its
complex scale designs and deep dependency on a scale resid-
ual discrete tokenizer. In contrast, our proposed FlowAR
employs a simple scale design and maintains compatibility
2
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
with any VAE tokenizer (Kingma & Welling,2014).
Flow- and Diffusion-based Models. Diffusion mod-
els (Rombach et al.,2022;Peebles & Xie,2023;Li et al.,
2024;Hoogeboom et al.,2023;Ho et al.,2020;Song et al.,
2020;Liu et al.,2024;Yang et al.,2024) have surpassed
earlier image generation methods like GANs (Goodfellow
et al.,2014;Sauer et al.,2022) by utilizing multistep dif-
fusion and denoising processes. Latent Diffusion Models
(LDMs) (Rombach et al.,2022) advance this approach by
transitioning diffusion from pixel space to latent space, en-
hancing efficiency and scalability. Building on this foun-
dation, DiT (Peebles & Xie,2023) and U-ViT (Bao et al.,
2023) replace the traditional convolution-based U-Net (Ron-
neberger et al.,2015) with Transformer architectures within
the latent space, further improving performance. Flow
matching (Lipman et al.,2022;Liu et al.,2022;Albergo &
Vanden-Eijnden,2022;Shin et al.,2025;He et al.,2025)
redefines the forward process as direct paths between the
data distribution and a standard normal distribution, offering
a more straightforward transition from noise to target data
compared to conventional diffusion methods. SiT (Atito
et al.,2021) leverages the backbone of DiT and employs
flow matching to more directly connect these distributions.
Scaling this concept, SD3 (Esser et al.,2024) introduces
a novel Transformer-based architecture trained with flow
matching for text-to-image generation (Deng et al.,2025).
MAR (Li et al.,2024) presents a diffusion-based strategy
to model per-token probability distributions in a continuous
space, enabling autoregressive models without relying on
discrete tokenizers and utilizing a specialized diffusion loss
function instead of the traditional categorical cross-entropy
loss. In contrast to MAR, our proposed FlowAR employs a
scale-wise flow matching model to capture per-scale prob-
abilities, utilizing coarse-to-fine scale-wise conditioning
derived from a scale-wise autoregressive model.
Discussion. Our FlowAR provides a more flexible frame-
work for next scale prediction, enhanced with flow matching.
In Figure 2, we compare our FlowAR model with VAR (Tian
et al.,2024), highlighting key differences in both the tok-
enizer and generator components. For the tokenizer, VAR
relies on a multi-scale VQGAN (Razavi et al.,2019;Esser
et al.,2021;Lee et al.,2022) that is tightly integrated with
its generator and trained on a complex set of scales (
{
1, 2,
3, 4, 5, 6, 8, 10, 13, 16
}
). In contrast, FlowAR can use
any off-the-shelf VAE (Kingma & Welling,2014) as its to-
kenizer, offering greater flexibility by constructing coarse
scale token maps through direct downsampling of the finest
scale token map. For the generator, VAR is constrained by a
complex and rigid scale structure, whereas FlowAR benefits
from a simpler and more general scale design, allowing the
integration of the modern flow matching model (Lipman
et al.,2022).
3. Method
In this section, we begin with an overview of autoregressive
modeling in Sec. 3.1, followed by a detailed introduction of
the proposed method in Sec. 3.2.
3.1. Preliminary: Autoregressive Modeling
Autoregressive Modeling in NLP. Consider a corpus repre-
sented as a sequence of words
U={w1, w2, . . . , wn}
. In
NLP, autoregressive models predict each word based on all
preceding words in the sequence:
p(U) =
n
Y
k=1
p(wk|w1, w2, . . . , wk−1,Θ),(1)
where the autoregressive model is parameterized by
Θ
. The
objective function to minimize is the negative log-likelihood
over the entire corpus:
L=−
n
X
k=1
log p(wk|w<k,Θ),(2)
where
< k
denotes all positions preceding
k
. This approach
serves as the foundation for successful large-scale language
models, as demonstrated in (Touvron et al.,2023a;OpenAI,
2022;2023).
Token-wise Autoregressive Modeling for Image Gener-
ation. Extending autoregressive modeling to images in-
volves tokenizing a 2D image
X∈R3×H×W
using vector
quantization (Van Den Oord et al.,2017). This process
converts the image into a grid of discrete tokens, which
is subsequently flattened into a one-dimensional sequence
X={t1, t2, . . . , tN}:
L=−
N
X
k=1
log p(tk|t<k,Θ).(3)
However, flattening the token grid disrupts the intrinsic two-
dimensional spatial structure of the image. To preserve this
spatial information, VAR (Tian et al.,2024) introduces a
scale-wise autoregressive modeling approach, as described
below.
Scale-wise Autoregressive Modeling for Image Genera-
tion. Rather than flattening images into token sequences,
VAR (Tian et al.,2024) decomposes the image into a series
of token maps across multiple scales,
S={s1, s2, . . . , sn}
.
Each token map
sk
has dimensions
hk×wk
and is obtained
by a specially designed multi-scale VQGAN with residual
structure (Razavi et al.,2019;Lee et al.,2022;Tian et al.,
2024;Esser et al.,2021). In contrast to single flattened
tokens
tk
that lose spatial context, each
sk
maintains the
two-dimensional structure with
hk×wk
tokens. The autore-
gressive loss function is then reformulated to predict each
3
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
Discrete Tokens Decoder
Complex and Rigid Scale Design
{1, 2, 3, 4, 5, 6, 8, 10, 13, 16}
Simple and General Scale Design
{1, 2, 4, 8, 16}
Decoder
Scale-wise
Autoregressive
Scale-wise
Autoregressive
Semantics
Scale-wise
Flow Matching
Scale-wise
Flow Matching
×
√
Multi-scale
VQGAN
Downsample
Bonded with Generator
Off-the-shelf
VAR
FlowAR
Multi-Scale
Residual
×
√
(a) Tokenizer Comparison (b) Generator Comparison
Discrete
Tokens
Continuous
Latents
Any
VAE
Any
VAE
Figure 2.
Comparison Between VAR and Our FlowAR in (a) Tokenizer and (b) Generator design. (a) VAR (Tian et al.,2024) utilizes a
complex multi-scale residual VQGAN discrete tokenizer, whereas FlowAR can leverage any off-the-shelf VAE continuous tokenizer,
constructing coarse scale token maps by directly downsampling the finest scale token map. (b) VAR’s generator is constrained by the
same complex and rigid scale design as its tokenizer, while FlowAR benefits from a simple and flexible scale design, enhanced by the
flow matching model.
scale based on all preceding scales:
L=−
n
X
k=1
log p(sk|s<k,Θ).(4)
In this framework, generating the
k
-th scale in VAR re-
quires attending to all previous scales
s<k
(i.e.,
s1
to
sk−1
) and simultaneously predicting all
hk×wk
tokens
in
sk
via categorical distributions. The chosen scales,
S={1,2,3,4,5,6,8,10,13,16}
, introduce significant
complexity to the scale design and constrain the model’s
generalization capabilities. This is due to the tight coupling
between the generator and tokenizer with the scale design,
reducing flexibility in updating the tokenizer or support-
ing alternative scale configurations. Furthermore, VAR’s
discrete tokenizer relies on a complex multi-scale residual
structure, complicating the training process, with essential
training code and details remaining publicly unavailable at
the time of our submission.
3.2. Proposed Method: FlowAR
Overview. To address the issues outlined above, we intro-
duce FlowAR seen in Figure 3, a more general scale-wise
autoregressive modeling, enhanced with flow matching (Lip-
man et al.,2022). Our method incorporates two primary
improvements over existing next scale prediction (Tian
et al.,2024): (1) replacing the multi-scale VQGAN dis-
crete tokenizer with any off-the-shelf VAE continuous to-
kenizer (Kingma & Welling,2014), and (2) modeling the
per-scale prediction (i.e., predicting all
hk×wk
tokens in
k
-th scale
sk
) using flow matching to learn the probabil-
ity distribution. The first change enables the flexibility to
leverage any existing VAE tokenizer, benefiting from recent
advances in VAE technology without being constrained by
a complex scale sequence design. The second improvement
enhances generation quality by utilizing the modern flow
matching algorithm.
Simple Scale Sequence with Any VAE Tokenizer. Given
an image, we extract its continuous latent representation
F∈Rc×h×w
using an off-the-shelf Variational Autoen-
coder (VAE) (Kingma & Welling,2014). We then construct
a set of coarse-to-fine scales by downsampling the latent
F
as follows:
S={s1, s2, . . . , sn}
={Down(F, 2n−1),Down(F, 2n−2),...,Down(F, 1)},
(5)
where
Down(F, r)
denotes downsampling the latent
F
by a
factor of r, and no downsampling is applied when r= 1.
Notably, our multi-scale token maps
S
are derived by di-
rectly downsampling the highest-resolution latent
F
, re-
moving the need for complex multi-scale residual VQGAN
design (Tian et al.,2024;Razavi et al.,2019;Lee et al.,
4
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
[C]
1
Upsample
2
3 4
Upsample
1 2
3 4
1 2
3 4
Flow
Matching
Flow
Matching
Flow
Matching
Semantics
Semantics
Semantics
1
Autoregressive
1
1 42 31 42 3
5 86 75 86 7
91210 1191210 11
13 1614 1513 1614 15
1 42 3
5 86 7
91210 11
13 1614 15
1 42 31 42 3
5 86 75 86 7
91210 1191210 11
13 1614 1513 1614 15
1 42 3
5 86 7
91210 11
13 1614 15
Downsample
Any
VAE
Any
VAE
𝑠 2
𝑠 1
𝑠 3
𝑠1
𝑠2
Velocity
Velocity
Velocity
𝑉
𝑡3
𝑉
𝑡2
𝑉
𝑡
1
Figure 3.
Overview of The Proposed FlowAR. FlowAR consists of three main components: (1) an off-the-shelf VAE that extracts a
continuous latent representation of the image. We then create a set of coarse-to-fine scales by downsampling this latent, forming a sequence
of token maps
s1, s2,· · · , sn
, where each subsequent scale doubles in size from the previous one. (2) A scale-wise autoregressive
Transformer that takes as input the sequence
{[C],Up(s1,2),...,Up(sn−1,2)}
, where
[C]
is a condition token and
Up(·,2)
denotes
upsampling by a factor of 2. This Transformer generates semantic representations for different scales,
ˆs1,...,ˆsn
. (3) A scale-wise flow
matching model, conditioned on the semantics
ˆsi
at each scale
i
(time step conditions are not shown for simplicity), predicts the velocity
given a random time step tthat moves the noises to the target data distribution.
2022;Esser et al.,2021). To distinguish our approach, we
use the superscript
i
to denote our
i
-th scale token map,
si
.
We set
n= 5
(i.e.,
S={1,2,4,8,16}
), making the scale
design in FlowAR significantly simpler and more versatile
than that of VAR (Tian et al.,2024). This design allows us
to integrate various off-the-shelf VAE tokenizers into our
framework, eliminating the need for a multi-scale tokenizer
trained on a predefined scale sequence.
With this simplified scale sequence and the flexibility to use
any off-the-shelf VAE tokenizer, we introduce our improve-
ment for next scale prediction. Unlike VAR (Tian et al.,
2024), which models the categorical distribution of each
scale using an autoregressive Transformer, we utilize the
Transformer to generate conditioning information for each
scale, while a scale-wise flow matching model captures the
scale’s probability distribution based on this information.
Below, we outline how the autoregressive Transformer gen-
erates conditioning information for each scale, followed by
details of the scale-wise flow matching module.
Generating Conditioning Information via Scale-wise Au-
toregressive Transformer. To produce the conditioning
information for each subsequent scale, we utilize conditions
obtained from all previous scales:
ˆsi=TC, Up(s1,2),...,Up(si−1,2),∀i= 1,· · · , n,
(6)
where
T(·)
represents the autoregressive Transformer model,
C
is the class condition, and
Up(s, r)
denotes the upsam-
pling of latent
s
by a factor of
r
. For the initial scale (
i= 1
),
only the class condition
C
is used as input. We set
r= 2
, fol-
lowing our simple scale design, where each scale is double
the size of the previous one. We refer to the resulting output
ˆsi
as the semantics for the
i
-th scale, which is then used
to condition a flow matching module to learn the per-scale
probability distribution.
Scale-wise Flow Matching Model Conditioned by Autore-
gressive Transformer Output. Flow matching (Lipman
et al.,2022) generates samples from the target data distribu-
tion by gradually transforming a source noisy distribution,
such as a Gaussian. For each
i
-th scale, FlowAR extends
flow matching to generate the scale latent
si
, conditioned
on the autoregressive Transformer’s output
ˆsi
. Specifically,
during training, given the scale latent
si
from the target data
distribution, we sample a time step
t∈[0,1]
and a source
noise sample
Fi
0
from the source noisy distribution, typically
setting
Fi
0∼ N (0,1)
to match the shape of conditioned la-
tent
ˆsi
, analogous to the “noise” in diffusion models (Rom-
bach et al.,2022). We then construct the interpolated input
Fi
tas:
Fi
t=tsi+ (1 −t)Fi
0.(7)
The model is trained to predict the velocity Vi
tusing Fi
t:
Vi
t=dF i
t
dt
=si−Fi
0,
(8)
where
Vi
t
indicates the direction to move from
Fi
t
toward
si
,
5
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
guiding the transformation from the source to the target dis-
tribution at each scale. Unlike prior approaches (Atito et al.,
2021;Esser et al.,2024) that condition velocity prediction
on class or textual information, we condition on the scale-
wise semantics
ˆsi
from the autoregressive Transformer’s
output. Notably, in prior methods (Atito et al.,2021;Esser
et al.,2024), the conditions and image latents often have
different lengths, whereas FlowAR shares the same length
(i.e.,
si
and
ˆsi
have the same shape,
∀i= 1,· · · , n
). The
training objective for scale-wise flow matching is:
L=
n
X
i=1
FM Fi
t,ˆsi, t;θ−Vi
t
2,(9)
where
FM
denotes the flow matching model parameterized
by
θ
. This approach allows the model to capture scale-
wise information bi-directionally, enhancing flexibility and
efficiency in image generation within our framework.
Scale-wise Injection of Semantics via Spatial-adaLN. A
key design choice is determining how best to inject the
semantic information
ˆsi
, generated by the autoregressive
Transformer, into the flow matching module. A straightfor-
ward approach would be to concatenate the semantics
ˆsi
with the flattened input
Fi
t
, similar to in-context condition-
ing (Bao et al.,2023) where the class condition is concate-
nated with the input sequence. However, this approach has
two main drawbacks: (1) it increases the sequence length
input to the flow matching model, raising computational
costs, and (2) it provides indirect semantic injection, poten-
tially weakening the effectiveness of semantic guidance. To
address these issues, we propose using spatially adaptive
layernorm for position-by-position semantic injection, re-
sulting in the proposed Spatial-adaLN. Specifically, given
the semantics
ˆsi
from the scale-wise autoregressive Trans-
former and the intermediate feature
Fi′
t
in the flow matching
model, we inject the semantics to the scale
γ
, shift
β
, and
gate
α
parameters of the adaptive normalization, following
the standard adaptive normalization procedure (Ba et al.,
2016;Zhang & Sennrich,2019;Peebles & Xie,2023):
α1, α2, β1, β2, γ1, γ2= MLP(ˆsi+t),
ˆ
Fi′
t= Attn γ1⊙LN(Fi′
t) + β1⊙α1,
Fi′′
t= MLP γ2⊙LN( ˆ
Fi′
t) + β2⊙α2,
(10)
where
Attn
denotes the attention mechanism, and
LN
de-
notes the layer norm,
Fi′′
t
is the block’s output, and
⊙
is the
spatial-wise product. Unlike traditional adaptive normaliza-
tion, where the scale, shift, and gate parameters lack spatial
information, our spatial adaptive normalization provides po-
sitional control, enabling dependency on semantics from the
autoregressive Transformer. We apply the Spatial-adaLN
to every Transformer block within the flow matching mod-
ule. Alternative design choices are explored in the ablation
study.
Inference Pipeline. At the beginning of inference, the au-
toregressive Transformer generates the initial semantics
ˆs1
using only the class condition
C
. This semantics
ˆs1
condi-
tions the flow matching module, which gradually transforms
a noise sample into the target distribution for
s1
. The result-
ing token map is upsampled by a factor of 2, combined with
the class condition, and fed back into the autoregressive
Transformer to generate the semantics
ˆs2
, which conditions
the flow matching module for the next scale. This process
is iterated
n
scales until the final token map
sn
is estimated
and subsequently decoded by the VAE decoder to produce
the generated image. Notably, we use the KV cache (Pope
et al.,2023) in the autoregressive Transformer to efficiently
generate each semantics ˆsi.
4. Experimental Results
In this section, we present our main results on the chal-
lenging ImageNet-256 and ImageNet-512 generation bench-
mark (Deng et al.,2009) (Sec. 4.1), followed by ablation
studies (Sec. 4.2).
4.1. Main Results
Following the settings in VAR (Tian et al.,2024), we train
FlowAR on ImageNet for class-conditional image gener-
ation. We evaluate the model using Fr
´
echet Inception
Distance (FID) (Heusel et al.,2017) and inception score
(IS) (Salimans et al.,2016), Precision (Pre.) (Powers,2020)
and Recall (rec.) (Powers,2020) as metrics.
Quantitative Results on ImageNet-256. As shown in Ta-
ble 1, when compared to previous generative adversarial
models (Brock et al.,2018;Sauer et al.,2022;Kang et al.,
2023), autoregressive models (Razavi et al.,2019;Esser
et al.,2021;Lee et al.,2022;Sun et al.,2024;Yu et al.,
2021), diffusion-based methods (Dhariwal & Nichol,2021;
Rombach et al.,2022;Peebles & Xie,2023;Bao et al.,
2023), and flow matching methods (Atito et al.,2021),
FlowAR achieves significant performance gains. Specif-
ically, our best model variant, FlowAR-H, attains an FID of
1.65, outperforming StyleGAN (Sauer et al.,2022) (2.30),
LlamaGen-3B (Sun et al.,2024) (2.18), DiT (Peebles & Xie,
2023) (2.27), and SiT (Atito et al.,2021) (2.06).
Compared to the closely related VAR (Tian et al.,2024),
FlowAR provides superior image quality at similar model
scales. For example, FlowAR-L, with 589M parameters,
achieves an FID of 1.90—surpassing both VAR-d20 (FID
2.95) of comparable size and even largest VAR-d30 (FID
1.97), which has 2B parameters. Furthermore, our largest
6
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
method type params FID ↓IS ↑Pre.↑Rec.↑
BigGAN (Brock et al.,2018) GAN 112M 6.95 224.5 0.89 0.38
GigaGAN (Kang et al.,2023) GAN 569M 3.45 225.5 0.84 0.61
StyleGAN (Sauer et al.,2022) GAN 166M 2.30 265.1 0.78 0.53
ADM (Dhariwal & Nichol,2021) diffusion 554M 10.94 101.0 0.69 0.63
LDM (Rombach et al.,2022) diffusion 400M 3.60 247.7 - -
U-ViT(Bao et al.,2023) diffusion 287M 3.40 219.9 0.83 0.52
DiT (Peebles & Xie,2023) diffusion 675M 2.27 278.2 0.83 0.57
SiT (Atito et al.,2021) flow matching 675M 2.06 270.3 0.82 0.59
VQVAE-2 (Razavi et al.,2019) token-wise autoregressive 13.5B 31.11 45.0 0.36 0.57
VQGAN (Esser et al.,2021) token-wise autoregressive 1.4B 15.78 74.3 - -
RQTransformer (Lee et al.,2022) token-wise autoregressive 3.8B 7.55 134.0 - -
LlamaGen-B (Sun et al.,2024) token-wise autoregressive 111M 5.46 193.6 0.83 0.45
ViT-VQGAN (Yu et al.,2021) token-wise autoregressive 1.7B 4.17 175.1 - -
LlamaGen-L (Sun et al.,2024) token-wise autoregressive 343M 3.07 256.1 0.83 0.52
LlamaGen-XL (Sun et al.,2024) token-wise autoregressive 775M 2.62 244.1 0.80 0.57
LlamaGen-3B (Sun et al.,2024) token-wise autoregressive 3.1B 2.18 263.3 0.81 0.58
VAR-d12 (Tian et al.,2024) scale-wise autoregressive 132M 5.81 201.3 0.81 0.45
FlowAR-S scale-wise autoregressive 170M 3.61 234.1 0.83 0.50
VAR-d16 (Tian et al.,2024) scale-wise autoregressive 310M 3.55 280.4 0.84 0.51
FlowAR-B scale-wise autoregressive 300M 2.90 272.5 0.84 0.54
VAR-d20 (Tian et al.,2024) scale-wise autoregressive 600M 2.95 302.6 0.83 0.56
FlowAR-L scale-wise autoregressive 589M 1.90 281.4 0.83 0.57
VAR-d30 (Tian et al.,2024) scale-wise autoregressive 2.0B 1.97 323.1 0.82 0.59
FlowAR-H scale-wise autoregressive 1.9B 1.65 296.5 0.83 0.60
Table 1.
Generation Results on ImageNet-256. Metrics reported include Fr
´
echet Inception Distance (FID), inception score (IS),
precision (Pre.) and recall (Rec.). The proposed FlowAR demonstrates state-of-the-art generation performance. VAR is evaluated using
code and pretrained weights from their official GitHub repository.
model, FlowAR-H (1.9B parameters, FID 1.65), sets a new
state-of-the-art benchmark for scale-wise autoregressive im-
age generation.
Quantitative Results on ImageNet-512. As demonstrated
in Table 2, we present quantitative results on ImageNet-512
generation. Our FlowAR-L model surpasses the closely
related VAR-d36 by 0.2 FID while utilizing only 25.6%
of its parameters. Compared to other state-of-the-art gen-
erative models—including GAN-based BigGAN (Brock
et al.,2018), mask-prediction-based MaskGiT (Chang et al.,
2022), token-wise autoregressive VQGAN (Esser et al.,
2021), and diffusion-based DiT-XL/2 (Peebles & Xie,2023),
our FlowAR-L achieves superior performance across both
FID and IS metrics, establishing consistent improvements
over all competing approaches.
Visualization. We visualize samples generated by FlowAR
using different tokenizers in Figure 4, showing that FlowAR
is capable of producing high-quality images with impressive
visual fidelity and is compatible with various off-the-shelf
VAEs. More samples at 256
×
256 and 512
×
512 resolution
are provided in the Appendix.
method params FID ↓IS ↑
BigGAN (Brock et al.,2018) 158M 8.43 177.9
MaskGiT (Chang et al.,2022) 227M 7.32 156.0
VQGAN (Esser et al.,2021) 227M 26.52 66.8
DiT-XL/2 (Peebles & Xie,2023) 675M 3.04 240.8
VAR-d36 (Tian et al.,2024) 2.3B 2.63 303.2
FlowAR-L 590M 2.43 295.1
Table 2.
Generation Results on ImageNet-512. Metrics reported
include Fr
´
echet Inception Distance (FID), inception score (IS).
The proposed FlowAR outperforms prior works.
4.2. Ablation Studies
We conduct ablation studies on tokenizer compatibility,
scale-wise flow matching model and injection of semantics.
More ablation studies on construction of scale sequence and
scale configurations can be found in the Appendix.
Tokenizer Compatibility. VAR (Tian et al.,2024) relies on
a complex multi-scale residual tokenizer that compresses
images into discrete tokens at different scales, with the scale
structure of the tokenizer directly mirroring VAR’s archi-
tectural scales. This tight coupling between the tokenizer
and VAR limits the framework’s flexibility and adaptabil-
7
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
MAR-VAE
SD-VAE
ostrich (9) macaw (88) wallaby (104) blenheim
spaniel (156)
lesser panda
(387)
lesser panda
(387)
barn (425)barn (425) stone wall (825)stone wall (825)
Figure 4.
Visualization of Samples Generated by FlowAR Using Different Tokenizers. FlowAR consistently produces high-quality
visual samples across various tokenizer configurations including VAE from MAR (Li et al.,2024) and SD (Rombach et al.,2022).
tokenizer Param. generator FID IS
multi-scale residual VQGAN 109.0M VAR 5.81 201.3
DC-AE (Chen et al.,2024) 323.4M FlowAR 4.22 220.9
SD-VAE (Rombach et al.,2022) 83.7M FlowAR 3.94 231.0
MAR-VAE (Li et al.,2024) 66.5M FlowAR 3.61 234.1
Table 3.
Ablation on Tokenizer Compatibility. While VAR (Tian
et al.,2024) depends heavily on a multi-scale residual discrete
tokenizer, FlowAR is compatible with a variety of continuous VAE
tokenizers. The final setting is marked in gray.
ity. In contrast, our proposed FlowAR is compatible with a
wide range of variational autoencoders (VAEs), enhancing
versatility and ease of integration. As shown in Table 3,
FlowAR achieves superior performance across various VAE
architectures, including DC-AE (Chen et al.,2024) (FID
of 4.22), SD-VAE (Rombach et al.,2022) (FID of 3.94),
and MAR-VAE (Li et al.,2024) (FID of 3.61), compared to
VAR’s multi-scale residual discrete tokenizer, which yields
an FID of 5.81. These results underscore FlowAR ’s effec-
tiveness and adaptability, highlighting its advantage over
VAR’s more rigid tokenizer dependency.
Scale-wise Flow Matching Model. The flow matching
model is used to learn the per-scale probability distribu-
tion, predicting all
hk×wk
tokens in the
k
-th scale
sk
.
We consider two design alternatives. First, per-scale pre-
diction could be replaced with per-token prediction using
Multi-Layer Perceptrons (MLPs) (Li et al.,2024), which,
however, lacks the ability to capture interactions between
tokens. Second, we could substitute the flow matching ap-
proach (Lipman et al.,2022) with a diffusion framework (Ho
et al.,2020). These design choices are explored in Table 4,
where per-scale prediction consistently outperforms per-
token prediction, regardless of whether flow matching or
diffusion is used. Additionally, flow matching provides
marginal improvements over the diffusion framework. Our
final model configuration employs per-scale prediction with
per-token per-scale diffusion flow-matching FID IS
✓ ✓ 6.85 155.6
✓ ✓ 6.15 184.1
✓ ✓ 3.93 223.9
✓ ✓ 3.61 234.1
Table 4.
Ablation on Scale-wise Flow Matching Model. We
investigate two design choices in this module: (1) per-token vs.per-
scale prediction and (2) diffusion vs.flow-matching framework.
The final setting is highlighted in gray.
the flow matching framework.
Injection of Semantics. There are several methods to inject
the semantics
ˆsi
, generated by the autoregressive Trans-
former, into the flow matching module. We summarize
these methods in Table 5: (1) ‘addition’: Element-wise ad-
dition of the semantics and the flattened input. (2) ‘cross
attention’: Using cross-attention where the flattened input
serves as the query and the semantics act as the key and
value. (3) ‘sequence concatenation’: Concatenating the
semantics with the flattened input along the sequence di-
mension. (4) ‘channel concatenation’: Concatenating the
semantics with the flattened input along the channel dimen-
sion. (5) ‘adaLN’: Adaptive LayerNorm conditioned on
the spatially averaged semantics. (6) ‘Spatial-adaLN’: The
proposed spatial adaptive LayerNorm, injecting semantics
position-by-position.
As shown in Table 5, the choice of semantic injection
method significantly impacts performance. The proposed
‘Spatial-adaLN’ achieves the best results, with an FID of
3.61 and an Inception Score (IS) of 234.1, outperforming all
other methods. These results indicate that methods preserv-
ing spatial structures and offering position-wise semantic
guidance yield superior image generation quality. The ex-
ceptional performance of Spatial-adaLN can be attributed to
its ability to inject semantics directly into the normalization
layers in a spatially adaptive manner.
8
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
injection of semantics FID IS
adaLN 14.28 111.9
sequence concatenation 9.22 146.9
addition 6.22 188.2
cross attention 5.37 202.5
channel concatenation 4.85 210.9
spatial-adaLN 3.61 234.1
Table 5.
Ablation on Semantic Injection Schemes for the Flow
Matching Module. The proposed Spatial-adaLN injects semantics
in a spatially adaptive manner, achieving the best performance. The
final setting is highlighted in gray.
5. Conclusion
In this work, we presented FlowAR, a flexible and gen-
eralized approach to scale-wise autoregressive modeling
for image generation, enhanced with flow matching for im-
proved quality. By adopting a streamlined scale design and
compatible with any VAE tokenizer, FlowAR addresses lim-
itations of prior models, offering greater adaptability and
superior image quality. With spatially adaptive layer normal-
ization, achieving state-of-the-art results on ImageNet-256
generation benchmark. We hope FlowAR will inspire more
future research in autoregressive image modeling.
Impact Statement
FlowAR represents a major advance in scale-wise autore-
gressive image generation: it eliminates VAR’s complex
multi-scale residual tokenizer, supports any off-the-shelf
Variational AutoEncoder, and adopts a streamlined scale de-
sign in which each successive resolution is simply twice that
of its predecessor. FlowAR not only elevates the quality of
synthetic imagery but also establishes a modular framework
that we anticipate will catalyze a new wave of innovation in
scalable, adaptable image generation.
Acknowledge
This work is supported by Office of Naval Research with
award N000142412696.
References
Albergo, M. S. and Vanden-Eijnden, E. Building normal-
izing flows with stochastic interpolants. arXiv preprint
arXiv:2209.15571, 2022.
Atito, S., Awais, M., and Kittler, J. Sit: Self-supervised
vision transformer. arXiv preprint arXiv:2104.03602,
2021.
Ba, J. L., Kiros, J. R., and Hinton, G. E. Layer normalization.
arXiv preprint arXiv:1607.06450, 2016.
Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan,
Y., Ge, W., Han, Y., Huang, F., et al. Qwen technical
report. arXiv preprint arXiv:2309.16609, 2023.
Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., and Zhu, J.
All are worth words: A vit backbone for diffusion models.
In CVPR, 2023.
Brock, A., Donahue, J., and Simonyan, K. Large scale gan
training for high fidelity natural image synthesis. arXiv
preprint arXiv:1809.11096, 2018.
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan,
J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G.,
Askell, A., et al. Language models are few-shot learners.
NeurIPS, 2020. URL
https://proceedings.
neurips.cc/paper/2020/file/
1457c0d6bfcb4967418bfb8ac142f64a-Paper.
pdf.
Chang, H., Zhang, H., Jiang, L., Liu, C., and Freeman,
W. T. Maskgit: Masked generative image transformer. In
CVPR, 2022.
Chen, J., Cai, H., Chen, J., Xie, E., Yang, S., Tang, H.,
Li, M., Lu, Y., and Han, S. Deep compression autoen-
coder for efficient high-resolution diffusion models. arXiv
preprint arXiv:2410.10733, 2024.
Chen, M., Radford, A., Child, R., Wu, J., Jun, H., Dhariwal,
P., Luan, D., and Sutskever, I. Generative pretraining
from pixels. In ICML, 2020.
Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra,
G., Roberts, A., Barham, P., Chung, H. W., Sutton, C.,
Gehrmann, S., et al. Palm: Scaling language modeling
with pathways. Journal of Machine Learning Research,
24(240):1–113, 2023.
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei,
L. Imagenet: A large-scale hierarchical image database.
In CVPR, 2009.
Deng, X., Yu, Q., Athar, A., Yang, C., Yang, L., Jin,
X., Shen, X., and Chen, L.-C. Coconut-pancap: Joint
panoptic segmentation and grounded captions for fine-
grained understanding and generation. arXiv preprint
arXiv:2502.02589, 2025.
Dhariwal, P. and Nichol, A. Diffusion models beat gans on
image synthesis. NeurIPS, 34, 2021.
Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle,
A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan,
A., et al. The llama 3 herd of models. arXiv preprint
arXiv:2407.21783, 2024.
Esser, P., Rombach, R., and Ommer, B. Taming transformers
for high-resolution image synthesis. In CVPR, 2021.
9
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
Esser, P., Kulal, S., Blattmann, A., Entezari, R., M
¨
uller, J.,
Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al.
Scaling rectified flow transformers for high-resolution
image synthesis. In ICML, 2024.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B.,
Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y.
Generative adversarial nets. NeurIPS, 2014.
He, J., Yu, Q., Liu, Q., and Chen, L.-C. Flowtok: Flowing
seamlessly across text and image tokens. arXiv preprint
arXiv:2503.10772, 2025.
Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and
Hochreiter, S. Gans trained by a two time-scale update
rule converge to a local nash equilibrium. NeurIPS, 30,
2017.
Ho, J., Jain, A., and Abbeel, P. Denoising diffusion proba-
bilistic models. NeurIPS, 2020.
Hoogeboom, E., Heek, J., and Salimans, T. simple diffusion:
End-to-end diffusion for high resolution images. In ICML,
2023.
Kang, M., Zhu, J.-Y., Zhang, R., Park, J., Shechtman, E.,
Paris, S., and Park, T. Scaling up gans for text-to-image
synthesis. In CVPR, 2023.
Kim, D., He, J., Yu, Q., Yang, C., Shen, X., Kwak, S., and
Chen, L.-C. Democratizing text-to-image masked gener-
ative models with compact text-aware one-dimensional
tokens. arXiv preprint arXiv:2501.07730, 2025.
Kingma, D. P. and Welling, M. Auto-encoding variational
bayes. In ICLR, 2014.
Lee, D., Kim, C., Kim, S., Cho, M., and Han, W.-S. Autore-
gressive image generation using residual quantization. In
CVPR, 2022.
Li, T., Tian, Y., Li, H., Deng, M., and He, K. Autoregressive
image generation without vector quantization. NeurIPS,
2024.
Lipman, Y., Chen, R. T., Ben-Hamu, H., Nickel, M., and
Le, M. Flow matching for generative modeling. arXiv
preprint arXiv:2210.02747, 2022.
Liu, Q., Zeng, Z., He, J., Yu, Q., Shen, X., and Chen, L.-
C. Alleviating distortion in image generation via multi-
resolution diffusion models. NeurIPS, 2024.
Liu, X., Gong, C., and Liu, Q. Flow straight and fast:
Learning to generate and transfer data with rectified flow.
arXiv preprint arXiv:2209.03003, 2022.
OpenAI. Introducing chatgpt.
https://openai.com/
blog/chatgpt/, 2022.
OpenAI. Gpt-4 technical report. arXiv preprint
arXiv:2303.08774, 2023. URL
https://arxiv.
org/abs/2303.08774.
Peebles, W. and Xie, S. Scalable diffusion models with
transformers. In ICCV, 2023.
Pope, R., Douglas, S., Chowdhery, A., Devlin, J., Bradbury,
J., Heek, J., Xiao, K., Agrawal, S., and Dean, J. Efficiently
scaling transformer inference. Proceedings of Machine
Learning and Systems, 5, 2023.
Powers, D. M. Evaluation: from precision, recall and f-
measure to roc, informedness, markedness and correla-
tion. arXiv preprint arXiv:2010.16061, 2020.
Radford, A., Narasimhan, K., Salimans, T.,
and Sutskever, I. Improving language under-
standing by generative pre-training.
https:
//cdn.openai.com/research-covers/
language-unsupervised/language_
understanding_paper.pdf, 2018.
Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G.,
Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark,
J., Krueger, G., and Sutskever, I. Learning transferable
visual models from natural language supervision. In
ICML, 2021.
Razavi, A., Van den Oord, A., and Vinyals, O. Generating
diverse high-fidelity images with vq-vae-2. NeurIPS,
2019.
Ren, S., Wang, Z., Zhu, H., Xiao, J., Yuille, A., and Xie, C.
Rejuvenating image-gpt as strong visual representation
learners. In ICML, 2024a.
Ren, S., Yu, Y., Ruiz, N., Wang, F., Yuille, A., and Xie,
C. M-var: Decoupled scale-wise autoregressive model-
ing for high-quality image generation. arXiv preprint
arXiv:2411.10433, 2024b.
Ren, S., Yu, Q., He, J., Shen, X., Yuille, A., and Chen, L.-C.
Beyond next-token: Next-x prediction for autoregres-
sive visual generation. arXiv preprint arXiv:2502.20388,
2025a.
Ren, S., Zhu, H., Wei, C., Li, Y., Yuille, A., and Xie, C.
Arvideo: Autoregressive pretraining for self-supervised
video representation learning. TMLR, 2025b.
Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and
Ommer, B. High-resolution image synthesis with latent
diffusion models. In CVPR, 2022.
Ronneberger, O., Fischer, P., and Brox, T. U-net: Convolu-
tional networks for biomedical image segmentation. In
MICCAI, 2015.
10
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V.,
Radford, A., and Chen, X. Improved techniques for
training gans. NeurIPS, 29, 2016.
Sauer, A., Schwarz, K., and Geiger, A. Stylegan-xl: Scal-
ing stylegan to large diverse datasets. arXiv preprint
arXiv:2201.00273, 2022.
Shin, I., Yang, C., and Chen, L.-C. Deeply super-
vised flow-based generative models. arXiv preprint
arXiv:2503.14494, 2025.
Song, J., Meng, C., and Ermon, S. Denoising diffusion
implicit models. arXiv preprint arXiv:2010.02502, 2020.
Sun, P., Jiang, Y., Chen, S., Zhang, S., Peng, B., Luo,
P., and Yuan, Z. Autoregressive model beats diffusion:
Llama for scalable image generation. arXiv preprint
arXiv:2406.06525, 2024.
Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.-B., Yu,
J., Soricut, R., Schalkwyk, J., Dai, A. M., Hauth, A., et al.
Gemini: a family of highly capable multimodal models.
arXiv preprint arXiv:2312.11805, 2023.
Tian, K., Jiang, Y., Yuan, Z., Peng, B., and Wang, L. Visual
autoregressive modeling: Scalable image generation via
next-scale prediction. NeurIPS, 2024.
Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux,
M.-A., Lacroix, T., Rozi
`
ere, B., Goyal, N., Hambro, E.,
Azhar, F., et al. Llama: Open and efficient foundation lan-
guage models. arXiv preprint arXiv:2302.13971, 2023a.
Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi,
A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P.,
Bhosale, S., et al. Llama 2: Open foundation and fine-
tuned chat models. arXiv preprint arXiv:2307.09288,
2023b.
Van den Oord, A., Kalchbrenner, N., Espeholt, L., Vinyals,
O., Graves, A., et al. Conditional image generation with
pixelcnn decoders. NeurIPS, 2016.
Van Den Oord, A., Vinyals, O., et al. Neural discrete repre-
sentation learning. NeurIPS, 2017.
Vaswani, A. Attention is all you need. NeurIPS, 2017.
Wang, F., Yang, T., Yu, Y., Ren, S., Wei, G., Wang, A., Shao,
W., Zhou, Y., Yuille, A., and Xie, C. Causal image mod-
eling for efficient visual understanding. arXiv preprint
arXiv:2410.07599, 2024.
Wang, F., Yu, Y., Wei, G., Shao, W., Zhou, Y., Yuille,
A., and Xie, C. Scaling laws in patchification: An im-
age is worth 50,176 tokens and more. arXiv preprint
arXiv:2502.03738, 2025.
Weber, M., Yu, L., Yu, Q., Deng, X., Shen, X., Cremers, D.,
and Chen, L.-C. Maskbit: Embedding-free image gen-
eration via bit tokens. arXiv preprint arXiv:2409.16211,
2024.
Yang, C., Liu, C., Deng, X., Kim, D., Mei, X., Shen,
X., and Chen, L.-C. 1.58-bit flux. arXiv preprint
arXiv:2412.18653, 2024.
Yu, J., Li, X., Koh, J. Y., Zhang, H., Pang, R., Qin, J., Ku,
A., Xu, Y., Baldridge, J., and Wu, Y. Vector-quantized
image modeling with improved vqgan. arXiv preprint
arXiv:2110.04627, 2021.
Yu, J., Xu, Y., Koh, J. Y., Luong, T., Baid, G., Wang, Z.,
Vasudevan, V., Ku, A., Yang, Y., Ayan, B. K., et al. Scal-
ing autoregressive models for content-rich text-to-image
generation. arXiv preprint arXiv:2206.10789, 2022.
Yu, Q., He, J., Deng, X., Shen, X., and Chen, L.-C. Ran-
domized autoregressive visual generation. arXiv preprint
arXiv:2411.00776, 2024a.
Yu, Q., Weber, M., Deng, X., Shen, X., Cremers, D., and
Chen, L.-C. An image is worth 32 tokens for reconstruc-
tion and generation. NeurIPS, 2024b.
Zhang, B. and Sennrich, R. Root mean square layer normal-
ization. NeurIPS, 2019.
11
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
Appendix
The appendix includes the following additional information:
• Sec. Alists the hyper-parameters of FlowAR.
•
Sec. Bprovides the architectural details of FlowAR
model variants.
• Sec. Cprovides more ablation studies.
• Sec. Dprovides more visualization results.
• Sec. Eincludes the impact statement.
A. Hyper-parameters
We list the hyper-parameters of our FlowAR in Table 6.
training hyper-parameters
optimizer AdamW
warmup epochs 100
total epochs 400
batch size 1024
peak learning rate 2e-4
minimal learning rate 1e-5
learning rate schedule cosine
class label dropout rate 0.1
max gradient norm 1.0
Table 6. Hyper-parameters of FlowAR.
B. Model Variants
In Table 7, we provide four kinds of different configurations
of FlowAR for a fair comparison under similar parameters
with VAR (Tian et al.,2024). The proposed FlowAR con-
tains two main modules: Autoregressive Model and Flow
Matching Model, both build on top of Transformer architec-
tures (Vaswani,2017).
variants autoregressive model flow matching model
FlowAR-S D=12, W=768 D=2, W= 1024
FlowAR-B D=16, W= 768 D=6, W=1024
FlowAR-L D=16, W= 1024 D=12, W= 1024
FlowAR-H D=30, W= 1536 D=18, W= 1536
Table 7.
Model Variants. “D” and “W” represent model depth
and width, respectively.
C. More Ablation Studies
Construction of Scale Sequence. Instead of using a multi-
scale VQGAN with residual connections (Tian et al.,2024),
scale construction FID IS
image 12.19 118.2
latent 3.61 234.1
Table 8.
Ablation on Construction of Scale Sequence. We com-
pare two approaches for constructing the scale sequence: down-
sampling the input image before feeding it into the VAE, or down-
sampling the latents extracted by the VAE. The final setting is
highlighted in gray.
method scales SFID IS
VAR {1,2,3,4,5,6,8,10,13,16}5.81 201.3
VARs {1,2,4,8,16}N/A N/A
FlowAR {1,2,4,8,16}3.61 234.1
FlowAR {1,4,8,16}4.88 200.1
FlowAR {1,4,16}6.10 194.2
Table 9.
Ablation on Scale Configurations. We compare VAR
and FlowAR under various scale configurations. VAR is con-
strained by its predefined scale sequence and fails to generalize
to other configurations (2nd row), whereas FlowAR demonstrates
flexibility across different scale setups. The final setting is high-
lighted in gray.
as in VAR (Tian et al.,2024), we propose a simpler ap-
proach by directly downsampling the latent representations
extracted by any off-the-shelf continuous VAE (Kingma
& Welling,2014) tokenizer. An alternative design choice
would be to downsample the image before feeding it into a
VAE. We explore this design choice in Table 8, where down-
sampling the latents (FID 3.61) significantly outperforms
downsampling the image (FID 12.19).
Scale Configurations. To demonstrate FlowAR’s flexibility
with respect to scale design, we perform an ablation study
by progressively reducing the number of scales used in the
model. Table 9presents the results of this study. VAR (Tian
et al.,2024) relies on a complex scale configuration with ten
scales (
{
1, 2, 3, 4, 5, 6, 8, 10, 13, 16
}
) to achieve its reported
performance. Simplifying VAR’s scale configuration to
{
1,
2, 4, 8, 16
}
leads to training failure, indicating a strong
dependency between its tokenizer and generator. In contrast,
FlowAR demonstrates strong robustness to scale reduction.
With the simplified sequence
{
1, 2, 4, 8, 16
}
, FlowAR
achieves an FID of 3.61, outperforming VAR even with its
full scale sequence. Reducing the scales further to
{
1, 4, 8,
16}, FlowAR still maintains competitive performance with
an FID of 4.88. Even with just three scales (
{
1, 4, 16
}
),
FlowAR achieves an FID of 6.10, comparable to VAR’s FID
of 5.81.
D. More Visualization Results
Additional visualization results generated by FlowAR-H are
provided from Figure 5to Figure 10.
12
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
Figure 5.
Generated samples from FlowAR. FlowAR generate
high-fidelity great grey owl (24) images.
Figure 6.
Generated samples from FlowAR. FlowAR generate
high-fidelity macaw (88) images.
Figure 7.
Generated samples from FlowAR. FlowAR generate
high-fidelity golden retriever (207) images.
Figure 8.
Generated samples from FlowAR. FlowAR generate
high-fidelity otter (360) images.
13
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
Figure 9.
Generated samples from FlowAR. FlowAR generate
high-fidelity artichoke (944) images.
Figure 10.
Generated samples from FlowAR. FlowAR generate
high-fidelity alp (970) images.
E. Impact Statement
Our main objective is to advance foundational research
on generative models. A direct next step is to extend our
method to large-scale visual generation tasks, such as text-to-
image or text-to-video models, where it can notably reduce
both training and inference costs while enhancing perfor-
mance. Nonetheless, since our approach derives statistics
directly from ImageNet, it may also inherit any biases inher-
ent in that data.
14
+