FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching

We thank the reviewer for the constructive feedback and address the concerns below.

Essential References

Thank you for the suggestions. FlowAR focuses on next-scale prediction through autoregressive modeling, whereas the referenced works explore diffusion-based approaches with multi-stage architectures [1], multi-scale designs [2], or independent noise levels combined with causal next-token prediction [3]. We will cite and discuss all related works accordingly.

W1: scale design analysis

We have ablated the scale design in Table 9 of the Appendix (page 11 in paper). For quick reference, we provide the results below. As shown, even with just three scales ({1, 4, 16}), FlowAR achieves an FID of 6.10—comparable to VAR, which uses a much more complex scale configuration ({1, 2, 3, 4, 5, 6, 8, 10, 13, 16}) and obtains an FID of 5.81.

W2: speed comparison with VAR

In the table below, we compare the generation speed and performance of FlowAR with VAR and other widely used diffusion- and flow-matching-based models, including MAR, DiT, and SiT.

As shown, although FlowAR involves additional denoising steps due to flow matching and is slightly slower than VAR, FlowAR-L remains over 10× faster than other diffusion- and flow-matching-based models such as MAR, DiT, and SiT while also achieving superior image generation quality.

W3: Precision and Recall metrics in Table 1

Precision and Recall metrics often saturate and offer limited additional insight beyond FID. Therefore, following prior works such as MAR, VAR, DiT, and SiT, we primarily focus on optimizing and reporting FID, which remains the most sensitive and informative metric for assessing image generation quality in our evaluations.

W4: Update Table 3 with tokenizer’s parameters

Thank you for the suggestion. In the updated table below, we include the tokenizer parameters to provide a more complete and transparent comparison.

W5: Change “Any VAE” to “Any continuous tokenizer (VAE)”

Thank you for the suggestion. We will update the text to “any continuous tokenizer (VAE)” accordingly.

We thank the reviewer for the constructive feedback and address the concerns below.

Q1: speed comparison with VAR

While FlowAR requires additional denoising steps due to the use of flow matching, it is only slightly slower than VAR, and FlowAR-L still achieves a significant 10× speedup over other diffusion- and flow-matching-based models such as MAR, DiT, and SiT.

Q2: more detailed description about the AR and flow-matching Transformer

Our autoregressive and flow matching modules are built upon the VAR and SiT codebases, respectively. As stated in the paper, we will fully open-source the training and inference code, along with model checkpoints, to enable the community to reproduce our results and examine all implementation details.

Q3: balance sizes of AR and flow-matching

Thank you for the question. We ablate the effect of AR and flow matching model sizes in the table below. Increasing the size of the flow matching module initially improves performance, but further scaling eventually leads to a performance drop.

Q4: visualization of samples

While our primary focus is on the image quality at the final scale, we provide intermediate-scale visualizations at anonymous link and will add more intermediate-scale visualizations in the final version.

Q5: VAE from MAR

To analyze the impact of different underlying VAEs, we report their reconstruction FID (rFID) in the table below:

As shown, MAR-VAE achieves a better rFID, indicating better reconstruction quality, which correlates with improved generation FID (gFID) in both our experiments and those reported in the MAR paper.

To further validate the effect of MAR-VAE versus SD-VAE, we also compare FlowAR with another generator, UViT, using both tokenizers. MAR-VAE consistently delivers better generation performance across different generator architectures.

We thank the reviewer for the constructive feedback and address the concerns below.

Essential References

We thank the reviewer for the suggestion. These works represent early efforts in applying autoregressive modeling to image generation in pixel space. We will cite and discuss them in the related work section.

Q1: scale with resolution

In our design, each subsequent scale is simply double the previous one—providing a simple and intuitive hierarchical structure. Additionally, our method supports non-square resolutions, though this may result in a slight drop in performance.

Q2: time steps across scales

Thank you for the question. By default, we use the same number of time steps (25 steps) for all scales. We ablate the effect of using different denoising time steps across scales (1x1, 2x2, 4x4, 8x8, 16x16) in the following table.

As shown, increasing the time steps from 25 to 50 across all scales results in only a marginal FID improvement of 0.01, while doubling the inference time. Reducing the steps to 15 or 10 significantly speeds up inference but comes at the cost of degraded image quality. We also evaluate varying the number of steps across scales. Using fewer steps for smaller scales leads to a slight FID drop of 0.06 without affecting inference speed. Conversely, assigning fewer steps to larger scales also slightly degrades performance but yields faster inference. Overall, these results suggest that a uniform 25-step setting offers a better trade-off between generation quality and efficiency.

Q3: long-text conditioning

Since FlowAR is an autoregressive model, we can either concatenate the long text condition in front of the image sequence—similar to other autoregressive models [A]—or incorporate cross-attention as done in diffusion models [B].

[A] Autoregressive Model Beats Diffusion: Llama for Scalable Image Generation

[B] High-Resolution Image Synthesis with Latent Diffusion Models

Q4: positional encodings

The positional embeddings are handled in the same way as prior methods [C, D, E, F] (e.g., absolute position embeddings are added at the input layer). As stated in the paper, the detailed implementation (for both training and inference) will be fully open-source, allowing the community to reproduce and examine all details.

[C] Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction

[D] Autoregressive Image Generation without Vector Quantization

[E] Alleviating Distortion in Image Generation via Multi-Resolution Diffusion Models and Time-Dependent Layer Normalization

[F] Scalable Diffusion Models with Transformers

Q5: level of controllability

We thank the reviewer for the insightful question. In this work, our goal is to develop a general and broadly applicable framework by treating all scales uniformly—for example, using the same number of denoising steps across scales—on the standard ImageNet benchmark, following the next-scale prediction framework introduced by VAR. While datasets such as LAION or COCO may benefit from dataset-specific scale designs, our focus is on establishing a simple and general scale design that can generalize across settings without requiring dataset-specific tuning.

We thank the reviewer for the constructive feedback.

C1: importance of AR

Unlike Matryoshka Diffusion Models (MDM), which use a nestedUNet to jointly denoise all scales, FlowAR conditions on previously generated scales. Below, we include a baseline, “diffuse all noisy scales” (without AR), which faces two issues: (1) slower inference, as denoising is applied to all scales, resulting in extremely long sequences; and (2) worse FID, due to the absence of guidance from previously generated scales.

C2: single-scale flow model

Below, we compare with SiT-XL, a pure single-scale flow model, and “FlowAR-L w/ single-scale.” Our multi-scale design outperforms SiT-XL while requiring 14% of the Flops and achieving a 13.9× speedup. Slightly slower than the single-scale variant, the multi-scale design achieves a 1.78 FID improvement.

FLOPs are computed across the entire generation process.

C3: token-wise prediction of semantics

This token-wise design results in slower inference and worse performance:

C4: no split between AR and flow models

Below, we experimented with directly predicting the flow between the upsampled previous scale and the next scale.

C5: performance vs. runtime

Unlike MAR, which models each token individually, FlowAR models the probability over entire scales. It also reduces inference time by using only 5 AR steps (with 25 denoising steps each), compared to MAR’s 256 AR steps (with 100 steps each). As shown below, FlowAR-H is 5× faster than MAR-B while achieving a 0.66 FID improvement.

Due to response length limitation, please also refer to our response to Reviewer 6Cuw’s Q3 for the ablation on AR and flow model sizes.

ImageNet train-set gFID

We thank the reviewer for raising this point. FID is a widely used metric in prior works (e.g., MAR, DiT, VAR, SiT, StyleGAN), and our setup follows these standards for fair comparison.

Training settings of VAR and FlowAR

For data augmentation, we follow VAR’s setup with RandomCrop and horizontal flipping. As noted in the paper, we will fully open-source our code and checkpoints for reproducibility and transparency. Below, we clarify the training settings of VAR and FlowAR. While diffusion and flow models typically require long training (e.g., MAR: 800 epochs, DiT/SiT: 1400), FlowAR achieves strong performance with just 400 epochs. Compared to VAR-d30/d24 (350 epochs), FlowAR-H trains for a similar duration but uses nearly half the GPU hours, thanks to its simpler scale design and shorter sequences. The table below summarizes A100 training hours and FID scores. Compared to VAR-d20/12, our 250-epoch version of FlowAR achieves better performance while incurring only half the training cost.

Essential References

We will cite and discuss all mentioned works in the related work section and compare with [3, 8, 9] in the experiments. Our work is most closely related to VAR and its follow-up [6], which autoregressively predict the next scale—unlike other methods that, while leveraging multi-scale information, do not adopt this formulation.

IS vs. FID

We observe an IS–FID trade-off when tuning the CFG scale. While we focus on optimizing FID, a slight adjustment still allows both FID and IS to outperform VAR by a notable margin.

Visualization

Due to response length limitation, please also refer to our response to Reviewer 6Cuw’s Q4 for the visualization of intermediate stage generated samples.

L288 claim

Our Spatial-adaLN injects scale- and position-wise semantics for fine-grained conditioning, unlike the simpler concat approach (L281), which provides only sequence-level, indirect guidance.