NeuralRemaster: Phase-Preserving Diffusion for Structure-Aligned Generation
Yu Zeng1Charles Ochoa1Mingyuan Zhou2Vishal M. Patel3Vitor Guizilini1Rowan McAllister1
1Toyota Research Institute 2University of Texas, Austin 3Johns Hopkins University
Input eFLUX-Kontext QWen-Edit Ours
Input Cosmos-Transfer 2.5 Ours Ours
Figure 1. We present Phase-Preserving Diffusion (ϕ-PD), a model-agnostic reformulation of the diffusion process that preserves an image’s
phase while randomizing its magnitude, enabling structure-aligned generation with no architectural changes or additional parameters.
Abstract
Standard diffusion corrupts data using Gaussian noise
whose Fourier coefficients have random magnitudes and
random phases. While effective for unconditional or text-to-
image generation, corrupting phase components destroys
spatial structure, making it ill-suited for tasks requiring ge-
ometric consistency, such as re-rendering, simulation en-
hancement, and image-to-image translation. We introduce
Phase-Preserving Diffusion (ϕ-PD), a model-agnostic re-
formulation of the diffusion process that preserves input
phase while randomizing magnitude, enabling structure-
aligned generation without architectural changes or addi-
tional parameters. We further propose Frequency-Selective
Structured (FSS) noise, which provides continuous control
over structural rigidity via a single frequency-cutoff param-
eter. ϕ-PD adds no inference-time cost and is compatible
with any diffusion model for images or videos. Across pho-
torealistic and stylized re-rendering, as well as sim-to-real
enhancement for driving planners, ϕ-PD produces control-
lable, spatially aligned results. When applied to the CARLA
simulator, ϕ-PD improves CARLA-to-Waymo planner per-
formance by 50%. The method is complementary to existing
conditioning approaches and broadly applicable to image-
to-image and video-to-video generation. Videos, additional
examples, and code are available on our project page.
1. Introduction
Recent advances in diffusion models have revolution-
ized image generation, achieving high-fidelity results for
unconditional or text-conditioned synthesis. Yet many prac-
tical applications do not require generating a scene from
scratch. Instead, they operate within an image-to-image set-
ting where the spatial layout, such as object boundaries, ge-
ometry and scene structures, should remain fixed while the
appearance is modified. Examples include neural rendering,
stylization, and sim-to-real transfer for autonomous driving
or robotics simulation. We refer to this broad class of prob-
lems as structure-aligned generation.
Although these tasks are conceptually easier than gener-
ating an image from scratch, existing solutions are unnec-
essarily complex. Methods such as ControlNet [42], T2I-
Adapter [21], and related variants attach auxiliary encoders,
or adapter branches to inject structural input into the model.
While effective, this introduces additional parameters and
computational cost, paradoxically making structure-aligned
generation harder than it should be.
We argue that this inefficiency stems not from the net-
work architecture, but from the diffusion process itself. The
forward diffusion process injects Gaussian noise, which de-
stroys both the magnitude and phase components in the fre-
quency domain. Classical signal processing [23,30,37],
however, tells us that phase encodes structure while magni-
tude encodes texture. Destroying the phase means destroy-
ing the very spatial coherence that structure-aligned gener-
ation depends on, forcing the model to reconstruct structure
from scratch.
Motivated by this insight, we propose Phase-Preserving
Diffusion (ϕ-PD). Instead of corrupting data with Gaussian
noise, ϕ-PD constructs structured noise whose magnitude
matches that of Gaussian noise while preserving the input
phase. This naturally maintains spatial alignment through-
out sampling (Figure 1) with no architectural modifica-
tion, no extra parameters (Figure 2), and is compatible
with any DDPM or flow-matching model for images or
videos.
To provide controllable levels of structural rigidity, we
further introduce Frequency-Selective Structured (FSS)
noise, which interpolates between input phase and pure
Gaussian noise via a single cutoff parameter (Figure 4).
This allows us to control the trade-off between strict align-
ment and creative flexibility.
We evaluate ϕ-PD across photorealistic re-rendering,
stylized re-rendering and simulation enhancement for
embodied-AI agents. ϕ-PD consistently maintains geom-
etry alignment while producing high-quality visual outputs,
outperforming prior methods across both quantitative and
qualitative metrics. When used to enhance CARLA sim-
ulations, ϕ-PD improves planner transfer to the Waymo
Open Dataset by 49%, substantially narrowing the sim-to-
real gap. In summary, our contributions include:
•Phase-preserving diffusion process: A diffusion pro-
cess that preserves phase while randomizing magni-
tude in frequency domain, maintaining spatial struc-
ture without architectural changes.
•Frequency-selective structured noise: A single-
parameter mechanism that enables continuous control
over structural alignment rigidity.
•Unified and efficient framework applicable to both
images and videos, compatible with DDPMs and flow-
matching, and requires no inference-time overhead.
2. Related Work
Diffusion Models. Diffusion models have become
a dominant paradigm for generative modeling, capable of
representing complex data distributions with remarkable fi-
delity [11,16]. They progressively corrupt data into Gaus-
sian noise through a forward diffusion process, then learn
to invert this process via iterative denoising. This frame-
work has demonstrated state-of-the-art performance across
diverse domains, including image, video, and audio genera-
tion [3,4,12,17,25,28,31], as well as reinforcement learn-
ing [14,26,36] and robotics [1,5,35].
Frequency-Domain Manipulation for Diffusion. Re-
cent work has explored frequency domain operations for
diffusion models. [7] argues that diffusion models of im-
ages perform approximate autoregression in the frequency
domain. [40] shows that modifying the UNet frequency do-
main features significantly improves the generating qual-
ity for image or video generation. FreeDiff [38] intro-
duces a fine-tuning free approach for image editing that em-
ploys progressive frequency truncation to refine the guid-
ance of diffusion models. [27] proposed to use a frequency-
dependent moving average during sampling. [2] proposes a
training-free approach for image inpainting that optimizes
the initial seed noise in the spectral domain. [9] proposes to
generate optical illusion images using phase interpolation
based on DDIM inversion. While these approaches modify
the diffusion sampling process to achieve the desired behav-
ior, ϕ-PD introduces minimal changes to existing diffusion
frameworks and preserves the original sampling dynamics.
Structure-Aligned Generation with Diffusion. Most
existing methods achieve structure-aligned generation by
modifying the network architecture and introducing addi-
tional adaptation components. ControlNet [42] copies the
entire U-Net encoder into a trainable encoder branch, which
adds significant computation overhead. T2I-Adapter [21]
reduces computation overhead using a lightweight adapter
module but sacrifices control precision. Uni-ControlNet
[43] enables simultaneous utilization of multiple local con-
trols by training two adapters. OmniControl [32] integrates
image conditions into Diffusion Transformer (DiT) archi-
tectures with only 0.1% additional parameters by re-using
Phase Preserving (PP)
Noise
Gaussian Noise Frame-wise PP
noise
Prior methods encode structural input with ad-
ditional modules (yellow), that depend on the
model (green) and incur additional computation.
Our method incurs no additional over-
head and works with any model (green).
Our method extends to video.
Figure 2. Unlike prior approaches that modify architectures and add overhead, ϕ-PD preserves structure via phase consistency, remaining
lightweight and model-agnostic, reflecting that image-conditioned generation should be simpler, not harder.
the VAE and transformer blocks of the base model. Con-
trolNeXt [24] uses a lightweight convolutional module to
inject control signals, and directly finetune selective pa-
rameters of the base model to reduce training costs and la-
tency increase. SCEdit [15] proposes an efficient finetuning
framework that edits skip connections using a lightweight
module. NanoControl [13] aims to achieve efficient con-
trol with a LoRA-style control module. The above methods
all rely on an additional module to incorporate the control
signal, though some are more lightweight than others. Cos-
mosTransfer [22] achieves multi-modal control by combin-
ing multiple ControlNet branches, demonstrating promising
applications on physical tasks; however, multiple branches
introduce significant computation overhead. In contrast, ϕ-
PD does not introduce any computation overhead or addi-
tional parameters while enabling universal spatial control.
Training-Free Guidance Methods. Recently, sev-
eral training-free methods have been developed. [41] intro-
duced FreeDoM, which leverages off-the-shelf pre-trained
networks to construct time-independent energy functions
that guide generation. [6] proposed ZestGuide for zero-
shot spatial layout conditioning, utilizing implicit segmen-
tation maps extracted from cross-attention layers to align
generation with input masks. [20] presented FreeControl, a
training-free approach that enforces structure guidance with
the base model feature extracted from the control signal.
Although these methods avoid training cost, they introduce
additional overhead at test time, either an external model,
DDIM inversion, or multiple inferences of the base model.
In contrast, ϕ-PD can achieve training-free spatial control
without any additional inference time overhead.
3. Method
3.1. Frequency Domain Fundamentals
In the frequency domain, any image I(x, y)can be rep-
resented through the 2D Fourier transform:
F(u, v) = F{I(x, y)}
=
W−1
X
x=0
H−1
X
y=0
I(x, y)e−2πj(ux/W +vy/H),(1)
where F(u, v)is a complex-valued function that can be
decomposed into magnitude and phase components:
F(u, v) = |F(u, v)| · ejϕ(u,v)=A(u, v)·ejϕ(u,v).(2)
Here, A(u, v) = |F(u, v)|is the magnitude spectrum and
ϕ(u, v)the phase spectrum. The inverse Fourier transform
uses magnitude and phase to reconstruct the original image:
I(x, y) = F−1{F(u, v)}
=
W−1
X
u=0
H−1
X
v=0
F(u, v)e2πj(ux/W +vy/H).(3)
Phase-Magnitude Separation in Signal Processing.
Foundational work by Oppenheim et al. [23] shows that
phase primarily determines spatial structure, while magni-
tude largely controls texture statistics. Mixing phase and
magnitude from different images produces reconstructions
whose spatial layout follows the source of the phase, not
magnitude (see Figure 3). This observation motivates our
approach: if diffusion destroys phase, it destroys spatial ge-
ometry; if we preserve phase, we preserve structure.
Images Phase Magnitude
Car Phase
Dog Magnitude
Figure 3. Mixing phase and magnitude from two images. The
mixture keeps the structure of the image where the phase is taken.
3.2. Phase-Preserving Diffusion
Standard diffusion corrupts data using Gaussian noise
whose Fourier coefficients have random magnitudes and
random phases. As a result, even early diffusion steps erase
spatial alignment. We propose a simple alternative: pre-
serve the input image’s phase and randomize the magni-
tude, by using structured noise that shares the input phase.
Structured Noise Construction. Given an input image
I, we compute its Fourier transform:
FI=AI·ejϕI.(4)
We construct phase-preserving noise by pairing the input
image phase with a random magnitude:
Fˆϵ=Aϵ·ejϕI,(5)
and invert it:
ˆϵ=F−1{Fˆϵ},(6)
where the random magnitude Aϵcan be either from the
Fourier transforms of Gaussian noise:
Aϵ=|F{ϵ}|, ϵ ∼ N(0,1),(7)
or sampled from Rayleigh distribution directly [10]:
Aϵ=√−2 ln U, U ∼Uniform(0,1).(8)
This structured noise is used in place of Gaussian noise
in forward diffusion for training. It injects randomness
while maintaining the phase of the input. At test time, we
achieve structure-aligned generation by starting sampling
from structured noise constructed with input image phase.
Frequency Selective Structured (FSS) Noise. In prac-
tice, we often want to control to what extent we keep
Figure 4. Frequency Selective Structured (FSS) Noise with in-
creasing cutoff radius r.
the structure from the input image. Some tasks require
strict structure preservation, while others benefit from par-
tial freedom to reinterpret the scene. To provide this control,
we introduce Frequency Selective Structured (FSS) noise,
which only keep the image within a radius rand use the
phase from the noise for the remainder. We define a smooth
frequency mask M(u, v)based on the cutoff radius r:
M(u, v) = (1if √u2+v2≤r
exp −(√u2+v2−r)2
2σ2if √u2+v2> r
(9)
where σcontrols the smoothness of the transition.
The FSS noise ˆϵis the combination of image phase and
noise phase using the mask:
Fˆϵ=Aϵ·ejϕI⊙M+jϕϵ⊙(1−M),(10)
where ⊙represents element-wise multiplication. We can
sample phase ϕϵas the Fourier transform of Gaussian noise:
ϵ∼ N(0,1),
ϕϵ=arg(F{ϵ})∼Uniform(−π, π).(11)
When the mask is all zero, we take the phase from Gaus-
sian noise for all frequencies, then this FFS noise becomes
Gaussian noise and ϕ-PD becomes standard diffusion. Fig-
ure 4visualizes FSS noise with different cutoff radius r. We
can see that the noise becomes increasingly more structured
with increasing r. Figure 5shows images generated from
the same input with different cutoff radius where the gener-
ated image aligns more tightly to the input with larger r.
3.3. Training Objective
ϕ-PD does not depend on model architecture or diffusion
formulation. In the experiment section, we demonstrate in-
tegration both DDPM and Flow Matching, without modify-
ing their architectures of loss functions.
The flow matching objective learns a vector field that
transports the structured noise distribution to the target im-
age distribution. During training, given a target image Iand
a structured noise ˆϵ, and a timestep t∈[0,1], an intermedi-
ate image xtis obtained using a linear combination between
Iand ˆϵfollowing Rectified Flows [19]:
xt=tˆϵ+ (1 −t)I. (12)
Input r= 1 r= 6 r= 10 r= 20 r= 30
Figure 5. Image generated with the same noise and different cutoff radius r. Results are based on SD1.5.
The ground-truth velocity is
vt=dxt
dt = ˆϵ−I. (13)
With this ground-truth, we can then train the model by min-
imizing the mean squared error between the model output
and the ground-truth:
L=EI,ˆϵ,t∥u(xt, t;θ)−vt∥2
2.(14)
In the Fourier domain, the velocity becomes
Fvt= (Aˆϵ−AI)ejϕI(15)
which has the same phase ϕIas the image; therefore, the
trained model always generates images with the same phase
as the input image, remaining structurally aligned by con-
struction.
In DDPMs [11], data x0is gradually corrupted by Gaus-
sian noise:
q(xt|xt−1) = N(p1−βtxt−1, βtI).(16)
The model learns the reverse process by predicting the
added noise ϵat each step using the loss
LDDPM =Ex0,ϵ,t∥ϵ−ϵθ(xt, t)∥2
2,(17)
where xt=√¯αtx0+√1−¯αtϵand ¯αt=Qt
s=1(1 −βs).
In our formulation, we replace the Gaussian noise ϵwith
structured noise ˆϵthat preserves the input phase:
xt=√¯αtx0+√1−¯αtˆϵ. (18)
The training objective remains identical,
Lϕ-PD =Ex0,ˆϵ,t∥ˆϵ−ϵθ(xt, t)∥2
2.(19)
The phase-preserving noise biases denoising toward
structure-consistent trajectories, maintaining spatial align-
ment without altering the DDPM training objective.
3.4. Extension to Videos
ϕ-PD can be directly repurposed for video generation by
constructing phase-preserving noise frame-by-frame. For
a video {I1, I2, ..., IT}, we construct structured noise for
each frame, and concatenate them along the time dimen-
sion. We found that the best video generation results are
when we first apply an image-based ϕ-PD on the first frame,
and then use the first-frame-conditioned video ϕ-PD to gen-
erate the rest of the video. Similar to image generation,
for video generation, our method requires no architectural
changes either and simply supplies structured noise for each
frame.
4. Experiments
We evaluate ϕ-PD across three settings: photoreal-
istic re-rendering, stylized re-rendering, and simulation
enhancement for autonomous driving, comparing against
state-of-the-art methods. Our goals are to (1) assess spa-
tial alignment under appearance changes, (2) measure vi-
sual realism, and (3) quantify the impact on downstream
embodied-AI tasks.
To demonstrate its broad applicability, we implement
ϕ-PD on three representative diffusion models: SD 1.5,
FLUX-dev, and WAN 2.2 14B, which vary in size, formu-
lation, and modality, covering both image and video gener-
ation.
4.1. Implementation Details
4.1.1 Datasets
UnrealCV1is a open-source tool that includes multiple as-
sets. We created a diverse test set consisting of 5,000 im-
ages across all available assets, for a total of around 200
scenes. Figure 6shows examples from this test set. This
dataset covers a diverse range of scenes, including outdoor
and indoor, city and natural etc, with geometry diversity
while lacking photorealism. This dataset evaluates photo-
realistic enhancement and structure preservation.
ImageNetR is a test set proposed by [34], including 29 im-
ages of various objects and styles. While the original dataset
1https://github.com/unrealcv/unrealcv
Input QWen-Edit FLUX.1 Kontext Ours
Figure 6. Results on UnrealCV compared to FLUX-Kontext and QWenEdit.
provides prompts, these are generic image editing prompts.
Since our work primarily focuses on re-rendering, we keep
the editing prompts with style hints in the original dataset
and added additional style prompts, resulting in a total of 8
prompts for each image. This benchmark assesses stylized
re-rendering and structure preservation.
CARLA is an open-source driving simulator [8]. We collect
5.5 hours of driving videos from CARLA Town 4 using the
simulator’s default autopilot. We then split the videos into
25 second clips and annotate a caption for each clip. For
simulation enhancement, we use these captions combined
with the style hint “A photorealistic video of driving”. We
evaluate the effectiveness of sim-to-real transfer by testing
the CARLA-trained planner on Waymo’s WOD-E2E [39]
validation set.
4.1.2 Model architecture
We integrate ϕ-PD into:
• SD 1.5 (image DDPM)
• FLUX-dev (image flow matching)
• Wan2.2-14B (video flow matching)
We either fully finetune or LoRA-finetune each model using
phase-preserving noise; no architectural changes are intro-
duced. Notably, this finetuning is highly efficient: adapting
the Wan2.2-14B video model with LoRA required only a
single GPU while still yielding high-quality results, further
demonstrating the lightweight nature of ϕ-PD. Please refer
to the Appendix for additional implementation details and
ablation studies.
4.1.3 Evaluation Metrics
For photorealistic re-rendering, we use our UnrealCV
dataset and apply different methods to make images photo-
realistic. We use the style prompt ”A high-quality picture”
combined with the caption extracted from the images using
ChatGPT.
Visual quality. We define an appearance score (AS) that
measures how successful re-rendering is by looking at
the ratio of CLIP similarity between positive and negative
prompts:
AS =x⊤tp
x⊤tn
(20)
where xindicates the CLIP embedding of the re-rendered
image; tnthe CLIP embedding of the negative prompt; and
tpthe CLIP embedding of the positive prompt. For the
positive prompt we use “Photo, camera captured, picture,
photorealistic”, and for the negative prompt we use “Game,
render, cartoon, unreal”.
Structural alignment. Besides visual quality, successful
re-rendering also requires preserving the structure from the
original image. To evaluate the structural alignment, we
compute the error between the depth map of the original
image and the generated image, using metrics SSIM (Struc-
Table 1. Quantitative evaluation results for photorealistic re-
rendering on UnreadlCV.
Model Input images ControlNet-Tile SDEdit Ours
AS 0.9485 0.9733 0.9782 1.0008
SSIM - 0.8781 0.8883 0.8982
ABSREL - 0.5936 0.4938 0.4569
Table 2. Quantitative evaluation for stylized re-rendering.
Model ControlNet-Tile SDEdit PNP Ours
AS 1.3167 1.4243 1.4726 1.4709
SSIM 0.8831 0.7638 0.8498 0.8502
ABSREL 0.6684 1.0336 0.8194 0.7949
tural Similarity Index Measure) and ABSREL (Absolute
Relative error).
For stylized re-rendering, we evaluate in a similar way
as for photorealistic re-rendering. We compute the appear-
ance score with the target style prompt as positive and the
original style hint of the original image as negative. We
also use the error between depth maps to evaluate structural
alignment.
For simulation enhancement, we train an end2end
planner for each compared method on the corresponding
re-rendered image. We then test the planner’s trajectories
on Waymo’s WOD-E2E validation set using distance-based
metrics from the demonstrated driving, including Average
Displacement Error (ADE) and Final Displacement Error
(FDE).
4.2. Results
ADE FDE
0
5
10
15
20
25
30
Error
8.2
17.1
11.2
28.8
4.1
9.1
4.2
10.0
CARLA
Cosmos Transfer 2.51
Ours1
Ours2
Figure 7. Planner error on Waymo validation set. Lower is better.
1zero-shot, 2finetuned on Waymo training set videos.
Photorealistic Re-Rendering. Quantitative results on
UnrealCV re-rendering are summarized in Table 1, where
all methods are implemented using the SD 1.5-based mod-
els for fair quantitative comparison. Qualitative exam-
ples are shown in Figure 6, which compares our Flux-
based model against stronger recent models such as FLUX-
Kontext, and Qwen-Edit. Across both settings, all methods
improve photorealism-as reflected by higher AS scores than
the input images while ϕ-PD achieves the highest photo-
realism and superior structure alignment. We observe that
QWen-Edit produces visually high-quality results but often
fails to maintain structural alignment with the input image;
for example, in the first four cases, it enlarges the main sub-
jects significantly. FLUX-Kontext aligns better with the in-
put structure but provides only limited improvement in vi-
sual quality. Our method achieves both high visual fidelity
and consistent structural alignment across frames.
Stylized Re-rendering. Results of stylized re-rendering
are presented in Table 2, with representative examples
shown in Figure 8. All models are based on SD 1.5.
This task evaluates the model’s ability to alter appearance
while preserving scene structure. As shown, ϕ-PD produces
visually coherent stylizations that maintain object bound-
aries and spatial consistency, while prior methods often dis-
tort geometry or introduce texture misalignment. Quan-
titatively, ϕ-PD achieves a better trade-off between style
strength and structural fidelity.
Simulation Enhancement. For this experiment, we
generate 5.5 hours of demonstration driving videos from
CARLA using its autopilot. Then we train an end2end plan-
ner on the re-rendered CARLA videos from each method
using a ResNet backbone with a GRU to take temporal input
and an MLP head to output a trajectory ∈ R16×2(4s predic-
tions at 4Hz in XY space). As baseline, we also present the
results from a purely CARLA-trained model. Open-loop
imitation driving results are given in Figure 7.ϕ-PD boosts
planner generalization by 50% in zero-shot setting, demon-
strating that structure-preserving appearance enhancement
significantly reduces the sim-to-real gap. Video examples
in Figure 9show that ϕ-PD maintains road boundaries, ve-
hicle shapes, and spatial layout consistently across frames,
whereas the compared method produces distorted trees and
multi-object artifacts.
5. Conclusion
We introduced Phase-Preserving Diffusion (ϕ-PD), a
simple yet effective reformulation of the diffusion process
that replaces Gaussian noise with structured noise that pre-
serves image phase while randomizing the magnitude in the
frequency domain. This simple change retains spatial align-
ment throughout sampling without modifying the architec-
ture, altering training objectives, or introducing inference-
time overhead. We also introduced Frequency-Selective
“Pencil Sketch of a Castle. ”
“Picture of a Husky”
Input PnP SDEdit ControlNet Ours
Figure 8. Results on stylized rerendering.
Figure 9. Video re-rendering results for simulation enhancement. From top to bottom are: input, Cosmos-Transfer2.5 [22], ours.
Structured (FSS) noise, which provides continuous con-
trol over structural alignment rigidity through a single fre-
quency cutoff parameter, making it broadly applicable to
different applications. Across photorealistic re-rendering,
stylized re-rendering, and simulation enhancement for driv-
ing, ϕ-PD demonstrates strong spatial fidelity and visual re-
alism. When applied to CARLA re-rendering, ϕ-PD signif-
icantly improves zero-shot planner transfer to the Waymo
dataset, narrowing the sim-to-real gap.
Limitation. ϕ-PD assumes image-like inputs; modalities
such as depth or normals may require a lightweight prior to
produce an initial image representation.
Future work. ϕ-PD is orthogonal to existing condition-
ing or adapter methods and can be integrated with them for
enhanced control. Future work includes extending ϕ-PD to
tasks such as deblurring, relighting, super-resolution, and
general image restoration.
References
[1] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua Tenenbaum,
Tommi Jaakkola, and Pulkit Agrawal. Is conditional gen-
erative modeling all you need for decision-making? arXiv
preprint arXiv:2211.15657, 2022. 2
[2] Seungyeon Baek, Erqun Dong, Shadan Namazifard, Mark J
Matthews, and Kwang Moo Yi. Sonic: Spectral optimiza-
tion of noise for inpainting with consistency. arXiv preprint
arXiv:2511.19985, 2025. 2
[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-
aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala,
Timo Aila, Samuli Laine, et al. eDiff-I: Text-to-image dif-
fusion models with an ensemble of expert denoisers. arXiv
preprint arXiv:2211.01324, 2022. 2
[4] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang,
Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu,
Qifeng Chen, Xintao Wang, et al. VideoCrafter1: Open
diffusion models for high-quality video generation. arXiv
preprint arXiv:2310.19512, 2023. 2
[5] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric
Cousineau, Benjamin Burchfiel, and Shuran Song. Diffu-
sion policy: Visuomotor policy learning via action diffusion.
arXiv preprint arXiv:2303.04137, 2023. 2
[6] Guillaume Couairon, Marl`
ene Careil, Matthieu Cord,
St´
ephane Lathuili`
ere, and Jakob Verbeek. Zero-shot spatial
layout conditioning for text-to-image diffusion models. In
Proceedings of the IEEE/CVF International Conference on
Computer Vision (ICCV), pages 2174–2183, October 2023.
3
[7] Sander Dieleman. Diffusion is spectral autoregression.
urlhttps://sander.ai/2024/09/02/spectral-autoregression.html,
September 2024. Accessed: 7 Dec 2025. 2
[8] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio
Lopez, and Vladlen Koltun. CARLA: An open urban driving
simulator, 2017. 6
[9] Xiang Gao, Shuai Yang, and Jiaying Liu. PTDiffusion: Free
lunch for generating optical illusion hidden pictures with
phase-transferred diffusion model. In Proceedings of the
Computer Vision and Pattern Recognition Conference, pages
18240–18249, 2025. 2
[10] Joseph W. Goodman. Statistical Optics. Wiley, 2 edition,
2015. Section 2.9.3. 4
[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu-
sion probabilistic models. Advances in Neural Information
Processing Systems, 33:6840–6851, 2020. 2,5
[12] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William
Chan, Mohammad Norouzi, and David J Fleet. Video dif-
fusion models. Advances in Neural Information Processing
Systems, 35:8633–8646, 2022. 2
[13] Huang et al. NanoControl: A lightweight framework for
precise and efficient control in diffusion transformer. arXiv
preprint arXiv:2508.10424, 2024. 3
[14] Michael Janner, Yilun Du, Joshua B Tenenbaum, and Sergey
Levine. Planning with diffusion for flexible behavior synthe-
sis. arXiv preprint arXiv:2205.09991, 2022. 2
[15] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and
Jingfeng Zhang. SCEdit: Efficient and controllable image
diffusion generation via skip connection editing. In Proceed-
ings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition (CVPR), pages 8995–9004, June 2024.
3
[16] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine.
Elucidating the design space of diffusion-based generative
models. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave,
and Kyunghyun Cho, editors, Advances in Neural Informa-
tion Processing Systems, 2022. 2
[17] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and
Bryan Catanzaro. DiffWave: A versatile diffusion model for
audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 2
[18] Black Forest Labs. Flux. https: //github .com/
black-forest-labs/flux, 2024. 11
[19] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow
straight and fast: Learning to generate and transfer data with
rectified flow. arXiv preprint arXiv:2209.03003, 2022. 4
[20] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu,
Bochen Guan, Yin Li, and Bolei Zhou. FreeControl:
Training-free spatial control of any text-to-image diffusion
model with any condition. In Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition
(CVPR), pages 7465–7475, June 2024. 3
[21] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian
Zhang, Zhongang Qi, and Ying Shan. T2I-Adapter: Learning
adapters to dig out more controllable ability for text-to-image
diffusion models. In AAAI, volume 38, pages 4296–4304,
2024. 2
[22] NVIDIA. Cosmos-Transfer1: Conditional world generation
with adaptive multimodal control. ArXiv, abs/2503.14492,
2025. 3,8
[23] Alan V. Oppenheim and Jae S. Lim. The importance of phase
in signals. Proceedings of the IEEE, 69(5):529–541, 1981.
2,3
[24] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, Ming-
Chang Yang, and Jiaya Jia. ControlNeXt: Powerful and effi-
cient control for image and video generation. arXiv preprint
arXiv:2408.06070, 2024. 3
[25] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima
Sadekova, and Mikhail Kudinov. Grad-TTS: A diffusion
probabilistic model for text-to-speech. In International Con-
ference on Machine Learning, pages 8599–8608. PMLR,
2021. 2
[26] Michael Psenka, Alejandro Escontrela, Pieter Abbeel, and
Yi Ma. Learning a diffusion model policy from rewards via
q-score matching. arXiv preprint arXiv:2312.11752, 2023.
2
[27] 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. 2
[28] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu,
and Mark Chen. Hierarchical text-conditional image gen-
eration with clip latents. arXiv preprint arXiv:2204.06125,
2022. 2
[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz,
Patrick Esser, and Bj¨
orn Ommer. High-resolution image syn-
thesis with latent diffusion models. In CVPR, pages 10684–
10695, June 2022. 11
[30] Daniel L. Ruderman and William Bialek. The statistics of
natural images. Network: Computation in Neural Systems,
5(4):517–548, 1994. 2
[31] Chitwan Saharia, William Chan, Saurabh Saxena, Lala
Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed
Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi,
Rapha Gontijo Lopes, et al. Photorealistic text-to-image
diffusion models with deep language understanding. arXiv
preprint arXiv:2205.11487, 2022. 2
[32] Zhenxiong Tan et al. OminiControl2: Efficient conditioning
for diffusion transformers. arXiv preprint arXiv:2503.08280,
2025. 2
[33] Alibaba Team Wan. Wan: Open and advanced large-scale
video generative models. arXiv preprint arXiv:2503.20314,
2025. 11
[34] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali
Dekel. Plug-and-play diffusion features for text-driven
image-to-image translation. In Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition
(CVPR), pages 1921–1930, June 2023. 5
[35] Julen Urain, Niklas Funk, Jan Peters, and Georgia Chal-
vatzaki. SE(3)-DiffusionFields: Learning smooth cost func-
tions for joint grasp and motion optimization through diffu-
sion. In 2023 IEEE International Conference on Robotics
and Automation (ICRA), pages 5923–5930. IEEE, 2023. 2
[36] Zhendong Wang, Jonathan J Hunt, and Mingyuan Zhou. Dif-
fusion policies as an expressive policy class for offline rein-
forcement learning. arXiv preprint arXiv:2208.06193, 2022.
2
[37] Zhou Wang and Eero P. Simoncelli. Translation insensitive
image similarity in complex wavelet domain. IEEE Transac-
tions on Image Processing, 14(4):466–479, 2005. 2
[38] Wei Wu, Qingnan Fan, Shuai Qin, Hong Gu, Ruoyu Zhao,
and Antoni B Chan. Freediff: Progressive frequency trunca-
tion for image editing with diffusion models. In European
Conference on Computer Vision, pages 194–209. Springer,
2024. 2
[39] Runsheng Xu, Hubert Lin, Wonseok Jeon, Hao Feng, Yu-
liang Zou, Liting Sun, John Gorman, Ekaterina Tolstaya,
Sarah Tang, Brandyn White, Ben Sapp, Mingxing Tan, Jyh-
Jing Hwang, and Dragomir Anguelov. WOD-E2E: Waymo
open dataset for end-to-end driving in challenging long-tail
scenarios, 2025. 6
[40] Cuihong Yu, Cheng Han, Chao Zhang, Yuewei Wang, Qi-
hang Hu, Yin Yan, Moran Zhan, Meng Li, and Guangjin Bi.
DMFFT: improving the generation quality of diffusion mod-
els using fast fourier transform. Scientific Reports, 15, March
2025. 2
[41] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and
Jian Zhang. FreeDoM: Training-free energy-guided condi-
tional diffusion model. In Proceedings of the IEEE/CVF In-
ternational Conference on Computer Vision (ICCV), pages
23174–23184, October 2023. 3
[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding
conditional control to text-to-image diffusion models. In
ICCV, pages 3836–3847, 2023. 2
[43] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin
Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K. Wong. Uni-
ControlNet: All-in-one control to text-to-image diffusion
models. In Advances in Neural Information Processing Sys-
tems 36 (NeurIPS), 2023. 2
Figure 10. Visual realism measured by Appearance Score.
A. Additional Implementation Details
We implemented ϕ-PD based on three models: SD
1.5 [29], FLUX-dev [18] and Wan2.2-14b [33]. We use the
implementation of these models from DiffSynthStudio2This
section describes the additional implementation details for
the experiment with each model.
A.1. Training and Inference Details
We start from the officially released checkpoints of
SD 1.5 [29], FLUX-dev [18], and Wan2.2-14B [33], and
finetune each model with phase-preserving noise. For
SD 1.5, we experiment with both full finetuning and LoRA
finetuning, while for FLUX-dev and Wan2.2-14B we use
LoRA finetuning due to computational constraints. At in-
ference time, for Wan2.2-14B we adopt the 4-step LoRA
from LightX2V3and apply it directly on top of our fine-
tuned LoRA weights to accelerate sampling.
We LoRA finetune Wan2.2-14B for 1,200 iterations and
FLUX-dev for 10,000 iterations, while SD 1.5 is fully fine-
tuned for 140,000 iterations. Each training run takes ap-
proximately 48 hours on an NVIDIA A100 GPU.
For each training iteration, we sample a cutoff radius r
from an exponential distribution and add a constant offset
r0to ensure a minimum amount of phase information is
always preserved:
r=r0+r′, r′∼Exp(λ),(21)
where λ > 0is the rate parameter of the exponential distri-
bution and r0>0controls the minimum cutoff. In our ex-
periments, we set λ= 0.1empirically. We set the transition
bandwidth parameter σ= 2, which controls the smoothness
of the frequency mask M(u, v)around the cutoff radius r.
B. Ablation Studies
We ablate the choices of r0, the minimal cutoff radius
at training time, and the inference-time cutoff radius rin
2https://github.com/modelscope/DiffSynth-Studio
3https://huggingface.co/lightx2v/Wan2.2-Lightning
Figure 11. Structural alignment measured by depth SSIM.
this section. All ablation experiments are conducted using
SD 1.5 with LoRA finetuning and evaluated on 1,000 ran-
domly selected samples from the UnrealCV test set. Fig-
ure 11 shows the Appearance Score (Sec. 4.1.3), which
measures the photorealism of re-rendered images. Fig-
ure 11 shows the depth consistency measured by SSIM, re-
flecting how well the re-rendered image aligns structurally
with the original input.
From both figures, we observe that increasing the cut-
off radius rat inference time improves structural alignment
at the cost of reduced photorealism. The minimal cutoff
threshold r0during training also affects performance across
different inference-time radii r. A higher r0during training
leads to better performance with higher rduring inference,
making it suitable for re-rendering images of decent qual-
ity where the model primarily refines details. Conversely,
a lower r0favors scenarios with smaller inference-time r,
performing better on low-quality inputs that require larger
visual changes to achieve photorealism. In our experiments,
we set r0= 4 for a balanced trade-off between structure
preservation and photorealism.
+