DiffIR2VR-Zero: Zero-Shot Video Restoration with Diffusion-based Image

Restoration Models

Chang-Han Yeh1Hau-Shiang Shiu1Chin-Yang Lin1Zhixiang Wang2

Chi-Wei Hsiao3Ting-Hsuan Chen1Yu-Lun Liu1

1National Yang Ming Chiao Tung University 2University of Tokyo 3MediaTek Inc.

InputResultInputResultInputResult

DenoisingSuper-resolutionDepth estimation

Figure 1. Zero-shot temporal-consistent diffusion model for video restoration and beyond. Given a pre-trained diffusion model for

single-image restoration, our method generates temporally consistent restored video with ﬁne details without any further training. Our

method applies to other video applications, such as depth estimation.

Abstract

We present DiffIR2VR-Zero, a zero-shot framework that

enables any pre-trained image restoration diffusion model

to perform high-quality video restoration without additional

training. While image diffusion models have shown remark-

able restoration capabilities, their direct application to

video leads to temporal inconsistencies, and existing video

restoration methods require extensive retraining for different

degradation types. Our approach addresses these challenges

through two key innovations: a hierarchical latent warping

strategy that maintains consistency across both keyframes

and local frames, and a hybrid token merging mechanism

that adaptively combines optical ﬂow and feature match-

ing. Through extensive experiments, we demonstrate that

our method not only maintains the high-quality restoration

of base diffusion models but also achieves superior tem-

poral consistency across diverse datasets and degradation

arXiv:2407.01519v5 [cs.CV] 31 Dec 2025

conditions, including challenging scenarios like 8

super-

resolution and severe noise. Importantly, our framework

works with any image restoration diffusion model, providing

a versatile solution for video enhancement without task-

speciﬁc training or modiﬁcations. Please see our project

page at jimmycv07.github.io/DiffIR2VR web.

1. Introduction

Video restoration — the task of transforming low-quality

videos into high-quality ones through denoising, super-

resolution, and deblurring — remains a signiﬁcant chal-

lenge in computer vision. While diffusion models have re-

cently revolutionized image restoration [

] by generat-

ing highly realistic details that surpass traditional regression-

based methods (Fig. 2(a)), extending these advances to video

has proven difﬁcult. Current approaches that directly apply

image diffusion models frame-by-frame suffer from severe

temporal inconsistencies and ﬂickering artifacts (Fig. 2(b)),

particularly with Latent Diffusion Models (LDMs).

Existing solutions typically attempt to bridge this gap by

ﬁne-tuning image diffusion models with video-speciﬁc com-

ponents like 3D convolution and temporal attention layers.

However, these approaches face signiﬁcant limitations: they

require extensive computational resources (e.g., 32 A100-

80G GPUs for video upscaling [

]), need task-speciﬁc re-

training, and often struggle to generalize across different

degradation types. This creates a pressing need for a more

efﬁcient and versatile approach to video restoration.

This paper presents DiffIR2VR-Zero, the ﬁrst training-

free framework that enables zero-shot video restoration using

any pre-trained image diffusion model. Unlike previous ap-

proaches that require extensive ﬁne-tuning or model modiﬁ-

cation, our method introduces two key innovations that work

in synergy: (i) Hierarchical latent warping that maintains

consistency at both global and local scales by intelligently

propagating latent features between keyframes and neigh-

boring frames. (ii) Hybrid ﬂow-guided spatial-aware token

merging that combines optical ﬂow, similarity matching,

and spatial information to achieve robust feature correspon-

dence across frames. These components work together to

enforce temporal consistency in both latent and token spaces

while preserving the high-quality restoration capabilities of

the underlying image diffusion model (Fig. 2(c)). Our ap-

proach requires no additional training or ﬁne-tuning, making

it immediately applicable to any pre-trained image diffusion

model.

As demonstrated in Fig. 1, our framework effectively

handles a diverse range of restoration tasks, including de-

noising, super-resolution, and even depth estimation, while

maintaining temporal consistency across frames. Extensive

experiments show that our method not only achieves state-

of-the-art performance in standard scenarios but also excels

Input frames (a) FMA-Net (b) DiffBIR (per-frame) (c) Ours

Current Frame

Next Frame

0.354 / 0.012 0.245 / 0.133 0.248 / 0.095

LPIPS (↓) / InterLPIPS (↓)

Figure 2. 4

video super-resolution results. (a) Traditional

regression-based methods such as FMA-Net [

] are limited to

the training data domain and tend to produce blurry results when

encountering out-of-domain inputs. (b) Although applying image-

based diffusion models such as DiffBIR [

] to individual frames

can generate realistic details, these details often lack consistency

across frames. (c) Our method leverages an image diffusion model

to restore videos, achieving both realistic and consistent results

without any additional training.

in extreme cases (such as 8× super-resolution and high-noise

denoising) where traditional methods struggle.

While our work builds upon recent advances in video

editing with diffusion models [

], we make several novel

contributions that speciﬁcally address the challenges of zero-

shot video restoration:

•

The ﬁrst zero-shot framework for adapting any pre-trained

image restoration diffusion model to video without ad-

ditional training, achieving a balance between temporal

consistency and detail preservation.

•

A training-free approach that innovatively combines hier-

archical latent warping with an improved token merging

strategy speciﬁcally designed for restoration tasks.

•

State-of-the-art performance across various restoration

tasks, demonstrating superior generalization and robust-

ness compared to existing methods, particularly in extreme

degradation scenarios.

2. Related Work

Video restoration. Video restoration encompasses trans-

forming degraded videos affected by noise, blur, and low

resolution into high-quality outputs [

]. Un-

like single-image restoration [

], video restoration faces

the additional challenge of maintaining temporal consistency

across frames. Current approaches primarily rely on motion-

based methods using optical ﬂow warping [

] or de-

formable convolutions [

] to align features temporally.

Other methods leverage attention mechanisms [

] to

model long-range dependencies across frames, while hybrid

approaches [

] combine multiple techniques to handle

complex degradations. However, these methods face several

critical limitations: they require extensive paired training

data [

], assume speciﬁc degradation models [

Figure 3. Pipeline of our proposed zero-shot video restoration method. We process low-quality (LQ) videos in batches using a diffusion

model, with a keyframe randomly sampled within each batch. (a) At the beginning of the diffusion denoising process, hierarchical latent

warping provides rough shape guidance both globally, through latent warping between keyframes, and locally, by propagating these latents

within the batch. (b) Throughout most of the denoising process, tokens are merged before the self-attention layer. For the downsample

blocks, optical ﬂow is used to ﬁnd the correspondence between tokens, and for the upsample blocks, cosine similarity is utilized. This

hybrid ﬂow-guided, spatial-aware token merging accurately identiﬁes correspondences between tokens by leveraging both ﬂow and spatial

information, thereby enhancing overall consistency at the token level.

and need retraining for different degradation levels [

These constraints signiﬁcantly limit their real-world applica-

bility and generalization capability.

Diffusion models for image restoration. Recent advances

in diffusion models [

] have led to breakthrough

improvements in image restoration. Current approaches ei-

ther train models from scratch [

], adapt pre-trained

models through guided sampling [

], or ﬁne-tune frozen

models with additional layers [

], as demonstrated by

StableSR and DiffBIR [

]. While these methods achieve

impressive results for single images, their direct application

to video leads to temporal inconsistencies due to the inher-

ent randomness of the diffusion process. Our work uniquely

bridges this gap by enabling zero-shot video restoration us-

ing these pre-trained image models without any additional

training or modiﬁcation.

Video consistency in diffusion models. Recent works

have explored extending image diffusion models to video

tasks [

]. Latent-space methods like Rerender-A-

Video [

] use warping and interpolation but struggle with

detail preservation in restoration. Token-level approaches

like VidToMe [

] and TokenFlow [

] often produce over-

smoothed results. Our work differs by combining hierarchi-

cal latent warping with hybrid correspondence mechanisms,

speciﬁcally designed for restoration tasks. This enables zero-

shot video restoration using any pre-trained image diffusion

model, without requiring task-speciﬁc training

3. Method

Given a low-quality video with

frames

y1, y2, . . . , yn

our goal is to restore it to high-quality

x1, x2, . . . , xn

using

image-based diffusion models. While direct frame-by-frame

application of these models causes temporal inconsistencies

due to inherent stochasticity, particularly in extreme degrada-

tion cases (Fig. 2and Fig. 6), our method (Fig. 3) addresses

this challenge through two key innovations: Hierarchical

Latent Warping (Sec. 3.2) and Hybrid Flow-guided Spatial-

aware Token Merging (Sec. 3.3). In this section, we ﬁrst

introduce the foundational concepts of diffusion models and

video token merging, then detail our novel components and

their integration.

3.1. Diffusion Models for Video Editing

The forward process of diffusion models progressively adds

noise to a clean image x0over Tsteps according to:

xt=√αtxt−1+√1−αtϵt−1⇒xt=√¯αtx0+√1−¯αtϵ,

(1)

where

t∼[1, T ]

ϵt, ϵ ∼ N(0,I)

, and

¯αt=Qt

s=1 αs

. A

UNet-based denoiser

ϵθ

learns to estimate and remove this

noise, with the inverse process gradually denoising

produce x0[11,34].

Recent video editing techniques like VidToMe [

] main-

tain temporal consistency by merging similar tokens within

frame chunks. Given a token chunk

T∈RB×A×C

where

A=w×h

, tokens are separated into source tokens

Tsrc

and a target token

Ttar

. The similarity between tokens is

Key frame Key frame Key frame

Step2: Latent Warping

Step 3: Token Merging with Optical Flow

Step 4: Token Merging with Similarity

Self

Attention

Step 1: Keyframe Latent Warping

Token

Unmerging

Cross

Attention

Token

Merging

Residual

Block

Correspondence

Figure 4. An illustration of our key modules. Without requiring

any training, these modules can achieve coherence across frames

by enforcing temporal stability in both latent and token space. Hier-

archical latent warping provides global and local shape guidance;

Hybrid spatial-aware token merging before the self-attention layer

improves temporal consistency by matching similar tokens using

optical ﬂow in the down blocks and cosine similarity in the up

blocks of the UNet.

computed as:

s(Tsrc,Ttar) = Tsrc ·Ttar

∥Tsrc∥∥Ttar∥, c = max

{t∈Ttar}(s(Tsrc,t)),(2)

where

s(·,·)

represents cosine similarity and

indicates cor-

respondences. The merging and unmerging operations are

deﬁned as:

Tmerge =M(Tsrc,Ttar, c, r),Tunmerge =U(Tmerge, c).

(3)

However, these existing techniques face signiﬁcant chal-

lenges in video restoration. Early-stage denoising produces

noisy latents that make traditional similarity measures un-

reliable, especially in UNet’s downsample blocks (Fig. 5).

Additionally, focusing solely on frame-to-frame consistency

misses global-local coherence, while aggressive token merg-

ing can lead to over-smoothing.

3.2. Hierarchical Latent Warping

Our key innovation in maintaining temporal consistency be-

gins with a hierarchical latent warping module that operates

at two distinct levels in the latent space. The ﬁrst level han-

dles global consistency across the video by warping between

keyframes, while the second level ensures local consistency

by propagating these warped latents within each processing

Step 10 Step 20 Step 30 Step 40

Figure 5. Token correspondences (cosine similarity and optical

ﬂow) across denoising steps. Early on (e.g., step 10), optical ﬂow

guides better due to noisy latents. Later (e.g., step 40), similarity

and ﬂow focus on different regions, showcasing the beneﬁt of our

hybrid approach for effective token merging throughout denoising.

batch. This hierarchical approach provides essential shape

guidance at both global and local scales, as illustrated in

Fig. 4(upper part). For a given latent

ˆxit→0

predicted for

the

ith

keyframe at denoising step

, we ﬁrst establish global

consistency through keyframe warping:

ˆxi

t→0←Mji ·ˆxi

t→0+(1−Mji)· W(ˆxj

t→0, fji),(4)

where

j=i−1

represents the previous keyframe. The

optical ﬂow ﬁeld

fji

and occlusion mask

Mji

are computed

from low-quality frames

lqj

lqi

using GMFlow [

]. This

formulation allows us to blend the original latent features

with warped features from the previous keyframe, weighted

by the occlusion mask to handle regions where warping may

be unreliable.

Following global alignment, we propagate these warped

latents to all frames within the current batch, establishing lo-

cal consistency. Unlike previous approaches that rely solely

on frame-to-frame warping, our hierarchical strategy ensures

that corresponding points maintain similar latent represen-

tations across both temporal scales from the early stages of

the denoising process. This comprehensive approach to tem-

poral consistency proves particularly effective in handling

complex motion patterns and maintaining coherent struc-

ture across the entire video sequence. To address potential

warping errors in severely degraded regions, we incorporate

forward-backward consistency checks and selective feature

propagation. This makes our approach more robust to op-

tical ﬂow failures and occlusions compared to single-scale

warping methods. Our experiments demonstrate that this

hierarchical strategy signiﬁcantly reduces temporal artifacts

while preserving ﬁne details in the restored video.

3.3. Hybrid Flow-guided Spatial-aware Token

Merging

While latent manipulation effectively maintains consistency

in early stages, it can produce blurry results when applied

during later denoising stages. To address this limitation, we

introduce a hybrid ﬂow-guided spatial-aware token merging

approach that operates in the semantically rich token space,

achieving both temporal consistency and detail preservation.

Flow guidance and spatial-awareness. Our key insight is

that different correspondence mechanisms are optimal at dif-

ferent stages of the denoising process. In early stages, when

latent representations are noisy, traditional cosine similar-

ity measures become unreliable, particularly in the UNet’s

downsample blocks (Fig. 5, top). During these stages, optical

ﬂow computed from low-resolution inputs provides more re-

liable guidance. As denoising progresses (e.g., steps 30-40),

ﬂow-based and similarity-based methods often identify com-

plementary correspondences (Fig. 5, bottom), motivating

our hybrid approach. For downsample blocks, we employ

ﬂow-guided correspondence with a forward-backward con-

sistency check to ensure reliability:

σ=e(−∥fsrc→tar(X(Tsrc))+ftar→src (X(Tsrc )+fsrc→tar (X(Tsrc )))∥2

2),

(5)

where

represents the conﬁdence score,

X(Tsrc)

denotes

the spatial location of source token

Tsrc

, and

fsrc→tar

ftar→src

represent forward and backward optical ﬂows. This conﬁ-

dence score guides the token merging process:

Tmerge =M(Tsrc,Ttar, fsrc→tar, σ, r).(6)

For upsample blocks, we enhance cosine similarity matching

with spatial awareness to prevent mismatches in uniform

texture regions (e.g., sky, grass). We weight similarity scores

based on spatial proximity:

s′

ij =sij ·e−τ,with τ=∥X(i)−X(j)∥2

2/R,(7)

where

X(i)

X(j)

are token spatial locations and

deﬁnes

the radius of inﬂuence. To prevent padding artifacts, we re-

move padding before merging and restore it after unmerging.

Merging ratio annealing. To maintain high-quality results

throughout the denoising process, we implement ratio an-

nealing that gradually reduces the merging ratio:

ri=r·cos π

2·max min δ·i−ibeg

iend −ibeg

,1,0,

(8)

where

ibeg

and

iend

deﬁne the annealing period, and

con-

trols the annealing speed. This annealing strategy helps bal-

ance temporal consistency with detail preservation, avoiding

the over-smoothing common in regression-based methods

while maintaining better temporal coherence than per-frame

processing (Fig. 2).

3.4. Scheduling

Our method orchestrates different components through the

denoising process. In early stages (steps 1-10), hierarchical

latent warping establishes global and local consistency by

warping between keyframes and propagating features within

batches. During the main denoising phase (steps 10-40), we

switch to hybrid spatial-aware token merging before each

attention layer. This mechanism adapts based on network

depth: optical ﬂow guides downsample blocks while simi-

larity matching handles upsample blocks. Throughout the

process, our annealing schedule gradually reduces the merg-

ing ratio, starting aggressively for consistency and becoming

more conservative to preserve details.

4. Experiments

We conduct extensive experiments to evaluate our zero-shot

video restoration framework across different tasks, datasets,

and degradation levels. We focus particularly on challeng-

ing scenarios that test both restoration quality and temporal

consistency.

Datasets and evaluation protocol. For video super-

resolution, we evaluate on three standard benchmarks:

REDS4 [

], Vid4 [

], and DAVIS [

]. We test at multiple

upscaling factors (

4 and

8) using the realistic degrada-

tion model from RealBasicVSR [

] to simulate real-world

conditions. For video denoising, we use REDS30 [

] and

Set8 [

], testing across a range of noise levels (std. = 50,

75, 100, 150) as well as random noise in the [50, 100] range

to evaluate robustness to varying degradation severity.

Evaluation metrics. Our evaluation considers both percep-

tual quality and temporal consistency through complemen-

tary metrics: (1) For perceptual quality, we use LPIPS for

assessing visual realism, alongside traditional PSNR and

SSIM metrics. (2) For temporal consistency, we employ

warping error (

Ewarp

), frame interpolation error, and our pro-

posed interpolation LPIPS. The latter metric extends the

interpolation error concept from [

] by using LPIPS to bet-

ter capture perceptual temporal consistency, measuring how

well interpolated frames match the actual frames.

Implementation details. We implement our framework us-

ing PyTorch and conduct experiments on an NVIDIA RTX

4090 GPU. To demonstrate the versatility of our approach,

we apply it to two different image restoration diffusion mod-

els: DiffBIR [

] and the SD

4 upscaler [

]. For achieving

super-resolution with models limited to 4

upscaling,

we cascade the process twice followed by bicubic downsam-

pling. While this approach works well with DiffBIR, we note

that memory constraints prevent its application with SDx4

on the REDS dataset due to the larger image sizes involved.

4.1. Comparisons with State-of-the-Art Methods

We conduct comprehensive comparisons with leading meth-

ods across different video restoration tasks, examining both

traditional learning-based approaches and recent diffusion-

based methods.

Video super-resolution. We compare against state-of-the-art

methods including BasicVSR++ [

], RVRT [

], and FMA-

Net [

]. Additionally, we evaluate against VidToMe [

]

applied to our base models and attempted comparisons with

Upscale-A-Video [

], though hardware limitations (48GB

A6000 GPU) prevented direct comparison due to memory

constraints.

Table 1. Quantitative comparisons of video super-resolution on the DAVIS [

], Vid4 [

] and REDS4 [

] datasets. The best and

second performances are marked in red and blue, respectively.

E∗

warp

denotes

Ewarp(×10−3)

and

Einter

, LPIPS

inter

denotes interpolation error

and LPIPS. - indicates out-of-memory.

BasicVSR++ [4] RVRT [23] FMA-Net [45] VidToMe [22] SD ×4 [1] DiffBIR [24] (ECCV 2024)

Metrics (CVPR 2022) (NeurIPS 2022) (CVPR 2024) (CVPR 2024) Frame Ours (Improve) Frame Ours (Improve)

DAVIS ×4

PSNR ↑26.576 26.595 25.215 23.014 23.504 23.843 (+0.339) 23.780 24.182 (+0.402)

SSIM ↑0.743 0.744 0.727 0.566 0.584 0.618 (+0.034) 0.601 0.621 (+0.020)

LPIPS ↓0.383 0.388 0.347 0.405 0.277 0.272 (-0.005) 0.264 0.262 (-0.002)

E∗

warp ↓0.090 0.090 0.186 0.520 0.912 0.745 (-0.167) 0.654 0.474 (-0.180)

Einter ↓9.115 9.135 11.558 13.676 18.125 17.431 (-0.694) 16.529 14.666 (-1.863)

LPIPSinter ↓0.058 0.058 0.078 0.329 0.292 0.274 (-0.018) 0.266 0.232 (-0.034)

DAVIS ×8

PSNR ↑24.301 24.504 22.690 22.097 20.268 20.519 (+0.251) 21.964 22.331 (+0.367)

SSIM ↑0.631 0.638 0.594 0.513 0.446 0.424 (-0.022) 0.502 0.519 (+0.017)

LPIPS ↓0.518 0.560 0.528 0.554 0.470 0.434 (-0.036) 0.362 0.367 (+0.005)

E∗

warp ↓0.132 0.127 0.351 0.440 2.199 1.759 (-0.440) 0.964 0.699 (-0.265)

Einter ↓9.882 9.725 13.978 12.624 24.496 21.746 (-2.750) 17.981 15.853 (-2.128)

LPIPSinter ↓0.088 0.081 0.132 0.388 0.457 0.442 (-0.015) 0.372 0.333 (-0.039)

REDS4 ×4

PSNR ↑27.227 27.244 25.829 23.134 24.189 24.226 (+0.037) 24.679 25.118 (+0.439)

SSIM ↑0.781 0.781 0.761 0.589 0.638 0.641 (+0.003) 0.657 0.683 (+0.026)

LPIPS ↓0.369 0.374 0.327 0.357 0.247 0.242 (-0.005) 0.211 0.222 (+0.011)

E∗

warp ↓0.134 0.133 0.392 0.579 0.817 0.811 (-0.006) 0.704 0.499 (-0.205)

Einter ↓15.799 15.838 19.014 17.869 22.906 22.889 (-0.017) 22.305 20.130 (-2.175)

LPIPSinter ↓0.106 0.101 0.133 0.356 0.295 0.281 (-0.014) 0.271 0.221 (-0.050)

REDS4 ×8

PSNR ↑26.109 26.226 22.842 21.894 20.601 20.622 (+0.021) 22.479 22.961 (+0.482)

SSIM ↑0.719 0.726 0.644 0.532 0.519 0.506 (-0.013) 0.559 0.590 (+0.031)

LPIPS ↓0.436 0.431 0.423 0.538 0.386 0.367 (-0.019) 0.311 0.306 (-0.005)

E∗

warp ↓0.127 0.129 0.753 0.423 1.928 1.735 (-0.247) 0.828 0.551 (-0.277)

Einter ↓15.753 15.822 21.519 15.502 26.886 25.503 (-1.383) 21.76 19.382 (-2.378)

LPIPSinter ↓0.099 0.099 0.159 0.412 0.370 0.388 (+0.018) 0.351 0.287 (-0.064)

REDS4 ×16

PSNR ↑23.579 23.715 21.569 20.520 18.706 18.858 (+0.152) 20.124 20.712 (+0.588)

SSIM ↑0.616 0.621 0.570 0.483 0.461 0.410 (-0.051) 0.461 0.509 (+0.048)

LPIPS ↓0.600 0.596 0.565 0.697 0.612 0.562 (-0.050) 0.446 0.438 (-0.008)

E∗

warp ↓0.084 0.085 0.619 0.296 2.664 2.030 (-0.634) 1.168 0.665 (-0.503)

Einter ↓14.069 14.267 18.758 12.945 28.478 24.000 (-4.478) 21.33 17.731 (-3.599)

LPIPSinter ↓0.088 0.088 0.139 0.417 0.559 0.493 (-0.066) 0.444 0.358 (-0.086)

Vid4 ×4

PSNR ↑23.142 23.160 23.209 19.622 20.047 20.134 (+0.087) 20.687 21.226 (+0.539)

SSIM ↑0.667 0.669 0.679 0.425 0.478 0.473 (-0.005) 0.497 0.525 (+0.028)

LPIPS ↓0.418 0.423 0.375 0.491 0.343 0.331 (-0.012) 0.329 0.326 (-0.003)

E∗

warp ↓0.173 0.167 0.203 0.687 1.502 1.397 (-0.105) 1.156 0.677 (-0.479)

Einter ↓3.398 3.399 4.442 11.754 17.234 16.921 (-0.313) 15.478 11.316 (-4.162)

LPIPSinter ↓0.015 0.015 0.026 0.337 0.275 0.271 (-0.004) 0.265 0.198 (-0.067)

Vid4 ×8

PSNR ↑21.601 21.707 21.033 18.811 17.813 17.992 (+0.179) 18.636 19.304 (+0.668)

SSIM ↑0.546 0.552 0.521 0.372 0.345 0.307 (-0.038) 0.367 0.406 (+0.039)

LPIPS ↓0.535 0.528 0.514 0.654 0.507 0.484 (-0.023) 0.440 0.435 (-0.005)

E∗

warp ↓0.139 0.151 0.221 0.477 2.523 1.972 (-0.551) 1.524 0.767 (-0.757)

Einter ↓3.170 3.193 5.269 9.942 22.881 19.970 (-2.911) 18.112 12.281 (-5.831)

LPIPSinter ↓0.011 0.011 0.032 0.393 0.423 0.419 (-0.004) 0.395 0.294 (-0.101)

Quantitative results in Tab. 1reveal several key ﬁndings:

regression-based methods like FMA-Net struggle with se-

vere degradation and large motion, while VidToMe achieves

temporal consistency but produces overly smooth results

with poor visual quality. Our method uniquely maintains

both the high-quality generation capabilities of the base dif-

fusion model and strong temporal consistency.

Visual comparisons in Fig. 6further demonstrate these

differences. FMA-Net’s results show clear limitations when

dealing with domain gaps between training and testing con-

ditions. While per-frame application of DiffBIR [

] and

SD×4 upscaler [

] produces sharp details, the results suffer

from temporal inconsistencies and jittering. Our approach

successfully combines high-ﬁdelity restoration with tem-

poral stability. For comparisons with Upscale-A-Video on

their test cases (Fig. 7), our method achieves superior detail

preservation, leveraging pre-trained diffusion priors more

effectively than their ﬁne-tuning approach.

Video denoising. In denoising experiments in Tab. 2, we ob-

serve that while regression models can perform adequately

VidToMe

DiffBIR

(per-frame)

Ours

(DiffBIR)

GT FMA-Net Input

BasicVSR++RVRT

Figure 6. Qualitative comparisons on 4

video super-resolution.

As shown in the ﬁrst row, the low-quality input lacks almost all

details. In the zoomed-in patches, our method produces clearer and

more consistent results.

Input Ours Upscale-A-Video Input Ours Upscale-A-Video

Figure 7. Qualitative comparisons with Upscale-A-Video [

]

on 4×video SR.

VidToMe

DiffBIR

(per-frame)

OursGT Shift-Net Input

Figure 8. Video denoising comparisons on the REDS30 [

]

dataset. Our method effectively denoises and generates detailed

results while maintaining temporal coherence.

with sufﬁcient batch sizes, our method consistently achieves

superior perceptual quality (LPIPS) and maintains this ad-

vantage even under severe degradation. As shown in Fig. 8,

Shift-Net [

] struggles with out-of-distribution noise lev-

els, and VidToMe produces temporally consistent but detail-

deﬁcient results. Per-frame DiffBIR generates high-quality

frames but suffers from temporal inconsistencies, particu-

larly noticeable in facial features and moving objects. Our

method successfully balances detail preservation with tem-

poral consistency.

Generalization to other tasks. To demonstrate the broad

applicability of our framework, we integrate it with

Marigold [

], a state-of-the-art monocular depth estimator

based on latent diffusion. Results in Fig. 9show signiﬁcant

improvements in temporal consistency of depth estimates

while maintaining accuracy. This successful adaptation to

Table 2. Quantitative comparisons of video denoising of various

noise levels on the REDS30 and Set8 [

] dataset. The best and

second performances are marked in red and blue, respectively.

E∗

warp

denotes

Ewarp(×10−3)

and

Einter

, LPIPS

inter

denotes interpolation

error and LPIPS.

VidToMe [22] Shift-Net [42] DiffBIR [24] (ECCV 2024)

σMetrics (CVPR 2024) (CVPR 2023) Frame Ours (Improve)

REDS30 75

PSNR ↑22.671 21.033 24.585 24.520 (-0.065)

SSIM ↑0.559 0.381 0.649 0.649 (+0.000)

LPIPS ↓0.397 0.735 0.276 0.275 (-0.001)

E∗

warp ↓0.727 0.765 0.751 0.706 (-0.045)

Einter ↓18.440 21.751 21.798 21.166 (-0.632)

LPIPSinter ↓0.375 0.501 0.275 0.264 (-0.011)

REDS30 100

PSNR ↑22.588 22.573 24.524 24.534 (+0.010)

SSIM ↑0.557 0.484 0.648 0.652 (+0.004)

LPIPS ↓0.404 0.518 0.275 0.271 (-0.004)

E∗

warp ↓0.733 1.126 0.763 0.696 (-0.067)

Einter ↓18.370 23.424 21.835 20.639 (-1.196)

LPIPSinter ↓0.380 0.375 0.281 0.267 (-0.014)

REDS30 random

PSNR ↑22.348 21.113 24.579 24.508 (-0.071)

SSIM ↑0.546 0.386 0.650 0.649 (-0.001)

LPIPS ↓0.429 0.728 0.276 0.270 (-0.006)

E∗

warp ↓0.681 1.896 0.755 0.713 (-0.042)

Einter ↓17.608 27.565 21.743 21.140 (-0.603)

LPIPSinter ↓0.384 0.542 0.282 0.272 (-0.010)

Set8 50

PSNR ↑21.531 23.433 23.197 23.713 (+0.516)

SSIM ↑0.501 0.482 0.594 0.630 (+0.036)

LPIPS ↓0.415 0.574 0.261 0.245 (-0.016)

E∗

warp ↓0.911 1.358 1.078 0.747 (-0.331)

Einter ↓17.217 19.845 19.732 16.814 (-2.918)

LPIPSinter ↓0.406 0.432 0.332 0.255 (-0.077)

Set8 100

PSNR ↑21.226 18.198 22.519 22.955 (+0.436)

SSIM ↑0.484 0.281 0.553 0.591 (+0.038)

LPIPS ↓0.472 0.733 0.338 0.323 (-0.015)

E∗

warp ↓0.918 2.229 1.13 0.802 (-0.328)

Einter ↓17.367 24.661 20.18 17.444 (-2.736)

LPIPSinter ↓0.421 0.619 0.372 0.286 (-0.086)

Set8 150

PSNR ↑20.209 16.136 21.005 21.418 (0.413)

SSIM ↑0.443 0.291 0.486 0.544 (0.058)

LPIPS ↓0.554 0.729 0.449 0.402 (-0.047)

E∗

warp ↓0.972 4.279 1.207 0.832 (-0.375)

Einter ↓17.872 22.343 20.729 17.616 (-3.113)

LPIPSinter ↓0.470 0.646 0.450 0.331 (-0.119)

Input

Marigold

Ours

Figure 9. Applying our techniques to consistent video depth.

Integrating our proposed framework into Marigold [

] helps im-

prove the temporal consistency of video depth estimation.

depth estimation, alongside our results in super-resolution

and denoising, highlights the versatility of our approach. As

the ﬁeld advances and more powerful image models emerge,

our framework’s zero-shot nature allows immediate leverage

of these improvements across video restoration tasks.

Table 3. Ablation studies for 8

VSR with different correspon-

dence matching methods on DAVIS [29] test sets.

Down

blocks

Spatial-

aware LPIPS ↓E∗

warp ↓LPIPSinter ↓

Flow Flow – 0.518 1.214 0.563

Cos Cos – 0.390 0.736 0.350

Cos Flow – 0.507 1.049 0.545

Flow Cos – 0.375 0.677 0.347

Flow Cos ✓0.367 0.699 0.333

Table 4. Ablation studies for 8

VSR with the proposed com-

ponents applied at different stages of the denoising process on

DAVIS [

] test sets. We apply our two proposed components,

hierarchical latent warping (HLW) and hybrid spatial-aware token

merging (HS-ToMe), at the early, mid, and late denoising stages.

HLW (Sec. 3.2) HS-ToMe (Sec. 3.3)

Early Mid Late Early Mid Late LPIPS ↓E∗

warp ↓LPIPSinter ↓

– – – – – – 0.362 0.964 0.372

✓– – ✓– – 0.368 0.887 0.369

✓ ✓ –✓ ✓ ✓ 0.43 0.804 0.383

✓ ✓ ✓ ✓ ✓ ✓ 0.411 0.704 0.339

✓– – ✓ ✓ ✓ 0.367 0.699 0.333

Table 5. Ablation studies on the merging ratio r.

Merging ratio rPSNR ↑SSIM ↑LPIPS ↓

0.3 22.814 0.483 0.358

0.6 23.308 0.522 0.477

0.9 23.169 0.518 0.478

0.6 →0 23.143 0.507 0.403

0.9 →0 (Ours) 23.302 0.518 0.428

4.2. Ablation Study

We conduct extensive ablation studies to validate our design

choices and analyze the impact of different components on

the ﬁnal performance. These experiments not only demon-

strate the effectiveness of our approach but also provide

insights into the interaction between different components.

Correspondence mechanism selection. We ﬁrst examine

different strategies for identifying token correspondences

across frames. As shown in Tab. 3, we systematically eval-

uate combinations of optical ﬂow and cosine similarity at

different stages of the UNet architecture. Our hybrid ap-

proach—using optical ﬂow in downsample blocks and co-

sine similarity in upsample blocks—achieves optimal results

across all metrics. This validates our insight that different

correspondence mechanisms are more effective at different

network depths. The addition of spatial awareness further

improves performance by preventing mismatches in texture-

less regions and ensuring locally coherent correspondence

matches.

Component scheduling analysis. We analyze the effec-

tiveness of our two key components—hierarchical latent

warping (HLW) and hybrid spatial-aware token merging

(HS-ToMe)—when applied at different stages of the denois-

ing process. Tab. 4shows that applying latent warping during

mid or late denoising stages signiﬁcantly degrades perfor-

mance. This conﬁrms our hypothesis that latent manipulation

is most effective in early stages when establishing coarse

structure, while token-level operations are crucial through-

out the process for maintaining ﬁne details and temporal

consistency.

A particularly interesting ﬁnding is that attempting to en-

force consistency in later stages through latent warping can

actually harm the restoration quality, likely due to the increas-

ing semantic richness of the latent space. This insight guided

our design choice to transition from latent-based to token-

based consistency enforcement as denoising progresses. Ex-

tensive temporal proﬁle comparisons and additional ablation

results are provided in the supplementary materials.

Token merging ratio r.The effectiveness of our method is

signiﬁcantly inﬂuenced by the token merging ratio r, which

controls the balance between temporal consistency and detail

preservation. We conduct extensive ablation studies on this

hyperparameter, comparing our annealing strategy (reduc-

ing r from 0.9 to 0) against ﬁxed ratios. Tab. 5show that

higher ﬁxed ratios (0.9, 0.6) tend to improve ﬁdelity metrics

(PSNR/SSIM) by enforcing stronger temporal consistency,

but at the cost of perceptual quality (higher LPIPS) due to

over-smoothing. Conversely, lower ﬁxed ratios (0.3) preserve

more details but sacriﬁce temporal coherence. Our anneal-

ing strategy achieves the best overall performance (PSNR:

23.302, SSIM: 0.518, LPIPS: 0.428) by adaptively reducing

the merging ratio throughout the denoising process, effec-

tively balancing the trade-off between temporal consistency

and detail preservation.

5. Conclusion

We have presented DiffIR2VR-Zero, a novel framework

enabling zero-shot video restoration using pre-trained im-

age diffusion models without additional training. Our ap-

proach combines hierarchical latent warping with hybrid

ﬂow-guided token merging to maintain temporal consistency

while preserving high-quality restoration capabilities. Exten-

sive experiments demonstrate state-of-the-art performance

across various restoration tasks and remarkable robustness

to severe degradations.

Limitations. Our framework faces two main limitations:

ﬂickering artifacts in dynamic scenes due to LDM decoder

sensitivity, and reduced performance under extreme degrada-

tion scenarios. Future work will focus on stabilizing decoder

output and enhancing degradation handling capabilities. As

more powerful image diffusion models emerge, our frame-

work’s zero-shot nature will allow immediate leverage of

these advances for improved video restoration.

References

[1] Stable diffusion x4 upscaler, 2023. 5,6,12

[2]

Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool.

Video super-resolution transformer. arXiv preprint

arXiv:2106.06847, 2021. 2

[3]

Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and

Chen Change Loy. Basicvsr: The search for essential compo-

nents in video super-resolution and beyond. In CVPR, 2021.

[4]

Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and

Chen Change Loy. Basicvsr++: Improving video super-

resolution with enhanced propagation and alignment. In

CVPR, 2022. 2,5,6

[5]

Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and

Chen Change Loy. Investigating tradeoffs in real-world video

super-resolution. In CVPR, 2022. 2,5

[6]

Ziyan Chen, Jingwen He, Xinqi Lin, Yu Qiao, and Chao Dong.

Towards real-world video face restoration: A new benchmark.

arXiv preprint arXiv:2404.19500, 2024. 11

[7]

Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang,

Han Hu, and Yichen Wei. Deformable convolutional networks.

In ICCV, 2017. 2

[8]

Prafulla Dhariwal and Alexander Nichol. Diffusion models

beat gans on image synthesis. In NeurIPS, 2021. 3

[9]

Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel.

Tokenﬂow: Consistent diffusion features for consistent video

editing. arXiv preprint arXiv:2307.10373, 2023. 2,3

[10]

Shi Guo, Zifei Yan, Kai Zhang, Wangmeng Zuo, and Lei

Zhang. Toward convolutional blind denoising of real pho-

tographs. In CVPR, 2019. 2

[11]

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu-

sion probabilistic models. In NeurIPS, 2020. 3

[12]

Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan,

Mohammad Norouzi, and David J Fleet. Video diffusion

models. In NeurIPS, 2022. 3

[13]

Yaosi Hu, Zhenzhong Chen, and Chong Luo. Lamd: La-

tent motion diffusion for video generation. arXiv preprint

arXiv:2304.11603, 2023. 3

[14]

Zhaoyang Huang, Xiaoyu Shi, Chao Zhang, Qiang Wang,

Ka Chun Cheung, Hongwei Qin, Jifeng Dai, and Hongsheng

Li. Flowformer: A transformer architecture for optical ﬂow.

In ECCV, 2022. 2

[15]

Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin

Wang, and Qi Tian. Video super-resolution with recurrent

structure-detail network, 2020. 2

[16]

Ozgur Kara, Bariscan Kurtkaya, Hidir Yesiltepe, James M.

Rehg, and Pinar Yanardag. Rave: Randomized noise shufﬂing

for fast and consistent video editing with diffusion models. In

CVPR, 2024. 3

[17]

Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming

Song. Denoising diffusion restoration models. In NeurIPS,

2022. 3

[18]

Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Met-

zger, Rodrigo Caye Daudt, and Konrad Schindler. Repurpos-

ing diffusion-based image generators for monocular depth

estimation. In CVPR, 2024. 7,11,12,14

[19]

Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li,

and Jinshan Pan. Efﬁcient frequency domain-based trans-

formers for high-quality image deblurring. In CVPR, 2023.

[20]

Dasong Li, Xiaoyu Shi, Yi Zhang, Ka Chun Cheung, Simon

See, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. A

simple baseline for video restoration with grouped spatial-

temporal shift. In CVPR, 2023. 2,7

[21]

Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya

Jia. Mucan: Multi-correspondence aggregation network for

video super-resolution. In ECCV, 2020. 2

[22]

Xirui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang.

Vidtome: Video token merging for zero-shot video editing. In

CVPR, 2024. 2,3,5,6,7,12

[23]

Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan,

Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Tim-

ofte, and Luc V Gool. Recurrent video restoration transformer

with guided deformable attention. In NeurIPS, 2022. 2,5,6

[24]

Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai,

Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diffbir:

Towards blind image restoration with generative diffusion

prior, 2024. 2,3,5,6,7,12

[25]

Ce Liu and Deqing Sun. On bayesian adaptive video super

resolution. IEEE TPAMI, 2013. 5,6

[26]

Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu

Chuang, and Jia-Bin Huang. Hybrid neural fusion for full-

frame video stabilization. In ICCV, 2021. 2

[27]

Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik

Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee.

Ntire 2019 challenge on video deblurring and super-

resolution: Dataset and study. In CVPRW, 2019. 5,6,7

[28]

Jinshan Pan, Haoran Bai, and Jinhui Tang. Cascaded deep

video deblurring using temporal sharpness prior. In CVPR,

2020. 2

[29]

F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M.

Gross, and A. Sorkine-Hornung. A benchmark dataset and

evaluation methodology for video object segmentation. In

CVPR, 2016. 5,8

[30]

Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc

Van Gool, Markus Gross, and Alexander Sorkine-Hornung.

A benchmark dataset and evaluation methodology for video

object segmentation. In CVPR, 2016. 6

[31]

Robin Rombach, Andreas Blattmann, Dominik Lorenz,

Patrick Esser, and Bj

orn Ommer. High-resolution image

synthesis with latent diffusion models. In CVPR, 2022. 3

[32]

Chitwan Saharia, Jonathan Ho, William Chan, Tim Sali-

mans, David J Fleet, and Mohammad Norouzi. Image super-

resolution via iterative reﬁnement. IEEE TPAMI, 2022. 3

[33]

Xiaoyu Shi, Zhaoyang Huang, Weikang Bian, Dasong Li,

Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei

Qin, Jifeng Dai, and Hongsheng Li. Videoﬂow: Exploiting

temporal cues for multi-frame optical ﬂow estimation. In

ICCV, 2023. 2

[34]

Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising

diffusion implicit models. arXiv preprint arXiv:2010.02502,

2020. 3

[35]

Matias Tassano, Julie Delon, and Thomas Veit. Dvdnet: A

fast network for deep video denoising. In ICIP, 2019. 7

[36]

Matias Tassano, Julie Delon, and Thomas Veit. Fastdvdnet:

Towards real-time deep video denoising without ﬂow estima-

tion. In CVPR, 2020. 5

[37]

Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK

Chan, and Chen Change Loy. Exploiting diffusion prior for

real-world image super-resolution. IJCV, 2024. 3

[38]

Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and

Chen Change Loy. Edvr: Video restoration with enhanced

deformable convolutional networks. In CVPRW, 2019. 2

[39]

Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong

Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Difﬁr:

Efﬁcient diffusion model for image restoration. In ICCV,

2023. 2,3

[40]

Liangbin Xie, Xintao Wang, Shuwei Shi, Jinjin Gu, Chao

Dong, and Ying Shan. Mitigating artifacts in real-world video

super-resolution models. In AAAI, 2023. 2

[41]

Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatoﬁghi, and

Dacheng Tao. Gmﬂow: Learning optical ﬂow via global

matching. In CVPR, 2022. 4

[42]

Zhaoyi Yan, Xiaoming Li, Mu Li, Wangmeng Zuo, and

Shiguang Shan. Shift-net: Image inpainting via deep fea-

ture rearrangement. In ECCV, 2018. 7,12

[43]

Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy.

Rerender a video: Zero-shot text-guided video-to-video trans-

lation. In SIGGRAPH Asia 2023 Conference Papers, 2023.

[44]

Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-

aware stable diffusion for realistic image super-resolution and

personalized stylization. In ECCV, 2024. 3

[45]

Geunhyuk Youk, Jihyong Oh, and Munchurl Kim. Fma-net:

Flow-guided dynamic ﬁltering and iterative feature reﬁne-

ment with multi-attention for joint video super-resolution and

deblurring. In CVPR, 2024. 2,3,5,6,12

[46]

Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Resshift:

Efﬁcient diffusion model for image super-resolution by resid-

ual shifting. In NeurIPS, 2024. 3

[47]

Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu-

nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang.

Restormer: Efﬁcient transformer for high-resolution image

restoration. In CVPR, 2022. 2

[48]

Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo,

and Chen Change Loy. Upscale-a-video: Temporal-consistent

diffusion model for real-world video super-resolution. In

CVPR, 2024. 2,5,7,12

(a) Correspondences without spatial-awareness and padding removal

(b) Correspondences without spatial-awareness

Figure 10. Correspondences at denoising step 40 for different

settings.

A. Appendix Section

In this supplementary material, we ﬁrst provide additional

details on the testing datasets and evaluation metrics. Then,

we report the computational complexity and inference time

comparison. Subsequently, we present more visual compar-

isons of various methods.

A.1. Ablation Studies on Correspondences Identi-

ﬁed by Cosine Similarity

Fig. 10 The ﬁgure shows the correspondences at denoising

step 40 for three scenarios: without spatial awareness and

padding removal, without spatial awareness, and with both

spatial awareness and padding removal (ours). It is evident

that padding values signiﬁcantly affect the matching quality.

However, even after removing padding, many mismatched

diagonal lines remain, leading to blurry results. In contrast,

our method effectively ﬁnds accurate correspondences by

leveraging spatial information from the video.

A.2. Severe Degradation Scenarios

Our balanced approach proves particularly effective in severe

degradation scenarios. For instance, in 8

super-resolution

tasks, our method not only avoids artifacts but can even

improve visual quality compared to per-frame approaches

(Fig. 11). Additionally, in the 4

video face super-resolution

DiffBIR

Ours

SDx4 upscaler

Ours

Figure 11. Applying our method on DiffBIR and SD

4 upscaler

for 8

SR task. In this case of severe degradation, our method

avoids artifacts and outperforms per-frame inference in terms of

visual quality.

Input frames FMA-Net DiffBIR (per-frame) Ours

Figure 12. Additional qualitative comparisons on 4

video

super-resolution. In the zoomed-in patches, our method produces

clearer and more consistent results.

dataset [

], our results contain more details compared to

FMA-Net and are temporally more consistent than per-frame

method DiffBIR as shown in Fig. 14. This underscores the

effectiveness of our ratio annealing technique in addressing

the over-smoothing tendency while maintaining the beneﬁts

of our token merging approach. Additional comparisons on

video super-resolution can be found at Fig. 12 and Fig. 13.

Other Video Tasks: Consistent Video Depth. Our zero-

shot framework is applicable to any pre-trained image-based

diffusion models and could improve the predicted video

consistency. Therefore, we integrate our proposed zero-

shot framework into a state-of-the-art latent diffusion-based

monocular depth estimator: Marigold [

]. Fig. 15 shows

that integrating our proposed framework into Marigold helps

improve the temporal consistency of video depth estimation.

A.3. Computational Complexity

While our method focuses on zero-shot video restoration

without additional training, it’s important to consider the

VidToMe

DiffBIR

(per-frame)

Ours

(DiffBIR)

GT FMA-Net Input

BasicVSR++

RVRT

Figure 13. Additional qualitative comparisons on 8

video

super-resolution. As shown in the ﬁrst row, the low-quality in-

put lacks almost all details. In the zoomed-in patches, our method

produces clearer and more consistent results.

InputFMA-NetDiffBIROurs

Figure 14. Additional qualitative comparisons on 4

video face

super-resolution.

computational requirements in comparison to other ap-

proaches. Tab. 6provides an overview of the training time

and GPU speciﬁcations for different methods, including

ours.

As shown in the table, our method stands out by not requir-

ing any training or ﬁne-tuning, which signiﬁcantly reduces

the computational resources needed. This is in stark con-

trast to other methods that require multiple high-end GPUs

and several days of training time. For inference, our method

introduces some computational overhead due to the hierar-

chical latent warping and hybrid token merging processes.

However, this overhead is relatively small compared to the

resources required for training or ﬁne-tuning video models.

Speciﬁcally, our method adds only approximately 6 seconds

to the inference time of the base image diffusion model per

frame.

A.4. Inference Time Comparison

We report the inference time for processing 10 video frames

at 854

480 resolution on a single 4090 GPU in Tab. 7. Our

method adds a reasonable overhead compared to per-frame

Table 6. Training time and used devices for different methods.

Method Training time GPU specs

Shift-Net [42] Not reported 8 NVIDIA A100-32G GPUs

FMA-Net [45] Not reported Not reported

Upscale-A-Video [48] Not reported 32 NVIDIA A100-80G GPUs

Ours No training needed -

Table 7. Inference time different methods.

Method Inference time

VidToMe [22] 1m 49s

FMA-Net [45] 4.7s

SDx4 [1] per-frame 41s

SDx4 [1] + Ours 1m 7s

DiffBIR [24] per-frame 1m 17s

DiffBIR [24] + Ours 2m 20s

Shift-Net [42] 12.7s

Marigold [18] (4-step) + Ours 10s

Upscale-a-Video [48] OOM

Table 8. Quantitative comparisons of different unmerging meth-

ods on Vid4 x4 SR task.

Unmerging Method LPIPS ↓

Averaging 0.337

Replacement 0.329

inference (around 26s for SDx4 and 63s for DiffBIR) while

maintaining strong temporal consistency. This is notably

more efﬁcient than training-based methods like Upscale-

A-Video, which requires 32 A100 GPUs and encounters

out-of-memory (OOM) issues even during inference on our

test setup. Furthermore, when applied to lightweight models

like Marigold with 4-step sampling, our method achieves

very fast inference at just 10 seconds total.

A.5. Additional Ablation Studies

Comparison of temporal proﬁles. The comparisons in

Fig. 16 also indicate that our results are smoother, demon-

strating better temporal stability.

Token Unmerging Strategies. We experimented with two

unmerging strategies: averaging paired tokens and direct re-

placement with keyframe tokens. Tab. 8shows the results

of these experiments on the Vid4 x4 SR task. As shown

in the table, the replacement method outperforms averag-

ing in terms of LPIPS, indicating better perceptual quality.

Our experiments consistently showed that averaging tends

to produce blurrier outputs in restoration tasks. Based on

these results, we adopted the replacement-based unmerging

process in our ﬁnal model, as it preserves more details and

leads to sharper outputs.

Limitations: Extreme Degradation Extreme degradation

(e.g., 32

super-resolution) or overly detailed facial fea-

tures may yield unsatisfactory results (Fig. 17). However,

our framework’s adaptability allows the incorporation of

future, more powerful image-based diffusion models. Fu-

ture improvements will focus on reﬁning keyframe selection,

stabilizing decoder output across LDM architectures, and

enhancing extreme degradation handling. These aim to im-

prove practical application and mitigate ﬂickering issues

inherent in LDM decoders.

Input

Marigold

Ours

Input

Marigold

Ours

Input

Marigold

Ours

Input

Marigold

Ours

Figure 15. Applying our techniques to consistent video depth. Integrating our proposed framework into Marigold [

] helps improve the

temporal consistency of video depth estimation.

Input Video

Flow + Cosine

Flow + Flow

Cosine + Cosine

Cosine + Flow

Ours

𝑥

𝑡

𝑥

Figure 16. Comparison of temporal proﬁle. We examine a row of

pixels and track changes over time. The proﬁles from Flow + Flow

and Cosine + Flow methods exhibit noise, indicating ﬂickering

artifacts. The Cosine + Cosine method shows smoother proﬁles

but contains some discontinuities. Flow + Cosine demonstrates

improved consistency but retains some distortions. Utilizing ﬂow,

cosine, and spatial-aware techniques, our method achieves the most

seamless and consistent transitions, effectively minimizing artifacts.

(a) DDNM (b) FMA-Net (c) Ours

Figure 17. Failure case under 32x SR. Most methods fail under

this extreme degradation. However, if more powerful image-based

diffusion models emerge in the future, our method can be easily

adapted, offering greater potential to achieve this task.