SWIFT: Sliding Window Reconstruction for Few-Shot Training-Free Generated Video Attribution

Chao Wang; Zijin Yang; Yaofei Wang; Yuang Qi; Weiming Zhang; Nenghai Yu; Kejiang Chen

cs.CVMar 9, 2026

SWIFT: Sliding Window Reconstruction for Few-Shot Training-Free Generated Video Attribution

Authors: Chao Wang, Zijin Yang, Yaofei Wang, Yuang Qi, Weiming Zhang, Nenghai Yu, Kejiang Chen

Organizations: University of Science and Technology of China · Hefei University of Technology

Abstract

Recent advancements in video generation technologies have been significant, resulting in their widespread application across multiple domains. However, concerns have been mounting over the potential misuse of generated content. Tracing the origin of generated videos has become crucial to mitigate potential misuse and identify responsible parties. Existing video attribution methods require additional operations or the training of source attribution models, which may degrade video quality or necessitate large amounts of training samples. To address these challenges, we define for the first time the "few-shot training-free generated video attribution" task and propose SWIFT, which is tightly integrated with the temporal characteristics of the video. By leveraging the "Pixel Frames(many) to Latent Frame(one)" temporal mapping within each video chunk, SWIFT applies a fixed-length sliding window to perform two distinct reconstructions: normal and corrupted. The variation in the losses between two reconstructions is then used as an attribution signal. We conducted an extensive evaluation of five state-of-the-art (SOTA) video generation models. Experimental results show that SWIFT achieves over 90% average attribution accuracy with merely 20 video samples across all models and even enables zero-shot attribution for HunyuanVideo, EasyAnimate, and Wan2.2. Our source code is available at https://github.com/wangchao0708/SWIFT.

Figures & tables

Figure 1: Comparison of existing attribution methods and SWIFT: (a) Watermark-based active attribution; (b) Training-based passive attribution; (c) SWIFT: few-shot training-free passive attribution.

Figure 2: The 3D VAE performs up and down sampling operations along the temporal dimension (Temporal Compression Ratio = 4) and executes two distinct reconstructions (Normal and Corrupted).

Figure 3: The framework of SWIFT consists of three key modules: determination of a fixed-length sliding window, normal and corrupted reconstruction, and threshold determination. SWIFT uses the average loss ratio of overlapping frames between two differential reconstructions as the attribution signal. This signal is then compared with the threshold obtained through KDE to derive the final attribution result.

Dataset / Model	Specific Type	Release Date	Resolution (W $\times$ H)	Frames	FPS	Quantity
OpenVid-1M [ 18 ]	Real	2024-07-01	512 $\times$ 360 $\sim$ 1024 $\times$ 1024	30 $\sim$ 196	20 $\sim$ 60	500
HunyuanVideo [ 13 ]	T2V	2024-12-03	960 $\times$ 544	129	24	700
Wan2.1 [ 26 ]	T2V-1.3B	2025-02-25	832 $\times$ 480	81	16	700
EasyAnimate [ 34 ]	T2V-7B-V5.1	2025-03-06	1008 $\times$ 576	49	8	700
LTX-Video [ 8 ]	TI2V-13B-0.9.7	2025-05-05	1216 $\times$ 704	121	30	700
Wan2.2 [ 26 ]	TI2V-5B	2025-07-28	1280 $\times$ 704	121	24	700

Table 1: Detailed composition information of the S-Video dataset.

$\mathcal{M}_{1}$	$\mathcal{M}_{2}$	AEDR [ 27 ]					SWIFT (ours)
$\mathcal{M}_{1}$	$\mathcal{M}_{2}$	TP	FP	FN	TN	Acc	TP	FP	FN	TN	Acc
HYV	Wan2.1	481	465	19	35	51.6%	473	6	27	494	96.7% (+45.1)
	EA	481	375	19	125	60.6%	473	7	27	493	96.6% (+36.0)
	LTX	481	467	19	33	51.4%	473	171	27	329	80.2% (+28.8)
	Wan2.2	481	476	19	24	50.5%	473	143	27	357	83.0% (+32.5)
	Real	481	433	19	47	54.8%	473	4	27	496	96.9% (+42.1)

Table 2: Evaluation of SWIFT’s attribution effectiveness.

Model	HYV [ 13 ]	Wan2.1 [ 26 ]	EA [ 34 ]	LTX [ 8 ]	Wan2.2 [ 26 ]
Precision	FP16	FP32	FP16	BF16	FP32
Generation	2369s	349s	421s	39s	656s
AEDR	227s	46s	31s	17s	186s
SWIFT	218s (-9s)	33s (-13s)	21s (-10s)	14s (-3s)	170s (-16s)
Speedup	4%	28%	32%	18%	9%

Table 3: Running time on different models.

Model	HYV [ 13 ]	Wan2.1 [ 26 ]	EA [ 34 ]	Wan2.2 [ 26 ]	LTX [ 8 ]
Spatial–Temporal Compression	8 $\times$ 8 $\times$ 4	8 $\times$ 8 $\times$ 4	8 $\times$ 8 $\times$ 4	16 $\times$ 16 $\times$ 4	32 $\times$ 32 $\times$ 8
Number of Output Channels	16	16	16	48	128
Total Compression	48	48	48	64	192
Functions of VAE	Up and Down Sampling				Up and Down Sampling & Final Denoising Step
SWIFT	90.7%	98.4%	97.8%	97.9%	85.3%

Table 4: The generalization capability of SWIFT across different types and functions of 3D VAE.

Loss Metric	TP	FP	FN	TN	Acc
MAE	483	495	17	5	97.8%
MSE	484	500	16	0	98.4 %
PSNR	478	0	22	500	47.8%
SSIM [ 32 ]	470	1	30	499	47.1%
LPIPS [ 38 ]	479	499	21	1	97.8%

Table 5: Attribution accuracy of different loss metrics.

Model	Normal	Corrupted	Acc	Normal	Corrupted	Acc
HYV $K=4$	$W_{0}$	$W_{1}$	82.3%	$W_{4}$	$W_{1}$	73.3%
		$W_{2}$	82.3%		$W_{2}$	72.4%
		$W_{3}$	90.7%		$W_{3}$	88.0%
LTX $K=8$	$W_{0}$	$W_{1}$	81.6%	$W_{8}$	$W_{1}$	49.5%
		$W_{2}$	76.8%		$W_{2}$	50.1%
		$W_{3}$	85.3%		$W_{3}$	53.4%

Table 6: Attribution accuracy of different window selection.

$S$	Attribution Accuracy					Avg Acc
$S$	HYV	Wan2.1	EA	LTX	Wan2.2	Avg Acc
0	90.1%	81.7%	89.3%	66.1%	98.4%	85.1%
5	78.6%	88.7%	80.3%	79.0%	91.8%	83.7%
10	76.1%	88.3%	80.8%	78.3%	90.4%	82.8%
20	85.2%	94.2%	95.0%	83.9%	92.7%	90.2%
35	90.3%	96.9%	97.4%	84.8%	93.2%	92.5%

Table 7: Impact of sample size on attribution accuracy.

Bandwidth	Kernel Function	Avg Acc
Scott	Gaussian	97.80%
	Uniform	97.40%
	Epanechnikov	97.40%
Silverman	Gaussian	97.80%
	Uniform	97.30%
	Epanechnikov	97.40%

Table 8: Impact of bandwidth and kernel function on accuracy.

Length	Frames	Runtime	Accuracy
100%	121	170s	98.0%
75%	93	134s	98.0%
50%	61	84s	97.1%
25%	33	40s	94.1%

Table 9: Impact of video length on accuracy and runtime.

Method	w/o	Crop	Compression	Flip	Noise
AEDR	63.1%	37.2%	42.6%	29.4%	49.9%
SWIFT	97.8%	81.9% (+44.7)	88.3% (+45.7)	52.0% (+22.6)	60.3% (+10.4)

Table 10: Performance of SWIFT on post-processing operations.

Video Forensics

See all 20 papers

Jul 31, 2026

Retrieval-Driven Training-Free AI-Generated Video Attribution

AI-generated videos are becoming increasingly realistic and difficult to distinguish from authentic ones, which facilitates malicious misuse and poses growing threats to cybersecurity and social governance. Attributing AI-generated videos to their specific generative sources is therefore of critical importance for forensic investigation and legal regulation. However, most existing visual attribution methods focus on images and particularly rely on the image generation model, thereby lacking the ability to generalize to large-scale AI-generated video data. To address these limitations, we introduce an training-free AI-generated video attribution paradigm. Specifically, we formulates AI-generated video attribution as an instance retrieval task, and design a generative fingerprint-based pipeline. This pipeline consists of an adapted orthogonal color transformation, multi-scale quantized residual generation, and temporal-semantic aggregation, progressively capturing and integrating artifacts introduced by generative models across video frames. Extensive experiments on the GenVidBench benchmark demonstrate that our method achieves strong performance in both AI-generated video detection and attribution, outperforming existing state-of-the-art methods with a Rank-1 accuracy of 20.5% and a mean Average Precision of 16.6%. The code is at https://github.com/renxi-seu/Video_Attribution.

Renxi Cheng, Chaolei Han, Jie Gui +1

School of Cyber Science and Engineering, Southeast University, Nanjing 210096, China · Purple Mountain Laboratories, Nanjing 210000, China · Engineering Research Center of Blockchain Application, Supervision And Management (Southeast University), Ministry of Education, China +2

Deepfake Detection

Ask PDF

Mar 16, 2026

Training-free Detection of Generated Videos via Spatial-Temporal Likelihoods

Following major advances in text and image generation, the video domain has surged, producing highly realistic and controllable sequences. Along with this progress, these models also raise serious concerns about misinformation, making reliable detection of synthetic videos increasingly crucial. Image-based detectors are fundamentally limited because they operate per frame and ignore temporal dynamics, while supervised video detectors generalize poorly to unseen generators, a critical drawback given the rapid emergence of new models. These challenges motivate zero-shot approaches, which avoid synthetic data and instead score content against real-data statistics, enabling training-free, model-agnostic detection. We introduce STALL, a simple, training-free, theoretically justified detector that provides likelihood-based scoring for videos, jointly modeling spatial and temporal evidence within a probabilistic framework. We evaluate STALL on two public benchmarks and introduce ComGenVid, a new benchmark with state-of-the-art generative models. STALL consistently outperforms prior image- and video-based baselines. Code and data are available at https://omerbenhayun.github.io/stall-video.

Omer Ben Hayun, Roy Betser, Meir Yossef Levi +2

Technion – Israel Institute of Technology, Haifa, Israel

Topic pending

Ask PDF

Feb 19, 2026

EA-Swin: An Embedding-Agnostic Swin Transformer for AI-Generated Video Detection

Recent advances in foundation video generators such as Sora2, Veo3, and other commercial systems have produced highly realistic synthetic videos, exposing the limitations of existing detection methods that rely on shallow embedding trajectories, image-based adaptation, or computationally heavy MLLMs. We propose EA-Swin, an Embedding-Agnostic Swin Transformer that models spatiotemporal dependencies directly on pretrained video embeddings via a factorized windowed attention design, making it compatible with generic ViT-style patch-based encoders. Moreover, we construct the EA-Video dataset, a benchmark dataset comprising 130K videos that integrates newly collected samples with curated existing datasets, covering diverse commercial and open-source generators and including unseen-generator splits for rigorous cross-distribution evaluation. Extensive experiments show that EA-Swin achieves 0.97-0.99 accuracy across major generators, outperforming prior SoTA methods (typically 0.8-0.9) by a margin of 5-20%, while maintaining strong generalization to unseen distributions, establishing a scalable and robust solution for modern AI-generated video detection.

Hung Mai, Loi Dinh, Duc Hai Nguyen +6

N2TP Technology Solution JSC, Hanoi, Vietnam · College of Technology, National Economics University, Hanoi, Vietnam · University of Science, Vietnam National University, Ho Chi Minh City, Vietnam +3

Spatio-Temporal Attention Deepfake Detection Vision Transformer

Ask PDF

Abstract

Video Forensics

Explore similar work

Retrieval-Driven Training-Free AI-Generated Video Attribution

Training-free Detection of Generated Videos via Spatial-Temporal Likelihoods

EA-Swin: An Embedding-Agnostic Swin Transformer for AI-Generated Video Detection