Free-Lunch Long Video Generation via Layer-Adaptive O.O.D Correction

Jiahao Tian; Chenxi Song; Wei Cheng; Chi Zhang

cs.CVMar 26, 2026

Free-Lunch Long Video Generation via Layer-Adaptive O.O.D Correction

Authors: Jiahao Tian, Chenxi Song, Wei Cheng, Chi Zhang

Organizations: Westlake University

Abstract

Generating long videos using pre-trained video diffusion models, which are typically trained on short clips, presents a significant challenge. Directly applying these models for long-video inference often leads to a notable degradation in visual quality. This paper identifies that this issue primarily stems from two out-of-distribution (O.O.D) problems: frame-level relative position O.O.D and context-length O.O.D. To address these challenges, we propose FreeLOC, a novel training-free, layer-adaptive framework that introduces two core techniques: Video-based Relative Position Re-encoding (VRPR) for frame-level relative position O.O.D, a multi-granularity strategy that hierarchically re-encodes temporal relative positions to align with the model's pre-trained distribution, and Tiered Sparse Attention (TSA) for context-length O.O.D, which preserves both local detail and long-range dependencies by structuring attention density across different temporal scales. Crucially, we introduce a layer-adaptive probing mechanism that identifies the sensitivity of each transformer layer to these O.O.D issues, allowing for the selective and efficient application of our methods. Extensive experiments demonstrate that our approach significantly outperforms existing training-free methods, achieving state-of-the-art results in both temporal consistency and visual quality. Code is available at https://github.com/Westlake-AGI-Lab/FreeLOC.

Figures & tables

Figure 1: Overview of FreeLOC. FreeLOC employs an offline, layer-wise probing procedure to identify DiT layers sensitive to two O.O.D sources: context-length O.O.D and frame-level relative position O.O.D. For layers only sensitive to frame-level relative position, we apply VRPR strategy to hierarchically remap out-of-range positions back into the pre-trained domain. For layers sensitive to context-length, we utilize TSA combined with VRPR to balance local detail and global coherence. Zoom in for better view.

Figure 2: Sensitivity analysis for each layer during frame-level relative position O.O.D probing. (a) Vision Reward for each probing layer. A lower score indicates that the layer is more sensitive to frame-level relative position O.O.D. (b) Attention Logits Difference (ALD) for each probing layer. A higher value indicates a significant change in the behavior of the attention mechanism, signifying high sensitivity to positional O.O.D.

Figure 3: Layer-wise sensitivity to frame-level relative-position O.O.D. The figure compares the original video with probing outputs from Layer 18 (low sensitivity) and Layer 28 (high sensitivity). Stronger sensitivity leads to noticeable distortions when relative-position shifts are applied.

Figure 4: Layer-wise sensitivity to context-length O.O.D measured via attention entropy differences.

Model	#Frames	Method	Video Consistency			Video Quality
Model	#Frames	Method	Subject Consistency( $\uparrow$ )	Background Consistency( $\uparrow$ )	Motion Smoothness( $\uparrow$ )	Imaging Quality( $\uparrow$ )	Aesthetic Quality( $\uparrow$ )	Dynamic Degree( $\uparrow$ )
Wan2.1 -T2V-1.3B	161 (2 $\times$ )	Direct Sampling	97.33	97.23	98.69	60.34	52.72	19.31
		Sliding Window	96.48	96.22	98.50	64.64	54.42	40.82
		RIFLEx [ 40 ]	97.21	97.14	98.70	60.62	53.67	23.45
		FreeLong [ 21 ]	97.05	97.19	98.84	63.91	54.56	32.19
		FreeNoise [ 25 ]	96.82	97.10	98.82	67.19	56.01	36.34

Table 1: Quantitative results at the extended (2× and 4

\times

) video lengths. In a training-free manner, our method maintains the quality regarding video consistency and video quality in multiple VBench dimensions when the generation length grows.

Method	SC ( $\uparrow$ )	BC ( $\uparrow$ )	MS ( $\uparrow$ )	IQ ( $\uparrow$ )	AQ ( $\uparrow$ )	DD ( $\uparrow$ )
Direct	98.50	97.89	98.83	59.21	49.43	4.32
Direct+TSA	97.41	96.76	98.67	65.87	57.05	37.01
Direct+VRPR	98.42	97.81	98.89	61.88	54.13	15.32
Direct+ $\text{(TSA+VRPR)}_{\text{uniform}}$	97.56	97.67	98.75	65.19	56.34	34.44
Direct+ $\text{(TSA+VRPR, VRPR)}_{\text{random}}$	98.03	97.61	98.91	63.90	54.44	33.13
Direct+ $\text{(TSA, VRPR)}_{\text{layer-wise}}$	98.38	97.66	98.95	66.92	60.88	35.05

Table 2: Ablation study of layer-wise strategy and main components.

Figure 5: Qualitative Results of Wan2.1-1.3B and Hunyuan Video

Method	SC ( $\uparrow$ )	BC ( $\uparrow$ )	MS ( $\uparrow$ )	IQ ( $\uparrow$ )	AQ ( $\uparrow$ )	DD ( $\uparrow$ )
Clipping	97.83	97.51	98.78	60.23	52.03	22.21
Group	97.95	97.49	98.66	59.91	51.32	18.32
Our VRPR	98.44	97.78	98.97	68.84	61.21	36.27

Table 3: Ablation study of different relative position re-encoding strategy for frame-level relative position O.O.D.

Method	SC ( $\uparrow$ )	BC ( $\uparrow$ )	MS ( $\uparrow$ )	IQ ( $\uparrow$ )	AQ ( $\uparrow$ )	DD ( $\uparrow$ )
Sliding Window	97.52	96.72	98.69	64.33	54.92	37.81
Selected Frame Attention	97.89	98.94	98.94	63.90	54.19	28.53
Our TSA	98.44	97.78	98.97	68.84	61.21	36.27

Table 4: Ablation study of different attention mechanism for context-length O.O.D.

Figure 1: Layer-wise sensitivity to context-length O.O.D measured via attention entropy differences.

Figure 2: Impact of

W_{1}

W_{2}

G_{1}

and

G_{2}

on Subject Consistency and Imaging Quality for VRPR .

Figure 4: Qualitative comparison of different allocation strategy . The (1/2, 1/2) split proved to be the robust "sweet spot" which balance detail preservation and temporal consistency.

Parameter	Wan2.1-T2V-1.3B	Hunyuan Video
Sampler	Unipc	Euler
Denoising Steps	50	50
CFG Scale	6.0	6.0
Resolution	$832\times 480$ (480p)	$832\times 480$ (480p)
Base Length (1x)	81 frames	127 frames
2x Length	161 frames	253 frames

Table 1: Generation Settings. Basic settings used for Wan2.1-T2V-1.3B and Hunyuan Video for quantitative and qualitative evaluation.

Figure 5: Impact of

D_{1}

D_{2}

and

\alpha

on Subject Consistency and Imaging Quality .

Method	SC ( $\uparrow$ )	BC ( $\uparrow$ )	MS ( $\uparrow$ )	IQ ( $\uparrow$ )	AQ ( $\uparrow$ )	DD ( $\uparrow$ )
Mid Frame	98.14	97.65	98.91	67.12	59.83	34.65
Last Frame	98.19	97.68	98.87	67.48	60.98	35.72
w/o Attention Sink	98.05	97.67	98.92	67.53	60.92	34.59
Our First Frame	98.44	97.78	98.97	67.44	61.21	36.27

Table 2: Ablation study of impact of attention sink . Attention could significantly improve overall video consistency and video quality.

Method	SC ( $\uparrow$ )	BC ( $\uparrow$ )	MS ( $\uparrow$ )	IQ ( $\uparrow$ )	AQ ( $\uparrow$ )	DD ( $\uparrow$ )
LI	97.93	97.62	98.64	51.11	50.91	16.32
YARN	98.16	97.64	98.89	62.33	57.98	30.87
NTK-aware	97.89	97.59	98.82	56.47	53.54	24.21
Our VRPR	98.44	97.78	98.97	68.84	61.21	36.27

Table 3: VRPR vs. other RoPE scaling methods.

Method	SC ( $\uparrow$ )	BC ( $\uparrow$ )	MS ( $\uparrow$ )	IQ ( $\uparrow$ )	AQ ( $\uparrow$ )	DD ( $\uparrow$ )
2-stage	98.15	97.69	98.90	65.33	58.91	33.42
4-stage	98.49	97.93	98.95	67.91	60.90	35.97
Our 3-stage	9 8.44	97.78	98.97	67.44	61.21	36.27

Table 4: Ablation study of different granularity of VRPR.

Method	SC ( $\uparrow$ )	BC ( $\uparrow$ )	MS ( $\uparrow$ )	IQ ( $\uparrow$ )	AQ ( $\uparrow$ )	DD ( $\uparrow$ )
Reverse	98.02	97.28	98.87	62.12	57.12	31.23
Half Implementation 1	98.18	97.41	98.91	64.21	59.12	32.19
Half Implementation 2	98.22	97.48	98.94	62.39	60.91	32.92
Interleaved	98.12	97.32	98.89	63.20	59.92	33.59
FreeLOC	98.44	97.78	98.97	67.44	61.21	36.27

Table 5: Ablation study of more different layer-wise strategy.

Method	Inference Time (s/step)	Mem (GB)
Direct Sampling	33.93	29.34
Sliding Window	16.72	30.33
FreeNoise [ 25 ]	17.21	30.34
Freelong [ 21 ]	48.15	40.71
RIFLEX [ 40 ]	33.94	29.34
FreeLOC	24.35	29.87

Table 6: Comparison of inference time and peak GPU memory usage for generating 321-frame videos on the Wan2.1-T2V-1.3B model. All measurements were conducted on an NVIDIA A100 GPU.

Method	Content Consistency	Video Quality	Video-Text Alignment
Direct Sampling	3.27	1.02	1.92
Sliding Window	1.17	3.25	2.34
FreeNoise [ 25 ]	1.20	2.88	2.68
Freelong [ 21 ]	3.59	2.11	2.98
RIFLEX [ 40 ]	3.31	1.09	1.89
FreeLOC	4.11	3.95	3.78

Table 7: User study results . Participants scored videos on a scale of 1 (worst) to 5 (best) across three criteria. Our method received the highest ratings across all three metrics.

Figure 6: Qualitative Results of Wan2.1-1.3B-T2V with $4\times$ extension (321-frame)

Figure 7: Qualitative Results of Wan2.1-1.3B-T2V with $2\times$ extension (161-frame)

Figure 8: Qualitative Results of HunyuanVideo with $4\times$ extension (509-frame)

Figure 9: Qualitative Results of HunyuanVideo with $2\times$ extension (253-frame)

Long-Horizon Video Generation

Long-horizon video generation creates coherent, controllable videos over extended durations despite limited context and accumulating errors.

See all 164 papers

Diff-VF: Training-free High-quality Long Video Generation via Diffusion Model
Aug 6, 2026Haoning Yang, Xinyuan Chen, Yaohui Wang +1Diffusion Model Guidance Video Diffusion Models
Shanghai Jiao Tong University, Shanghai, China · Shanghai Artificial Intelligence Laboratory, Shanghai, China
FreeSpec: Training-Free Long Video Generation via Singular-Spectrum Reconstruction
May 7, 2026Fangda Chen, Shanshan Zhao, Longrong Yang +3Video Diffusion Models Long Video Generation
College of Computer Science and Technology, National University of Defense Technology · Alibaba International Digital Commerce · Zhejiang University +1
Enhancing Train-Free Infinite-Frame Generation for Consistent Long Videos
May 18, 2026X. Feng, J. Zhu, M. Wu +6Video Diffusion Models Long Video Generation
School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China · The Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences, Beijing, China · AMAP, Alibaba Group, Beijing, China.

Aug 6, 2026J/K move · Enter open · S save

Diff-VF: Training-free High-quality Long Video Generation via Diffusion Model

Haoning Yang, Xinyuan Chen, Yaohui Wang, Guo Lu

Shanghai Jiao Tong University, Shanghai, China · Shanghai Artificial Intelligence Laboratory, Shanghai, China

Diffusion Model Guidance Video Diffusion Models Long Video Generation Temporal Consistency in Video Generation

Open paper AskPDF arXiv

Recently, diffusion models have made great progress in video generation. However, most existing video diffusion models are trained with short videos, and degrade when extrapolated to long videos, struggling to maintain long-range temporal coherence while retaining diverse motions. To generate consistent, high-quality and dynamic long videos, we propose Diff-VF, a training-free, plug-and-play and model-agnostic framework that converts existing short-video diffusion backbones into long-video generators without modifying or fine-tuning the base model. Diff-VF couples three complementary strategies: Hybrid Noise Initialization (HNI) to constrain global semantics, Weighted Window Sampling (WWS) to remove inter-window discontinuities, and Temporal Extended Sampling (TES) to establish long-range dependencies with a timestep-varying fusion. We further extend Diff-VF to long-video enhancement via Skip Residual Guidance that balances fidelity and realism through timestep-dependent guidance. VBench-Long evaluation results show that Diff-VF achieves a more favorable balance between temporal coherence and motion diversity than base models and recent training-free long video generation baselines, including FreeNoise, FreeLong, and RIFLEx, while maintaining competitive frame-wise quality. Experiments on two base models demonstrate the applicability to video diffusion models with different spatial-temporal modeling strategies. Extensive ablations validate the contribution of each component and hyperparameters.

Abstract

Long-Horizon Video Generation

Explore similar work

Diff-VF: Training-free High-quality Long Video Generation via Diffusion Model

FreeSpec: Training-Free Long Video Generation via Singular-Spectrum Reconstruction

Enhancing Train-Free Infinite-Frame Generation for Consistent Long Videos