Hierarchical Diffusion for Camera-controlled Autoregressive Video Generation

Introduction

Recent advances in video diffusion models Blattmann et al. [4]; Yang et al. [59]; Team et al. [47]; HaCohen et al. [20] have enabled photorealistic video generation with increasing levels of control. In this work, we study the problem of long-video generation from a single image given future camera trajectories. This setting is particularly appealing for applications such as virtual scene exploration and content creation, where controllable viewpoint changes must remain 3D consistent over extended time horizons. However, generating long videos that faithfully follow camera motion while maintaining spatial and temporal coherence remains a significant challenge.

Why is this hard? Existing approaches predominantly rely on autoregressive generation, where future frames are conditioned on previously generated outputs over a sequence of rollouts. While this is effective for short sequences, such methods accumulate errors over time, leading to drift in long videos. A second challenge lies in how camera motion is incorporated. One widely adopted approach is to condition on explicit 6-DOF pose representations, such as Plücker coordinates or ray maps. One subtle challenge is that many flagship benchmarks (like RealEstate10K Zhou et al. [68]) provide cameras only up-to-scale, making camera-conditioned video generation somewhat undefined - e.g., the world looks different when moving forward 10cm versus 10m.

Status quo. Prior work has attempted to address these challenges along two main directions. To mitigate autoregressive drift, several methods adopt self-forcing strategies Huang et al. [25], where models are periodically trained or evaluated on their own generated outputs. While effective in reducing error accumulation, these approaches significantly increase computational cost and are difficult to scale to long sequences; there is a linear increase in time and space for generating training data as rollouts can only be sequentially obtained. Most camera-controlled video generation methods appear to circumvent the up-to-scale ambiguity by overfitting to the scale of a particular dataset, making such camera-conditioned generation models unlikely to generalize.

Vector (PDF)

Our approach. We address these challenges with a simple and lightweight framework that combines online reconstruction with hierarchical denoising. First, to respect the fixed but unknown scale of input camera trajectories, we draw inspiration from feed-forward novel-view synthesis and reformulate video generation as a warping-and-inpainting problem, aligning intermediate depth predictions to the scale of the ground-truth geometry (typically provided in the same scale as camera inputs, often obtained via COLMAP). Second, to mitigate autoregressive drift, we propose a hierarchical denoising strategy in which different stages condition on different temporal histories. Concretely, we generate frames in a fine-to-coarse manner: an initial denoiser conditions on the immediate past to produce high-frequency details, while a second denoiser conditions on a temporally subsampled long-range history to enforce global consistency. This design allows each generated frame to efficiently incorporate information from the entire past without incurring the cost of conditioning on an exhaustive set of historical frames.

In summary, our contributions are as follows: (1) we identify scale inconsistency as a key failure mode of pose-conditioned video generation and propose a warping-based formulation that enforces scale consistency between the scene and camera trajectory; (2) we introduce a hierarchical denoising scheme that reduces long-horizon drift by conditioning on multi-scale temporal context; and (3) we present a simple combination of models, WanAR and FixAR, that achieves state-of-the-art performance on RealEstate10k and DL3DV for camera-controlled long-video generation, producing videos that remain consistent with the scale of input camera trajectories.

Related Work

Autoregressive long video generation. Modern video diffusion models Ho et al. [23]; Blattmann et al. [4]; Ha et al. (2024); Yang et al. [59]; Team et al. [47]; Chen et al. [10] can generate photorealistic videos, but are typically limited to short temporal horizons. To extend generation length, autoregressive formulations are widely adopted to roll out longer sequences Chen et al. [8]; Teng et al. [48]; Deng et al. [13]; Jin et al. [27]; Chen et al. [7]; Yin et al. [60]; Song et al. [44]; Huang et al. [25]; Cui et al. [11]. However, a fundamental challenge in autoregressive video diffusion is drift, which progressively degrades visual quality over long rollouts. Recent approaches attempt to mitigate this issue by either exposing models to drifted distributions during training [7, 44, 25, 37] (); Zhu et al. 2026), or by modifying attention mechanisms and introducing memory compression modules to reference earlier, cleaner context during inference [65, 6, 66, 58, 31]. While these methods improve long-range consistency, they still suffer from limitations. For example, diffusion-forcing [7, 44] relies on approximated drift distributions and exhibits a train–test drift distribution mismatch, while self-forcing [25] incurs increased training cost due to multi-step rollouts at every training iteration. In this work, we propose to decouple autoregressive generation (WanAR) and drift correction (FixAR) via a hierarchical denoising framework, where FixAR is trained on precomputed data pairs with real drift.

Camera-conditioned video world model. By introducing action controllability into long video generation and making the underlying world behind visual content explorable, video world models have emerged and demonstrated strong performance [5, 3, 49, 45, 24, 46]. Most video world models primarily used keyboard inputs as a simple form of action control [22, 19, 15], while many recent works [61, 57, 28, 33, 69, 56, 17] increasingly emphasizes camera motion as the control signal to enable more precise manipulation and support specific application scenarios. A natural question is how to incorporate camera information into video generative models. Existing approaches can be broadly categorized into three groups: (i) direct camera encoding, (ii) injecting 3D geometry through positional encoding fields [32, 2], and (iii) injecting 3D geometry via rendering in pixel space. Most existing methods Yu et al. (2025a); Xiao et al. (2025); Shen et al. (2026a); Li et al. (2025b); Team et al. (2026b;a); Hong et al. (2025); He et al. (2025b); Hong et al. (2025); Yu et al. (2025a) fall into the first category, where normalized representations (e.g., Plücker rays or ray maps) are encoded into latent space to guide the generation process, allowing camera control to be learned from large-scale camera annotations. However, such latent-space conditioning often leads to imprecise control when reflected in pixel-space generation, and tends to generalize poorly to out-of-distribution camera trajectories. The second line of work [42, 64] models 3D geometric transformations induced by camera motion through modifications to positional encoding. This approach achieves more accurate camera control, but the modified architecture introduces additional training overhead. In this work, we adopt the third line of work [56, 28], which directly models camera control by estimating 3D geometry using a metric depth estimator and warping it into future frames in an online fashion. This approach enables accurate camera control while naturally reformulating autoregressive generation as an inpainting task.

Generative novel-view synthesis. Current state-of-the-art novel-view synthesis (NVS) methods are largely based on generative diffusion models Gao et al. (2024); Wu et al. (2025b); Watson et al. (2022); Yu et al. (2025c;b); Chen et al. (2025b); Bai et al. (2025a); Ren et al. (2025); Li et al. (2026a); Bai et al. (2025b); Seo et al. (2024); Zhao et al. (2024). They are closely related to camera-conditioned video world models, but their spatial and temporal horizons for novel-view generation remain limited. Early NVS approaches also widely adopted direct camera encoding in latent space [18, 55], but lacked precise camera control. More recent methods instead employ 3D geometry-based warping-and-inpainting formulations Yu et al. (2025c;b); Chen et al. (2025b); Ren et al. (2025), leading to improved visual quality and more accurate camera control.

In addition, there is a line of work in NVS that uses diffusion models to correct rendering artifacts caused by large viewpoint changes [54] (); de Lутio et al. 2026). This perspective is closely related to the motivation behind our FixAR. In our setting, we interpret autoregressive drift in long-horizon video generation as another form of accumulated artifact over time, which can be corrected using video diffusion.

Vector (PDF)

Method

Preliminaries: Flow matching video diffusion

We adopt the open-source DiT-based Peebles & Xie [39] video diffusion model Wan-I2V-14B Team et al. [47] as our base model. Unlike traditional diffusion models that rely on score matching and stochastic reverse diffusion, Wan is trained with rectified flow matching [14, 38, 36], which formulates generation as solving an ordinary differential equation (ODE) without stochastic terms. This formulation provides a stable and scalable training objective while preserving a likelihood-based interpretation.

Specifically, given a clean video latent $x_1$, a random noise latent $x_0 \sim\mathcal{N}(0,I)$, a conditioning signal $c$, and a timestep $t \in[0,1]$, flow matching defines the interpolated latent as

The model is trained to predict the constant velocity field that transports noise $x_0$ to data $x_1$:

WanAR: Camera control for autoregressive video generation

We consider the problem of camera-conditioned long video generation: given an input image $I_0$ and a target camera trajectory $\{C_t\}_{t=0}^{T}$, the goal is to generate future observations $\{I_t\}_{t=1}^{T}$ that follow the specified camera motion, where $T$ can be large. General-purpose video diffusion models such as Wan-I2V-14B cannot be directly applied to this setting, because they (i) do not explicitly support camera conditioning, and (ii) are limited to a short generation horizon $L$, where typically $L \ll T$. To address this, we build WanAR on top of Wan. WanAR achieves camera control by propagating explicit 3D geometry (depth and camera), and extends generation length through an autoregressive formulation.

Given the initial observation $I_0$, we first estimate its metric depth $D_0$ using the monocular depth estimator MoGe-2 Wang et al. [51]. Since the corresponding camera $C_0$ is known, this gives us sufficient 3D geometry for the first frame. We then propagate this information to future viewpoints within the base model horizon by forward warping:

where $\widetilde{I}_t$ denotes the warped partial observation at time $t$, $M_t$ is the corresponding inpainting mask, and $\mathcal{W}$ is the warping function.

Although the warped frames contain only partial RGB observations, they already provide accurate camera-conditioned spatial guidance. As a result, camera-conditioned video generation is reformulated as an inpainting problem that completes these partial observations:

where $\widehat{I}_t$ denotes the generated full frame. To adapt Wan-I2V to this task, we modify its conditioning mechanism. In the original I2V model, the future frames and masks are represented by zero-padded latents in VAE space. In contrast, we replace this zero padding with the actual latent encodings of the warped frames and their corresponding inpainting masks.

So far, this formulation enables camera control only within the base diffusion horizon. To extend it beyond length $L$, we apply the same inpainting process autoregressively. At rollout $k$, we take the latest anchor frame $\widehat{I}_{k(L-1)}$, estimate its depth online using MoGe-2 Wang et al. [51], and warp it to the next chunk of target cameras. Formally, for

we generate

In this way, WanAR produces a long video sequence that follows the entire target camera trajectory.

FixAR: Hierarchical denoising for mitigating drift

The autoregressive formulation of WanAR enables long-horizon generation, but it also introduces drift: except for the first rollout, each new chunk is conditioned on previously synthesized frames, so errors accumulate over time and gradually degrade visual quality. To mitigate this effect, we introduce FixAR, a second-stage fixing module that is applied after every few WanAR rollouts.

Specifically, after every $K$ rollouts, we uniformly sample a sparse set of keyframes from the generated sequence and feed them into FixAR for content correction. Let $\{\widetilde{I}_{T_j}\}_{j=0}^{L-1}$ denote the selected sparse keyframes. FixAR outputs the corrected frames

where $g_\psi$ denotes the FixAR model and $I^*_{T_j}$ are the corrected keyframes.

To train FixAR on realistic drift patterns, we construct its training set by following the same WanAR rollout process and the same sparse keyframe sampling strategy used at inference time. Unlike self-forcing approaches Huang et al. [25], FixAR training does not require multi-step inference during each training iteration; instead, it uses precomputed input-target pairs, which substantially reduces computational overhead. Once trained, FixAR can be applied either globally to an entire generated sequence or autoregressively at a low frequency.

FixAR mitigates drift for two reasons. First, due to its hierarchical design, it can directly reference earlier frames with higher visual quality, or even ground-truth frames during training, allowing appearance and style to be propagated over a longer horizon. Second, because it is trained on realistic drift distributions, it learns to identify artifact-prone regions and correct them accordingly.

Recovering the original video framerate. Because of the limited base model horizon and for computational efficiency, FixAR operates only on sparse keyframes. To recover the full video sequence at original framerate, we train an additional video diffusion model, InterpolatAR, to interpolate between neighboring corrected keyframes. This is a substantially easier task, since adjacent corrected keyframes are already visually and temporally close, and little hallucination is required. Importantly, InterpolatAR does not introduce additional drift, because it only fills in intermediate content rather than performing autoregressive extrapolation as in WanAR.

Empirical Analysis

Experimental setup

Datasets. We train and evaluate our model on two real-world datasets, RealEstate10K Zhou et al. [68] and DL3DV Ling et al. [35], which cover diverse indoor and outdoor scenes. RealEstate10K Zhou et al. [68] contains approximately 60K videos, with ground-truth camera annotations available for both training and testing. For DL3DV Ling et al. [35], since the released version does not provide camera annotations, we use a COLMAP-preprocessed training set Wang et al. [50] with approximately 7K videos. For the test set, we process the videos with MegaSAM Li et al. [34] to obtain the corresponding camera poses and depth maps.

Baselines. We compare our method with baselines from both camera-conditioned long video generation and novel-view synthesis. For camera-conditioned video generation, we compare with VMem Li et al. [33], DFoT Song et al. [44], Lingbot-World Team et al. [46], and WorldWarp Kong et al. [28]. VMem, DFoT, and Lingbot-World use Plucker rays or ray maps as camera-control representations, while WorldWarp uses 3D Gaussian rendering as camera guidance. Lingbot-World is an open-source industrial model, and we use its flagship bidirectional model for evaluation.

For novel-view synthesis, we adopt video diffusion-based methods that also use warp-and-inpaint-style camera control, including ViewCrafter Yu et al. [63], TrajCrafter Yu et al. [62], and CogNVS Chen et al. [9]. We further include VACE Jiang et al. [26], a Wan-14B-based video inpainter, as an additional baseline. Since these methods are designed for fixed-length rendering, we evaluate them in an image-to-video autoregressive manner, following the same rollout setting as our method.

Metrics. We follow standard evaluation protocols used in related work [57, 28, 33, 43], including pairwise reconstruction metrics PSNR, SSIM, and LPIPS, as well as distribution-level metrics FID and KID. We also report FVD to evaluate video generation quality. In addition, we adopt masked PSNR (mPSNR) Gao et al. [16], which computes PSNR only over the covisible region from the first frame. This metric better reflects camera-control accuracy without scale ambiguity.

Since some baselines [44, 33] only support a resolution of $256 \times256$ and a temporal frame skip of 10, we crop and resize the outputs of all methods, including ours, to the same spatial and temporal resolution for fair comparison.

Implementation details. We initialize our models from the official Wan-I2V-14B-480P checkpoint Team et al. [47] and fine-tune them using LoRA with rank 32. We use a batch size of 8 and a learning rate of $1 \times10^{-4}$. For each model, including WanAR, FixAR, and InterpolatAR, we fine-tune for 10K steps on each dataset using 8 H100 GPUs at a resolution of $256 \times256$. Training each model takes approximately one day. Further inference details are provided in Appendix 1.1.

Comparison to state-of-the-art

RealEstate10K. Table 1 presents a detailed comparison between existing methods and ours on RealEstate10K. Overall, NVS methods achieve stronger reconstruction metrics, including PSNR, SSIM, LPIPS, and mPSNR, mainly because their warp-and-inpaint formulation provides more accurate camera control. In contrast, camera-conditioned long video generation methods perform better on visual fidelity metrics, such as FID, KID, and FVD, since they are typically trained with dedicated anti-drift strategies and are therefore better at preserving long-term visual quality. The importance of accurate camera control is also illustrated in Figure 3. For example, Lingbot-World, despite being an industrial-level bidirectional video diffusion model capable of generating highly photorealistic videos, still obtains relatively poor PSNR due to inaccurate camera control. This suggests that

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$	FVD $\downarrow$	mPSNR $\uparrow$
ViewCrafter Yu et al. [63]	12.45	0.361	0.499	67.36	0.0267	595.25	13.44
TrajCrafter Yu et al. [62]	14.35	0.428	0.410	30.54	0.0042	347.75	15.54
CogNVS Chen et al. [9]	13.22	0.379	0.474	48.58	0.0116	457.35	14.39
VACE Jiang et al. [26]	15.23	0.487	0.426	49.62	0.0119	393.30	16.89
Lingbot-World Team et al. [46]	12.90	0.362	0.513	25.46	0.0012	398.31	13.36
VMem Li et al. [33]	12.13	0.336	0.583	49.12	0.0099	465.13	12.46
DFoT Song et al. [44]	13.95	0.423	0.422	38.28	0.0084	387.21	15.17
WorldWarp Kong et al. [28]	13.81	0.425	0.475	38.59	0.0042	451.01	14.55
WanAR + FixAR	16.69	0.525	0.341	23.16	0.0018	282.59	18.31

Plücker-ray-based conditioning is not sufficiently accurate or generalizable for this setting. In contrast, our method achieves state-of-the-art performance across major metrics by combining accurate 3D-informed warp-and-inpaint camera control with hierarchical denoising for drift mitigation.

DL3DV is a more challenging dataset, containing fast viewpoint transitions, extreme camera trajectories, and diverse scene content. As shown in Figure 3, many NVS methods suffer from distorted content, largely because depth estimation becomes unreliable on drifted generated frames during autoregressive rollout. In contrast, our method uses FixAR to suppress drift accumulation, ensuring that the corrected frames remain meaningful inputs for subsequent depth estimation and warping. As reported in Table 2, our method again achieves state-of-the-art overall performance, despite marginally lower FID and KID compared with Lingbot-World. These results demonstrate that our method generalizes well to diverse indoor and outdoor scenarios with challenging camera motion.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$	FVD $\downarrow$	mPSNR $\uparrow$
ViewCrafter Yu et al. [63]	13.55	0.307	0.459	74.76	0.0114	847.25	13.19
TrajCrafter Yu et al. [62]	13.49	0.306	0.428	53.45	0.0042	703.82	13.93
CogNVS Chen et al. [9]	12.70	0.268	0.467	81.09	0.0166	882.71	13.04
VACE Jiang et al. [26]	13.26	0.302	0.476	92.48	0.0209	781.41	13.77
Lingbot-World Team et al. [46]	12.87	0.246	0.457	40.26	0.0013	742.09	13.07
VMem Li et al. [33]	12.02	0.232	0.563	80.76	0.0138	867.67	12.13
DFoT Song et al. [44]	9.19	0.143	0.603	82.44	0.0232	931.15	9.23
WorldWarp Kong et al. [28]	12.63	0.265	0.540	74.81	0.0072	1002.4	12.73
WanAR + FixAR	13.84	0.320	0.421	47.75	0.0034	640.64	14.05

Ablations and analysis

Effect of hierarchical denoising. We conduct ablation studies on a 50-sequence subset of RealEstate-10K. Figure 4 and Table 3 show the qualitative and quantitative effects of the hierarchical denoiser. By mitigating drift, the denoiser improves generation realism and consistently improves the evaluation metrics.

Method	PSNR ↑	SSIM ↑	LPIPS ↓	FID ↓	KID ↓	FVD ↓
WanAR	16.63	0.519	0.342	36.68	0.0022	455.87
WanAR + FixAR	16.71	0.521	0.338	34.55	0.0018	429.14

Ablations on camera pose conditioning. We compare WanAR’s warp-and-inpaint conditioning with Plucker ray conditioning using the same backbone and training setup. As shown in Table 4, warp-and-inpaint improves reconstruction quality, camera-control accuracy, and video quality, while achieving comparable FID and KID. This suggests that warp-and-inpaint provides more precise camera control in our setting. Additional ablations are reported in Appendix .§ 3.

Method	PSNR ↑	SSIM ↑	LPIPS ↓	FID ↓	KID ↓	FVD ↓	mPSNR ↑
Warp-and-inpaint	16.63	0.519	0.342	36.68	0.0022	455.87	18.20
Plucker ray	14.95	0.438	0.416	35.62	0.0021	505.49	15.33

Evaluation of multi-modal generation. A key advantage of video diffusion models is their multi-modal generation capability, which we also observe in WanAR + FixAR. As shown in Figure 5, the same corridor scene can be plausibly generated as either a narrow straight pathway or a larger room space connected to another room by an additional door. Both outcomes are reasonable under the given input. We therefore evaluate Top-$K$ generation metrics by sampling multiple outputs with different random seeds.

Vector (PDF)

Discussion

We present a simple framework for long-video, camera-controlled generation that combines autoregressive generation with hierarchical denoising. Our results show that reformulating generation as a warping-and-inpainting problem helps maintain consistency with input camera trajectories, particularly in the presence of up-to-scale ambiguity, while hierarchical denoising improves temporal stability by allowing each frame to incorporate both short- and long-range temporal information. Compared to prior approaches that rely on self-forcing or explicit pose conditioning alone, our method achieves stronger long-horizon consistency. These findings suggest that both scale consistency and long-range temporal conditioning are important for generating stable, camera-controlled videos over extended sequences.

Limitations

We note that our proposed warp-and-inpaint hierarchical video generation can only be used for static scenes. WanAR relies on the quality of underlying geometry, and errors in depth or reconstruction can propagate through the warping process, leading to artifacts or inconsistencies. WanAR rollouts only warp the last frame of the previous rollout and therefore do not have any memory mechanism baked in. This means that loop closure in long sequences may not be accurate. While hierarchical denoising with FixAR reduces drift, it does not fully eliminate it as our second stage operates over a finite horizon of frames too. Finally, our models trained on one dataset cannot be used for another dataset if the scale distributions between the two datasets are drastically different. We cover a detailed inference cost of our method with respect to other baselines in appendix.

Broader Impacts

Efficient and temporally consistent camera-controlled video synthesis could broaden access to tools for creative media production, 3D exploration in virtual reality applications, and embodied AI research, though care should be taken to mitigate risks associated with realistic synthetic video generation.

Appendix

In this supplement, we include further implementation details of implementation (Section .1), computational cost (.2), and more ablations .3 of our WanAR + FixAR approach. We further present a nuanced discussion of camera control in video generation (Section .4).

More implementation details

Given one initial image and a sequence of $L$ target cameras, our objective is to generate an $L$-frame video following the target camera trajectory. WanAR, FixAR, and InterpolatAR all use the same Wan2.1-14B architecture, but have separately trained checkpoints for their respective tasks. Unless otherwise specified, each stage operates at a resolution of $336 \times592$ with a temporal window of $T=49$ frames.

For InterpolatAR, we place the corrected FixAR keyframes at their corresponding temporal locations and replace the missing intermediate frames with blank frames. The first frame of each input window is always an observed keyframe. InterpolatAR then predicts the missing frames conditioned on the visible keyframes (i.e. if FixAR was applied at a keyframe interval of 4 then frames 1, 5, 9 . . . are input to InterpolatAR which interpolates to get frames 2, 3, 4, 6, 7, 8 . . . ).

For RealEstate10K, we generate ($L=280$) frames at 10 FPS: 1) WanAR: We subsample the target trajectory using an interval of 2 to generate the sequence with 3 autoregressive diffusion calls. 2) FixAR: We select 49 generated keyframes with an average temporal interval of approximately 6 and jointly refine them using 1 diffusion call. 3) InterpolatAR: We use an interval of 1 to reconstruct the complete 280-frame sequence, requiring 6 diffusion calls. The complete RealEstate10K pipeline uses 10 diffusion calls: 3 WanAR, 1 FixAR, and 6 InterpolatAR calls.

For DL3DV, we generate ($L=600$) frames at 30 FPS: 1) WanAR and FixAR: We subsample the target trajectory using an interval of 3 to generate the first 10s of the sequence with 2 WanAR diffusion calls. We then select 49 keyframes with an average temporal interval of 6 and jointly refine them using 1 FixAR diffusion call. Starting from the refined result, we repeat the same process with another 2 WanAR calls followed by 1 FixAR call to generate the remaining sequence. 2) InterpolatAR: We use an interval of 1 to reconstruct the complete 600-frame sequence, requiring 12 diffusion calls. The complete DL3DV pipeline uses 18 diffusion calls: 4 WanAR, 2 FixAR, and 12 InterpolatAR calls.

Analysis of computational cost

We compare the computational cost of our method with the baselines. Specifically, we run inference on 280-frame RealEstate10K videos using one 80GB H100 GPU and report the average runtime, peak GPU memory, number of diffusion calls, and denoising steps per call. As shown in Tab. 5, our method has moderate runtime and GPU memory usage. We additionally report our method at $480 \times832 \times49$ only for a resolution-matched cost comparison; all experiments in the paper use a training resolution of $336 \times592 \times49$.

Although our system contains three stages, all stages reuse the same Wan2.1-14B architecture and the same input shape of $336 \times592 \times49$. We sequentially reload the corresponding LoRA checkpoint for each stage without introducing additional architectural modules. This unified design reduces engineering complexity. Moreover, our hierarchical design applies FixAR at a low frequency before final interpolation, limiting the total number of diffusion calls.

Additional ablations

Sensitivity of FixAR’s keyframe interval. Our default FixAR processes 49 frames, is trained with keyframe intervals ranging from 2 to 6, and uses an interval of 6 during inference. We additionally train a 25-frame FixAR with keyframe intervals ranging from 4 to 12 to study the effects of input length and larger keyframe intervals. As shown in Table 6, the 49-frame FixAR performs consistently across intervals from 2 to 6, indicating that it is not sensitive to the precise interval within its training range. In contrast, the performance of the 25-frame FixAR decreases when the interval increases from 4 to 8 or 12. With very sparse keyframes, the temporal correspondence between

Setting	Keyframe Interval	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$	FVD $\downarrow$
49-length FixAR	2	16.38	0.508	0.336	34.33	0.0021	424.94
	4	16.78	0.522	0.333	34.06	0.0017	428.02
	6	16.71	0.521	0.338	34.55	0.0018	429.14
25-length FixAR	4	16.11	0.490	0.340	38.02	0.0021	461.49
	8	15.91	0.487	0.351	38.72	0.0022	488.46
	12	15.46	0.471	0.384	40.07	0.0030	561.53

caption: tbl:5: Computational cost comparison with baselines. Runtime and peak GPU memory are measured on 280-frame RealEstate10K videos using a single 80GB H100 GPU. We also report our method at $480 \times832 \times49$ for a resolution-matched comparison.

:::

adjacent input frames becomes weaker, making refinement more difficult. The 49-frame FixAR also consistently outperforms the 25-frame variant; a longer input sequence can cover a longer temporal range while retaining relatively dense keyframes.

Effect of different depth estimators. We estimate the depth of the first frame using each method and then perform one WanAR rollout. As shown in Table 7, the performance differences among the three depth estimators is relatively small, indicating that WanAR is robust to the choice of depth estimator – it can recover from errors introduced by depth estimators. Depth Anything V3 and MegaSAM depth provide slightly better reconstruction and video-quality metrics.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$	FVD $\downarrow$
MoGe2	20.22	0.654	0.177	21.90	0.0013	130.58
Depth Anything v3	20.46	0.666	0.172	21.77	0.0014	129.02
MegaSAM	20.74	0.679	0.170	21.92	0.0015	125.77

Extended discussion on camera-controlled video generation

In this section, we elucidate our claim that camera-controlled video generation is a special case of novel-view synthesis (NVS) and discuss relevant paradigms for evaluation which are inspired by the literature in NVS.

Case I: Image-to-video generation with pose / geometry information

When input observations are accompanied by reconstruction information, such as calibrated camera poses and/or 3D geometry defined in a fixed coordinate system, future camera trajectories should also be specified in that same coordinate system. This corresponds to the standard NeRF-style setting, where methods are conditioned on input images together with calibrated camera poses. More generally, this setup may also be instantiated using a single image paired with depth or XYZ information. For example, given an estimate of scene geometry, future camera motion can be defined relative to the reconstructed scene structure (e.g., moving a fixed distance toward an observed object). Framing camera control in this coordinate-aware manner removes the need to reference future ground-truth geometry during alignment and provides a cleaner formulation of camera-conditioned video generation.

Case II: Image-to-Video Generation Without Pose Information

In contrast, when only input images are provided without associated poses or geometric information of the input (such as the current setup of camera-controlled video generation works including ours), the problem more closely resembles unposed novel-view synthesis. In this setting, one possible approach is to require future camera trajectories to be specified in a metric coordinate system to avoid key pitfalls of prior art (c.f. Figure 6). An alternative is to follow protocols of recent benchmarks Kuo et al. [29] under whose formulation: (i) methods reconstruct the input observations within a locally consistent coordinate system; (ii) the benchmark provides a similarity transform aligning the reconstruction to the ground-truth frame; and (iii) future camera queries are specified within the local reconstructed coordinate system. Our current training, inference and benchmarking protocol is based on this formulation.