MV-Forcing: Long Multi-View Video Diffusion

• MV-Forcing is a multi-view video generation framework that uses a persistent 4D reconstruction state to couple temporal and view-wise autoregression.
• It integrates a 4D geometric bridge with joint denoising regimes, Distribution Matching Distillation, and Spatio-Temporal Self-Forcing to ensure cross-view and temporal consistency.
• Experimental results show that MV-Forcing outperforms baselines in camera accuracy, synchronization, and visual quality across long, unbounded multi-view video sequences.

MV-Forcing is a framework for long multi-view video generation that composes temporal and view-wise autoregression within a single diffusion model by introducing a 4D geometric bridge between sequentially generated views. Its central claim is that an autoregressive 3D reconstruction model can naturally interface between autoregressively generated views: after one source view is completed, its 3D structure is reconstructed and rendered as a geometric prior for the next target viewpoint, and the diffusion model refines that prior into a high-quality video. The framework further combines a joint denoising regime with Distribution Matching Distillation and Spatio-Temporal Self-Forcing so that generation can extend beyond a teacher’s fixed temporal window and remain geometrically consistent across arbitrary lengths and viewpoint counts (Fiebelman et al., 6 Jul 2026).

1. Problem formulation and design objective

MV-Forcing is motivated by a specific gap in video generation. Recent video diffusion models have enabled either long single-view generation through temporal autoregression, or short multi-view synthesis through bidirectional attention, but not long, multi-view consistent videos of dynamic scenes. The reported obstacles are structural: attention-based multi-view models are limited to short, fixed temporal windows because of the quadratic scaling of all-to-all attention across time and views, while autoregressive single-view temporal models lack geometric priors, so autoregressive view sequencing accumulates drift and loses 3D consistency (Fiebelman et al., 6 Jul 2026).

The framework therefore treats long-horizon temporal coherence and cross-view geometric consistency as a joint problem rather than as two separate extensions. A common misconception is that multi-camera generation can be obtained by simply adding view sequencing to a single-view autoregressive model. The reported failure mode is precisely that such sequencing, without an explicit geometric bridge, produces temporal or cross-view inconsistencies as fixed-length windows are chained together. MV-Forcing is organized around the opposite premise: temporal recursion and view recursion must be coupled through a persistent geometric state (Fiebelman et al., 6 Jul 2026).

2. Core architecture and the 4D geometric bridge

The architecture consists of a base video diffusion model, an autoregressive 4D reconstruction model, a geometric prior conditioning path, joint autoregression over time and view, and Distribution Matching Distillation with Spatio-Temporal Self-Forcing. The reconstruction module is CUT3R, described as a dynamic, recurrent 3D reconstructor that incrementally builds and queries a persistent scene state as views and frames are generated (Fiebelman et al., 6 Jul 2026).

At each autoregressive view step, decoded frames from previous views are integrated into CUT3R, which maintains a persistent recurrent state $\mathcal{S}$. The state update and query are given as

$$\mathcal{S}' = \mathrm{CUT3R\_update}(\mathcal{S}, I)$$

and

$$\hat{I}, \hat{C} = \mathrm{CUT3R\_query}(\mathcal{S}, \mathcal{R}),$$

where $\mathcal{R}$ encodes the camera as a raymap. The accumulated state is rendered from the target camera to yield a geometric prior consisting of an image and a per-pixel confidence map, and this prior is encoded and injected into the diffusion model as

$$g_k = \text{Concat}\left(\mathcal{E}(\hat{V}_k), \hat{C}_k\right).$$

The conditioning tensor is then added to the latent features through a 3D convolutional path (Fiebelman et al., 6 Jul 2026).

This 4D grounding is not merely a geometric initialization. The paper frames it as a continuous bridge between sequentially generated views, allowing each target view to be conditioned on explicit accumulated structure rather than only on latent history. This suggests that the framework’s notion of “4D” is operational rather than purely representational: geometry is maintained across both spatial viewpoint changes and temporal progression (Fiebelman et al., 6 Jul 2026).

3. Temporal autoregression, view-wise autoregression, and synchronization

MV-Forcing composes two causal processes. Along the temporal axis, future frames are generated by conditioning on past frames using autoregressive video diffusion. Along the view axis, each new camera view is generated sequentially and is conditioned on the fully denoised output of the previous view. The crucial addition is that view-wise conditioning is not only token-based; it is geometrically grounded by the CUT3R reconstruction state (Fiebelman et al., 6 Jul 2026).

Cross-view communication inside the transformer is handled by the Multi-View Synchronization module. At each transformer block, it adds camera embeddings and enables cross-view self-attention between feature tokens of the current and previous view at corresponding timesteps: $$\overline{F}_1^i, \overline{F}_2^i = \mathrm{MVS}\!\left(F_1^i + \mathcal{E}_c(\mathrm{cam}_1),\; F_2^i + \mathcal{E}_c(\mathrm{cam}_2)\right).$$ This camera-aware synchronization is presented as part of the mechanism that preserves scene structure and appearance across multiple viewpoints (Fiebelman et al., 6 Jul 2026).

The generation order is also described as flexible. Inference can proceed by time within a view or advance to a new view at any time, because the model combines a KV cache for temporal conditioning with a persistent scene state and explicit geometric priors for view transitions. The intended outcome is not only consistent novel-view synthesis, but arbitrary-order generation of long multi-view videos (Fiebelman et al., 6 Jul 2026).

4. Joint denoising, distillation, and Spatio-Temporal Self-Forcing

A central training issue is exposure bias. Autoregressive models are typically trained with clean ground-truth context but must condition on their own imperfect outputs at inference, which induces drift and instability over long sequences or view chains. MV-Forcing addresses this with two coupled mechanisms: a joint denoising regime and Distribution Matching Distillation with Spatio-Temporal Self-Forcing (Fiebelman et al., 6 Jul 2026).

The joint denoising regime is designed to unify text-to-video generation for the first view and view-conditioned generation for subsequent views in one model. During training, both view slots are randomly initialized from noise with some probability, rather than only the future slot. This is reported to enable text-to-video generation for the first view while also supporting view-conditioned generation for later views, and to extend generation beyond the teacher’s fixed temporal window (Fiebelman et al., 6 Jul 2026).

Distillation is performed from a bidirectional teacher such as SynCamMaster or ReCamMaster to a causal student using Distribution Matching Distillation. The reported loss is

$$\nabla_\phi \mathcal{L}_\mathrm{DMD} \approx -\mathbb{E}_{\tau}\left[\left(s_{\text{data}}(\hat{z}_\tau, \tau) - s_{\text{gen}}(\hat{z}_\tau, \tau)\right) \frac{\partial \hat{z}_\tau}{\partial \phi}\right].$$

Spatio-Temporal Self-Forcing then unrolls autoregressive generation along both temporal and view axes during training, so the student is scored on its own self-generated outputs rather than only on teacher-forced contexts. The reported effect is to close the train-inference exposure bias gap for both temporal and view-sequential autoregression (Fiebelman et al., 6 Jul 2026).

An important implication is methodological. The framework does not treat long-horizon generation as a pure scaling problem; it treats it as a distribution-shift problem created by recursive conditioning. This suggests that “self-forcing” here refers to deliberate training on self-generated contexts, not merely to self-conditioning in the narrower diffusion sense (Fiebelman et al., 6 Jul 2026).

5. Experimental results, scaling behavior, and ablations

The evaluation uses metrics for visual quality, camera accuracy, and cross-view synchronization: FID, FVD, CLIP-T, CLIP-F, RotErr, TransErr, Mat. Pix., CLIP-V, and FVD-V. On short sequences with $2$ views and $81$ frames, the reported comparison is:

Method	FID $\downarrow$	FVD $$\mathcal{S}' = \mathrm{CUT3R\_update}(\mathcal{S}, I)$$0	RotErr $$\mathcal{S}' = \mathrm{CUT3R\_update}(\mathcal{S}, I)$$1	TransErr $$\mathcal{S}' = \mathrm{CUT3R\_update}(\mathcal{S}, I)$$2	Mat. Pix. $$\mathcal{S}' = \mathrm{CUT3R\_update}(\mathcal{S}, I)$$3	FVD-V $$\mathcal{S}' = \mathrm{CUT3R\_update}(\mathcal{S}, I)$$4	CLIP-V $$\mathcal{S}' = \mathrm{CUT3R\_update}(\mathcal{S}, I)$$5
SynCamMaster	166.6	1451	3.83	8.83	236.7	1697	90.13
MV-Forcing	167.9	1469	3.64	8.26	251.1	1691	91.81

On long sequences, exemplified by $$\mathcal{S}' = \mathrm{CUT3R\_update}(\mathcal{S}, I)$$6 views and $$\mathcal{S}' = \mathrm{CUT3R\_update}(\mathcal{S}, I)$$7 frames, the reported comparison is:

Method	FID $$\mathcal{S}' = \mathrm{CUT3R\_update}(\mathcal{S}, I)$$8	FVD $$\mathcal{S}' = \mathrm{CUT3R\_update}(\mathcal{S}, I)$$9	RotErr $$\hat{I}, \hat{C} = \mathrm{CUT3R\_query}(\mathcal{S}, \mathcal{R}),$$0	TransErr $$\hat{I}, \hat{C} = \mathrm{CUT3R\_query}(\mathcal{S}, \mathcal{R}),$$1	Mat. Pix. $$\hat{I}, \hat{C} = \mathrm{CUT3R\_query}(\mathcal{S}, \mathcal{R}),$$2	FVD-V $$\hat{I}, \hat{C} = \mathrm{CUT3R\_query}(\mathcal{S}, \mathcal{R}),$$3	CLIP-V $$\hat{I}, \hat{C} = \mathrm{CUT3R\_query}(\mathcal{S}, \mathcal{R}),$$4
SF+ReCamMaster	157.6	1397	4.74	10.12	146.8	1759	87.26
SF+ReCamMaster+SF	156.8	1363	4.89	10.61	127.5	1771	86.61
MV-Forcing	153.3	1309	3.88	8.78	239.4	1554	90.83

These results are reported as showing that MV-Forcing matches or outperforms the teacher in camera accuracy and synchronization on short sequences, and outperforms all baselines in all metrics for long, multi-view, unbounded videos. The scaling results are stated as remaining stable up to $$\hat{I}, \hat{C} = \mathrm{CUT3R\_query}(\mathcal{S}, \mathcal{R}),$$5 views and over an $$\hat{I}, \hat{C} = \mathrm{CUT3R\_query}(\mathcal{S}, \mathcal{R}),$$6 increase in video length, up to $$\hat{I}, \hat{C} = \mathrm{CUT3R\_query}(\mathcal{S}, \mathcal{R}),$$7 frames (Fiebelman et al., 6 Jul 2026).

The ablations isolate the role of the main design choices. Without view-unrolling, there is a large degradation in cross-view consistency, attributed to exposure bias. Without the CUT3R geometric prior, synchronization and camera accuracy drop significantly. Without state accumulation, performance degrades slightly, which is used to support the importance of geometric context from multiple views. Classical manual rendering is also reported as less robust than learned raymap query (Fiebelman et al., 6 Jul 2026).

6. Scope, limitations, and terminological context

MV-Forcing is explicitly a long multi-view video generation framework, and the “MV” in its title denotes multi-view. This matters because “MV” has distinct meanings elsewhere. In mathematical origami, “MV assignment” refers to mountain-valley assignment, and forcing sets determine when a subset of creases uniquely determines the full foldable assignment in Miura-ori (Ballinger et al., 2014). In Fast Macroscopic Forcing Method, “MV-Forcing” is used in the sense of matrix-vector-based operator recovery for turbulence closure operators (Bryngelson et al., 2023). The term “forcing” also appears in supervisory control with event forcing (Reniers et al., 2024) and in forcing terms for Cascaded-LBM via central moments (Premnath et al., 2012). These are separate technical lineages rather than variants of the same framework.

A nearby but distinct generative framework is Motion Forcing, which addresses the trilemma of visual quality, physical consistency, and controllability by decoupling physical reasoning from appearance synthesis through a “Point-Shape-Appearance” paradigm and a Masked Point Recovery strategy (Xu et al., 11 Mar 2026). MV-Forcing instead centers on sequential view generation grounded by a persistent 4D reconstruction state, joint denoising, and self-forced autoregressive distillation (Fiebelman et al., 6 Jul 2026).

The stated limitation is that the teacher for Distribution Matching Distillation must support the desired domain, and adaptation to real-world video is performed via distillation from ReCamMaster. The paper also states that further improvements are possible with richer teachers and real-world multi-view datasets (Fiebelman et al., 6 Jul 2026). This suggests that the framework’s generalization is closely tied to the quality and domain coverage of the teacher-student distillation pipeline, even though the student itself is intended to operate as a single few-step model at arbitrary sequence lengths and viewpoint counts (Fiebelman et al., 6 Jul 2026).