Recent advances in video diffusion models have enabled either long single-view generation through temporal autoregression, or short multi-view synthesis through bidirectional attention. However, generating long, multi-view consistent videos of dynamic scenes remains unsolved. In this work, we present MV-Forcing, a framework that composes temporal and view-wise autoregression within a single diffusion model by introducing a 4D geometric bridge between sequentially generated views. Our key insight is that an autoregressive 3D reconstruction model naturally interfaces between autoregressively generated views. Given a completed source view, we reconstruct its 3D structure and render a geometric prior of the next target viewpoint, which the diffusion model refines into a high-quality video. To extend generation beyond the teacher's fixed temporal window, we introduce a joint denoising regime where both view slots are initialized from noise during training, enabling temporally unbounded generation. We distill the model via Distribution Matching Distillation with Spatio-Temporal Self-Forcing, closing the train-inference exposure bias gap for both temporal and view-sequential autoregression. Extensive experiments on both synthetic and real-world data demonstrate that MV-Forcing produces geometrically consistent multi-view videos of dynamic scenes at arbitrary lengths and viewpoint counts using a single few-step student model.
Alex: Welcome to another episode of ResearchPod. Today, we're looking at a new framework called MV-Forcing—a system designed to solve one of the trickier problems in AI-generated video.
Sam: And what's the core problem it's trying to solve?
Alex: So imagine you want an AI to generate a video of a scene—say, a street corner—from several camera angles at once, and you want that video to be long enough to actually be useful. Right now, that's surprisingly difficult. You can have a short clip that shows the scene from multiple angles simultaneously, but it falls apart if you try to extend it. Or you can have a long video, but it loses track of the 3D world the moment the camera moves. You can have length, or you can have 3D consistency, but not both.
Sam: Why is that such a hard problem? What's actually going wrong under the hood?
Alex: Most AI video tools generate frames one by one, each slightly informed by the last. The trouble is that tiny mistakes compound. By the time you're a hundred frames in, the scene can look completely different from where it started—objects shift, lighting changes, the geometry stops making sense. Researchers call this "exposure bias." The model was trained on clean, real data, but at generation time it's feeding its own slightly imperfect output back into itself. The errors accumulate until the scene drifts.
Sam: Like a game of telephone, where each person passes the message along with a small error, and by the end it's unrecognisable.
Alex: That's a good analogy. And MV-Forcing is built specifically to break that cycle.
Sam: So how does it do that?
Alex: The central idea is what the researchers call a "4D geometric bridge." Think of a sketch artist who keeps a physical 3D model of a room on their desk. Every time they draw the room from a new angle, they glance at the model to check where the furniture actually sits—rather than just guessing from memory. MV-Forcing does something similar. The system maintains a continuously updated 3D map of the scene, and every new frame it generates has to be consistent with that map.
Sam: So the map is the anchor. Instead of each frame only looking back at the previous frame, it checks the full 3D record of the scene.
Alex: Exactly. And that solves a second problem at the same time. Earlier approaches tried to maintain consistency by comparing every frame to every other frame simultaneously—which becomes computationally overwhelming very quickly. By routing everything through the 3D map instead, the system only needs to consult one shared reference. It's far more efficient.
Sam: Right, because the map already contains the relevant information. You don't need to re-examine every frame if the map is keeping score.
Alex: Precisely. Now, the map alone handles the spatial problem—keeping the geometry consistent across camera angles. But there's still the temporal problem: how do you stop errors from building up over time as the video gets longer? That's where a technique called Spatio-Temporal Self-Forcing comes in.
Sam: What does that actually involve?
Alex: During training, the model is shown its own generated frames—not just clean, real data—and it's penalised whenever those frames drift from the 3D logic established earlier. So it learns, through practice, to stay on track even when it's working from its own imperfect output. It's a bit like a student who has to mark their own homework and correct their own mistakes before moving on. The key insight is that the model is being trained specifically on the conditions it will face at generation time, rather than only on ideal data.
Sam: That makes sense. So it's not just learning what good video looks like—it's learning how to recover when things start to go slightly wrong.
Alex: Exactly. And once the model is trained that way, there's still the question of speed. The full training process is slow and computationally heavy. So the researchers use a separate technique to address that.
Sam: How do they speed it up without losing quality?
Alex: They use something called Distribution Matching Distillation. The idea is to take a large, slow, highly capable model—the "teacher"—and use it to train a much smaller, faster model—the "student." The student learns to produce outputs that are statistically indistinguishable from the teacher's, but in a fraction of the time. Think of it like an experienced chef writing down a simplified recipe so a kitchen assistant can produce the same dish much faster. The student isn't just copying individual outputs—it's learning the teacher's overall approach and accuracy.
Sam: So you end up with something that's both high quality and practical to actually run.
Alex: That's the goal. And the paper suggests that combining these three elements—the 3D map, the self-forcing training, and the distillation—allows the system to generate videos of arbitrary length from any number of camera angles. That combination was not previously achievable in a single framework.
Sam: One thing I'm wondering: if the system generates frames one after another, does it have to complete one full camera view before moving to the next angle?
Alex: That was a limitation of older approaches. Here, the researchers use a grid-based structure, where time and camera viewpoint are treated as two axes. The system can move across that grid in any order—it's not locked into finishing one row before starting the next. The 3D map stays consistent regardless of the order frames are generated.
Sam: So it's genuinely flexible. The map keeps everything coherent no matter which angle or moment the system is working on at any given time.
Alex: Exactly. And because the generation is additive—meaning it can always append new frames to what it's already built—the video can, in principle, continue indefinitely. It's like a novelist keeping a detailed story bible. No matter which chapter they're writing, they check the bible to make sure the details stay consistent. The 3D map is that story bible—a persistent record that keeps everything coherent across time and viewpoint.
Sam: It sounds robust. Are there limitations worth noting?
Alex: A few. The model is only as good as its starting point. If the very first frame contains an error, the 3D map will carry that error forward—it corrects drift, but it doesn't fix a bad foundation. And the training data used here was largely synthetic. The system can be fine-tuned on real-world footage, but training on larger, real-world multi-view datasets is identified as the natural next step.
Sam: So it's a meaningful step forward, but not a complete solution to every problem in this space.
Alex: That's a fair assessment. What the paper demonstrates is a clear path from short, disconnected clips toward longer, geometrically coherent video from multiple angles—and a practical way to make that generation fast enough to be useful. Thanks for listening to ResearchPod.