We propose OPSD-V, an on-policy self-distillation paradigm for post-training few-step autoregressive (AR) video diffusion models. Existing few-step AR video generators can produce long videos with low latency, but still suffer from error accumulation and weakened motion dynamics during long autoregressive rollout. OPSD-V reduces long-horizon degradation while preserving the original few-step inference path. The key idea is to introduce real long-video data as temporal context during training and use it to provide dense trajectory-level supervision. Specifically, the student follows the exact inference-time rollout, generating each chunk conditioned on its own previously generated KV cache. In parallel, the teacher is evaluated at the same student-visited denoising states, but uses a cleaner AR-consistent temporal cache in which older history can be replaced by real-video context. This provides dense denoising-level corrective targets under on-policy AR cache dynamics, without changing the sampler, number of denoising steps, or inference-time cache mechanism. We apply OPSD-V to representative few-step AR video models, including Self-Forcing and LongLive. Experiments show consistent improvements in visual quality, motion dynamics, and VBenchLong scores. A user study with 10 participants comparing 20 video pairs shows that OPSD-V is preferred over the base models in 66.0% of overall-preference judgments (82.5% excluding ties).
Alex: Welcome to another episode of ResearchPod. Today, we're looking at a new method for AI video generation called OPSD-V, which stands for On-Policy Self-Distillation for Video generation.
Sam: So this paper is asking how we stop AI-generated videos from falling apart after just a few seconds of playback?
Alex: Exactly. When an AI tries to generate a long video, it doesn't create the whole thing at once. It predicts one small chunk at a time, using its own previous work as a guide for what comes next.
Sam: Like a game of telephone, where the message gets distorted a little every time it's passed on?
Alex: That's a good way to put it. The model builds up a kind of memory of everything it has already created, so it can keep the video consistent. But if it makes a tiny mistake early on, that error gets baked into the memory.
Sam: So it ends up trapped in a loop of its own errors. And because it keeps using that flawed memory, the video quality just degrades over time?
Alex: That's the core issue. It keeps tripping over its own past mistakes, and the longer the video runs, the worse it gets.
Sam: And this paper proposes a way to fix that memory—without making the model slower to run?
Alex: Right. To understand the fix, it helps to know what that memory actually looks like. The model stores bits of previous frames in a buffer so it doesn't have to recalculate everything from scratch each time. Think of it like a scratchpad the AI uses to remember what it just drew. If that scratchpad gets cluttered with errors, the next frame will be messier too. Researchers call this buffer a KV Cache.
Sam: So the question is: how do you keep the scratchpad clean, if the AI is the one generating the mess in the first place?
Alex: That's exactly the problem OPSD-V is designed to solve. The approach is to have the AI act as both a student and a teacher at the same time. The student generates the video as it normally would—making those typical small errors along the way.
Sam: And the teacher? Does it have some perfect version to compare against?
Alex: Not a perfect version, no. The teacher looks at the same moment in time, but it gets to use a clean version of the history—drawn from real video footage—to see what should have happened. So it's not judging the student against an ideal; it's showing the student a more stable path from the same starting point.
Sam: So the teacher isn't just grading the final result. It's helping the student correct course while it's still in the middle of generating?
Alex: That's the key insight. And this is why the "on-policy" part of the name matters. The teacher always meets the student where the student actually is—in the middle of its own messy, error-prone trajectory. If the teacher only ever showed the student clean, perfect paths, the student would never learn how to recover from its own mistakes.
Sam: You're not forcing the AI to copy the real video, but you're giving it a stable reference point so it doesn't drift into nonsense?
Alex: Exactly. Think of a student practicing a speech from memory. The teacher listens with the script in hand and gives a quiet nudge whenever the student starts to wander—not to make them recite it word for word, but to keep them on track.
Sam: You mentioned the method uses something called velocity matching. What does that add?
Alex: Right, so instead of just comparing finished frames, the teacher also looks at the direction and rate of change at every step of the generation process. Think of it like correcting a drawing while the pen is still moving, rather than waiting for the whole picture to be done. That step-by-step guidance keeps fine structural details—the edge of a building, the movement of a person's hand—much sharper over time.
Sam: That makes sense. But does it come with any downsides? If you're training the model to prioritize stability, do you lose some of its ability to follow the original prompt?
Alex: There is a mild trade-off, and the paper is upfront about it. By emphasizing stable, coherent motion, the model can become slightly less precise about following the specific details of a text prompt. It's a balance between moving smoothly and sticking strictly to the user's instructions.
Sam: And I'd imagine the quality of this training depends heavily on having good real-world video data to work from?
Alex: That's correct, and it's a genuine limitation. The quality of what the teacher can show the student is only as good as the real-world footage used during training. That could be a constraint for developers who don't have access to large, high-quality video datasets.
Sam: Still, the core idea is notable. You don't need a bigger model or more computing power to get better results—you just need to teach the model how to remember more carefully.
Alex: That's the central takeaway. By managing the model's history more effectively during training, the method achieves longer, more coherent video generation without adding any extra cost when you actually run it. The speed and structure of the model stay exactly the same—it's just been taught to handle its own mistakes more gracefully. Thanks for listening to ResearchPod.