Autoregressive Video Gen

Preface: AR in LLMs

Modern language models turn generation into a sequence of AR predictions. Instead of solving an entire response at once, an AR LLM factorizes it from left to right:

p(y₁:T | x) = ∏ₜ pθ(yₜ | x, y<t)

The causal attention mask is more than a modeling choice. It creates a systems contract: a generated token never depends on future tokens, so completed prefix states remain valid. The server can cache their keys and values, emit output immediately, and advance the sequence without recomputing the whole prefix.

ModelOne-way information flow

Each new token reads the prefix, while the already emitted prefix stays fixed.

StateReusable KV cache

Past attention states become persistent session memory rather than repeated computation.

ServingA persistent decode loop

Streaming, cancellation, batching, and memory management operate at token boundaries.

This AR interface is the hidden reason LLM serving scales as a long-running system rather than a collection of fixed-length batch jobs. AR video generation asks whether video models can expose the same kind of stable boundary—at the level of frames or latent blocks—while retaining the visual quality of diffusion.

1. Why AR video generation matters

Full diffusion is a great clip generator—and a poor streaming primitive

A standard video diffusion transformer receives a noisy latent clip and repeatedly updates the whole clip with bidirectional spatiotemporal attention. This is a powerful inductive bias for a fixed-duration sample: every frame may use evidence from both past and future, and the model can repair globally inconsistent motion over multiple denoising iterations.

But the same coupling creates three deployment problems. First, the model cannot emit the beginning of the video while the future is still unresolved. Second, extending the duration increases the token set processed at every denoising step. Third, a prompt change in the middle of generation has no natural boundary: the sampler was designed to solve one global clip, not to maintain a live session.

AR diffusion factorizes time into frames or blocks:

p(x₁:T | c) = ∏ₖ pθ(x[Bₖ] | x[<Bₖ], c),    with a small diffusion/flow solver inside each block Bₖ

This factorization introduces a sequential dependency across blocks, but it also creates a reusable boundary. Once block k is clean, it can be streamed to the user and summarized into a KV cache while block k+1 is denoised. The active working set can be bounded with a rolling window, and the prompt can be updated between blocks.

Full-sequence diffusion compared with AR diffusionFull diffusion repeatedly updates an entire bidirectional clip. AR diffusion emits blocks sequentially and reuses a bounded KV cache.Full-sequence diffusionAR diffusionBidirectional clip, repeated global updatesAR blocks, streaming updatesEvery step revisits every frameNo frame is final until the clip is finalemitdenoisenextrolling KV cacheBounded state; output appears block by block
Figure 1. The architectural trade: full diffusion buys global bidirectional repair, while AR diffusion buys a streaming boundary, reusable state, and bounded memory. Few-step distillation is what makes each new block cheap enough to serve.

The wall-clock gap: bidirectional diffusion vs causal streaming

For a concrete long-video comparison, the LongLive paper uses SkyReels-V2. SkyReels-V2 is a Diffusion-Forcing system rather than a vanilla full-clip sampler, but its bidirectional attention preserves the same serving bottleneck: completed history cannot be reused as a KV cache under AR decoding. On a single H100, the reported gap for producing a 60-second video is:

Same prompt, same 60-second sequence. 0–10 s: “In a warmly lit home office, a bearded Asian man types intently on a laptop. An orange tabby cat sits beside him.”
Full / bidirectional diffusion SkyReels-V2
Computing the complete video before anything can be shown…
Generation progress0%
Video becomes available only after generation reaches 100%.
Causal video LongLive