Nvidia’s research division, working with the National University of Singapore, unveiled Flex-Forcing, a video diffusion training method that lets a single model switch between two competing generation styles instead of locking in one at training time. The technique targets a persistent tradeoff in video generation: bidirectional diffusion produces globally coherent, high-fidelity clips but runs slowly, while autoregressive generation streams frames fast but drifts and loses consistency over longer sequences. Flex-Forcing does not choose a side. It restructures how the model chunks a video during training so the same weights can operate at either extreme, or anywhere between them, depending on what a deployment needs. Nvidia published the work as an ICML 2026 spotlight paper.
The core mechanism is a chunking scheme applied jointly across two dimensions: the timeline of video frames and the sequence of denoising steps. A single large chunk spanning the whole clip recovers standard bidirectional diffusion, generating every frame with full context of what comes before and after. Splitting that same clip into many small chunks recovers autoregressive generation, where each segment is produced in order using only past context, the mode used for real-time streaming. Chunk size becomes a dial. A developer can turn it after training instead of locking an architecture decision in beforehand. At inference, the model can also mix regimes: it plans globally by attending across chunks, then generates the frames inside each chunk in autoregressive order for speed.
That dial matters because compute budgets vary sharply depending on where video generation runs. A cloud rendering pipeline can absorb the latency of bidirectional diffusion to maximize quality. A live avatar or interactive generation product cannot. It needs frames streamed as fast as they are consumed. Nvidia describes this as supporting different device budgets, meaning one trained checkpoint can serve both a high-end offline renderer and a low-latency interactive product without separate training runs.
The researchers also found that the best setting sits between the two extremes. Testing five-second clips split into three chunks, they ran a search over chunk-size configurations. A coarse-to-fine pattern, a large first chunk followed by two smaller ones, landed on the best combination of speed and Nvidia’s VBench quality score. That configuration beat both a uniform chunking baseline and the Self-Forcing autoregressive baseline on speed and quality simultaneously, and in some setups matched or exceeded fully bidirectional generation despite running faster.
Flex-Forcing also targets exposure bias, a known failure mode in streaming video models where small errors compound as a model generates frame after frame with no ability to look ahead. Nvidia compared thirty-second generations against the Infinity-RoPE baseline and reported that Flex-Forcing held up better over the extended horizon while preserving more motion. The chunking structure additionally enables any-order editing across timesteps. A developer can revise one chunk of a generated video without that edit spreading unpredictably through the rest of the clip, because the method keeps high-level structure separate from low-level detail.
For teams building real-time or interactive video products, Flex-Forcing is a training approach worth testing before locking a fixed autoregressive architecture into a roadmap. A single checkpoint that adjusts its speed-quality tradeoff at inference time removes the need to maintain separate bidirectional and streaming models for different products. The ICML 2026 spotlight selection is an early signal that unified chunking will draw scrutiny from the wider video-generation research field this year.
Nvidia Research, in collaboration with the National University of Singapore, published the Flex-Forcing findings on its research site on July 9, 2026, as an ICML 2026 spotlight paper.
