Skip to main content
Back

NVIDIA's Flex-Forcing Lets One Model Stream and Render Video at 48 FPS

NVIDIA's Flex-Forcing lets a single video diffusion model switch between fast streaming and high-quality generation at inference time, earning an ICML 2026 spotlight.

Paper, Full Access
  • Single model, all modes: Flex-Forcing lets one video diffusion model run as bidirectional, autoregressive, or anything in between at inference time.
  • ICML 2026 spotlight: The paper earned a spotlight designation at ICML 2026, one of the conference's top recognition tiers.
  • Speed + quality wins: Best config scores 85.07 VBench at 25.8 FPS; fastest config hits 29.4 FPS — both beating Self-Forcing baseline.
  • Key innovation — K-Projection: A lightweight timestep-conditioned layer aligns noisy and clean key states in attention, enabling stable mixed-mode inference.
  • Any-order video editing: The model can re-edit arbitrary middle segments of a video without regenerating the whole clip, using both past and future frames as context.
  • No public weights yet: Built on open Wan2.1-T2V-1.3B; training details are published but model checkpoints have not been released.

Video generation has long forced a hard choice: do you want quality or speed? Bidirectional diffusion models look at every frame at once and produce coherent, visually rich clips, but they are slow and expensive to run. Autoregressive models generate frame by frame, enabling real-time streaming and arbitrarily long videos, but they accumulate errors over time and lose track of what happened earlier in the clip. NVIDIA Research's new Flex-Forcing argues you should not have to choose at all.

One model, every mode

Flex-Forcing is a unified training and inference framework that enables a single video diffusion model to seamlessly operate under both bidirectional and autoregressive generation regimes. The key insight is that these two paradigms are not fundamentally different architectures , they are just two extreme points on a continuous spectrum of how much context each frame is allowed to see.

The core idea is a flexible chunking mechanism jointly defined over the temporal axis and denoising steps. This design allows the model to perform flexible chunking according to different device budgets, perform bidirectional inference across chunks for global structure planning while generating frames autoregressively within each chunk for efficient synthesis, and perform any-order, any-timestep autoregressive generation without the strict causal constraint.

Think of it like this: in fully autoregressive mode, every frame is its own chunk of size one. In fully bidirectional mode, all frames are one giant chunk. Flex-Forcing lets you pick anything in between , and crucially, the same trained model handles all of it.

The two problems it had to solve

Unifying these paradigms in a single model requires solving two non-trivial technical problems.

The first is training the model to handle both causal (left-to-right) and non-causal (full-context) attention simultaneously. Flex-Forcing introduces a unified training and inference framework that enables the model to seamlessly operate under both regimes. During training, the team randomly samples different chunk sizes at each step, exposing the model to the full spectrum of attention patterns from strictly causal to fully bidirectional.

The second problem is more subtle. When a frame attends to both past frames (which are already clean, fully denoised) and future frames (which are still noisy), the attention mechanism sees key-value states at wildly different noise levels. Standard self-attention treats all of them the same way, which breaks things. The paper's fix is a lightweight component called K-Projection: a timestep-conditioned linear layer that projects the clean cached keys from past frames into the same noisy representation space as the current denoising step, so all keys are noise-consistent before attention is computed.

How it was built

Bidirectional diffusion models excel at global coherence and visual fidelity but suffer from slow inference, while autoregressive models offer efficient and streaming generation at the cost of long-range consistency and exposure bias. Flex-Forcing is built on top of Wan2.1-T2V-1.3B as the student model, with Wan2.1-T2V-14B as the teacher. Training runs for only 600 iterations with a batch size of 64, randomly sampling chunk sizes between 2 and 10 frames. The training pipeline follows a two-stage process: first, ODE initialization injects causal attention masks into the bidirectional base model; then, asymmetric distillation with DMD (Distribution Matching Distillation) strengthens the causal behavior while preserving bidirectional capacity through stochastic chunking during rollouts.

The numbers

The paper benchmarks against Self-Forcing, the prior state-of-the-art for fast autoregressive video generation, on VBench (a standard video quality benchmark) and VBench-Long for 30-second clips. The results are strong across the board:

  • Best quality config (15-3-3 frame split): VBench total score of 85.07 at 25.8 FPS , beating Self-Forcing's 84.31 at the same speed.
  • Fastest config (7-7-7 equal split): 29.4 FPS with a score of 84.63, a meaningful throughput gain over the baseline's 24.9 FPS.
  • 3-step denoising: Pushing to only 3 function evaluations, Flex-Forcing hits 45.6–48.9 FPS while maintaining competitive quality scores around 83.9–84.3.
  • 30-second videos: On VBench-Long, Flex-Forcing outperforms Infinity-RoPE on most metrics, with the most dramatic gain in dynamic degree (+21 points), meaning the generated videos have significantly more motion diversity rather than repetitive loops.

Recent progress in large-scale generative models has substantially advanced video generation, yet existing methods remain constrained by a rigid inference paradigm. Bidirectional diffusion models excel at global coherence and visual fidelity but suffer from slow inference, while autoregressive models offer efficient and streaming generation at the cost of long-range consistency and exposure bias. Flex-Forcing's hybrid configs consistently push the Pareto frontier , you get more quality per unit of compute than either pure approach.

A bonus capability: surgical video editing

The flexible attention structure unlocks something that pure autoregressive models cannot do cleanly: editing an arbitrary segment of an already-generated video without re-generating everything else. Because the model can attend to both past and future frames, you can regenerate frames 26–49 of an 81-frame clip while keeping frames 1–25 and 50–81 fixed as context. The paper calls this any-order, any-timestep editing. It also supports editing only at late denoising steps (fine detail refinement) while keeping the early steps (global structure) frozen, which prevents local edits from cascading into global changes , a common failure mode in standard autoregressive editing.

Where it falls short

The paper is candid about limitations. Autoregressive models offer efficient and streaming generation at the cost of long-range consistency and exposure bias , and Flex-Forcing does not fully eliminate this. Error accumulation in very long videos is still present, just reduced. The method also depends heavily on the bidirectional priors inherited from pre-training, which means it may not transfer cleanly to models that were trained purely autoregressively from scratch. And like all generative video models, it carries the usual misuse risks.

Why this matters for the field

The video generation field has been bifurcating: one camp optimizes bidirectional models for quality (Sora, Veo, Wan), another optimizes autoregressive models for speed and streaming (CausVid, Self-Forcing, MAGI-1). Flex-Forcing challenges the assumption that these require separate architectures and separate training runs. NVIDIA had 74 papers accepted at ICML 2026 , and Flex-Forcing earned a spotlight , a designation given to a small fraction of accepted papers. The practical implication is significant: instead of maintaining two specialized models for different deployment scenarios, a single Flex-Forcing model can serve both a high-quality offline rendering pipeline and a real-time interactive streaming application, with the tradeoff dialed at inference time based on available compute.

The paper is available on arXiv, and the project page includes video examples. No weights or code have been released publicly yet, but the training setup is described in enough detail to reproduce on top of the open Wan2.1 base model.