Standard autoregressive video generation algorithms based on Diffusion and Flow Matching rely on rigid training objectives and static sampling schedules, limiting inference procedures from adapting to the data. We introduce Equilibrium Forcing (EqF), a simplified framework for video denoising generative models without noise level conditioning. EqF pioneers modular training- and inference-time designs for noise-unconditional generation that decouple learning the denoising field from sampling. This flexibility allows for inference-time algorithms that operate in a closed loop by adapting to feedback from the sample, improving video quality and consistency on challenging autoregressive video generation benchmarks. Extensive analysis elucidates exactly how removing the noise level conditioning enables EqF's data-dependent inference properties to surpass the performance of standard noise level-conditional denoising video methods.
Figures & tables
Figure 1: Equilibrium Forcing Overview. EqF is trained with a simple unmodulated equilibrium denoising objective (left), which can be shaped at inference time through the η warp (middle). The sampling landscape here is shaped to be an attractor, with the magnitude decreasing as the sample descends and converges (right).
Figure 2: Left: FM follows externally-defined noise schedule conditioning, which diverges from the true denoising field. Center: By training without explicit noise level conditioning, EqF learns a correct denoising field everywhere by estimating the noise level internally. Right: EqF leverages the internal noise level estimate for adaptive autoregressive video generation inference.
Figure 3: Left : Absolute noise conditioning error between scheduled σ(i) and the predicted σ~(i)≈gϕ(x(i)) increases over 256 autoregressive rollouts. Right : FVD worsens over 256 autoregressive rollouts with FM due to externally prescribed noise conditions. Conditioning on estimated noise levels (dotted) helps slightly. Removing the noise condition altogether ( EqF ) performs best.
Figure 4
Figure 5: Budget-adaptive inference results. EqF with budget-adaptive sampling achieves the lowest FVD across Minecraft, Re10K, and Droid.
Method
Sampler
Inference Warp
FVD ↓
VBench ↑
EqF (Ours)
NAG, μ=0.3
c -function
64.34±1.78
0.7733±0.0008
EqM ( Wang and Du, 2025 )
NAG μ=0.1
–
77.02±5.21
0.7679±0.0009
EqF
GD
c -function
77.63±3.31
0.7721±0.0005
FM
Euler
c -function
104.13±4.33
0.7703±0.0012
NU-FM ( Sun et al., 2025 )
GD
–
101.16±3.57
0.7698±0.0011
Diffusion
DDIM
–
106.00±1.22
0.7681±0.0008
Table 1: Autoregressive Minecraft 300-frame rollout results across 5 seeds. EqF with NAG sampling achieves the best FVD and VBench scores.
Figure 6: Visualizations of autoregressively generated RealEstate10K videos using EqF -FT and FM-FT finetuned models from Wan 2.1.
Figure 7: Generated videos from the Droid dataset using EqF -FT finetuned from Wan 2.2.
Figure 9Figure 10
Minecraft
Re10K
Droid
Method
FVD ↓
VBench ↑
FVD ↓
VBench ↑
FVD ↓
VBench ↑
EqF Adaptive NAG
91.45±0.92
0.771±0.000
127.69±1.57
0.761±0.000
107.38±0.82
0.812±0.002
EqF NAG
92.08±0.30
0.774±0.000
166.73±2.98
0.759±0.000
113.58±3.51
0.810±0.002
FM
135.37±1.74
0.769±0.000
145.42±5.93
0.761±0.000
108.69±3.71
0.809±0.000
Table 3: Comparison of FVD and VBench at 50 sampling steps across Minecraft, RealEstate10K, and Droid datasets using the budget-adaptive EqF algorithm. NAG μ=0.1 .
Table 12
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 12: Pearson Correlation convergence across training of the Re10K FM and EqF models.
Dataset
# Samples
FM/ EqF Field MSE
Validation v -Loss
Relative Field MSE (%)
Pearson Correlation
E∥fFM−f\textscEqF∥22
E∥fFM−v∥22
E∥f\textscEqF−v∥22
Minecraft
2,084
0.0425±0.00505
0.576±0.00270
0.576±0.00271
7.378
0.984
Re10K
4,096
0.00347±0.00003
0.142±0.00146
0.143±0.00147
2.432
0.999
Droid
1,417
0.0135±0.000385
0.307±0.00972
0.318±0.00993
4.325
0.996
Appendix
Table 6: Empirical agreement between trained FM and EqF denoisers. All quantities are evaluated on matched held-out samples using the same clean inputs, noise levels, and Gaussian noise. Relative field MSE denotes the magnitude of E∥fFM−f\textscEqF∥22 normalized by 21(E∥fFM∥22+E∥f\textscEqF∥22) . Across all three domains, the trained fields have small normalized discrepancy and high Pearson correlation despite EqF receiving no explicit noise-level condition.
Figure 13: Noise level readout across sampling steps. Ideally, sampling ends at noise level 0 , but Flow Matching samples terminate at a significant, nonzero noise level, which is a potential contributing factor to autoregressive error accumulation. The right shows a zoomed view of the red box from the left plot.
Figure 14: Noise conditioning error introduced artificially, with the loss calculated only on the active sampling horizon. Left: horizon only. Right: gain from adding context corruption. Noise conditions are clamped to [0,1] , and errors are standardized with respect to δ∈[−0.07,0.07] .
Method
Sampler
Inference Schedule
Conditioning
FVD ↓
Flow Matching
Euler
Identity
σ(i)
126.633
Flow Matching
Euler
Identity
σ~(i)≈gϕ(x(i))
121.145
Flow Matching
Euler
c -Function
σ(i)
103.610
Flow Matching
Euler
c -Function
σ~(i)≈gϕ(x(i))
94.488
Flow Matching
NAG, μ=0.2
c -Function
σ(i)
1045.633
Flow Matching
NAG, μ=0.2
c -Function
σ~(i)≈gϕ(x(i))
81.193
Appendix
Table 7: Flow Matching autoregressive generation results under different inference schedules and conditioning choices. Conditioning on the inferred noise level gϕ(x(i)) substantially improves FVD, especially when combined with NAG under the c Function inference schedule.
Figure 15: Comparison of the noise level that sampling terminates at across FM, NU-FM, EqF , and EqF with NAG. Ideally, sampling ends at a noise level of 0 . Closed-loop sampling with EqF terminates at a lower noise level, and NAG accelerates the schedule and ends at an even lower noise level. The right shows a zoomed view of the red box from the left plot.
Figure 16: Early Stopping on Minecraft with EqF . The early stopping algorithm using noise level readouts optimizes the compute/performance tradeoff. Each value of μ is evaluated with stopping thresholds σthresh=0.1,0.2,0.3,0.5,0.7
Inference Mode
Sampler
Stopping Noise Level σthresh
N
FVD ↓
Adaptive Early Stopping
EqF GD
0.7
27.7
161.67
0.5
38.9
94.87
0.3
49.8
86.70
0.2
55.4
89.88
0.1
67.6
94.75
Adaptive Early Stopping
EqF NAG μ =0.1
0.7
26.0
161.59
Appendix
Table 8: Adaptive early stopping experiments on Minecraft with EqF . For adaptive runs, the stopping noise level induces an empirical number of function evaluations N . Higher setting of μ for NAG makes the adaptive stopping condition trigger sooner. All runs use c -function as the η warp. These are the values plotted in Figure 16 .
Table 21
Config
Minecraft
Re10K
Droid
Training
Effective Batch Size
8
8
16
Learning Rate
2e-4 (Linear Warmup)
1e-4 (Linear Warmup)
1e-5 (Linear Warmup)
Warmup Steps
2,000
2,000
2,000
Weight Decay
1e-3
1e-3
1e-3
Training Steps
580k
100k
60k
Appendix
Table 12: Training settings for Minecraft, Re10K, and Droid across models. The Rectified Flow schedule is from Liu et al. (2023) and the Cosine schedule is from Nichol and Dhariwal (2021) . Minecraft is trained from scratch. For Re10K, we finetune Wan T2V 1.3B ( Wan et al., 2025 ) . For Droid, we finetune Wan TI2V 5B.
Table 23
Figure 17: Re10K Validation Loss and FVD across equilibrium finetuning steps.
Figure 18: Training loss for noise level estimation. The readout predictor hω , trained on EqF activations, converges substantially faster than the framewise predictor gϕ trained directly from noisy data.
Figure 19: Marginalized residuals of noise level estimates from the framewise noise level predictor gϕ trained directly on noisy data (left) and the readout predictor hω trained on EqF activations (right). Both predictors are accurate, but hω is especially accurate near the low-noise boundary.
Figure 20: Conversions between noise level, discretized solver time, and induced warp. a) Conversion between σ and update multiplier under different choices of the warp function η . b) Induced conversion between solver time s and noise level σ through the trajectory ρ . Equally sized steps in solver time can correspond to different changes in noise level. c) Warp multiplier as a function of solver time through the composition η(ρ(s)) . Due to the vanishing update multiplier for attractor-style choices of η , equally sized steps in solver time take increasingly smaller steps as the sampler approaches data.
Figure 21: Online replanning on Minecraft. The replanning threshold τ controls when the retained future horizon is discarded and regenerated. τ=∞ never replans, while τ=0 always replans. Intermediate thresholds trade off reuse of the current plan against responsiveness to new observations.
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. We introduce Flex-Forcing, a unified training and inference framework that enables a video diffusion model to seamlessly operate under both bidirectional and autoregressive generation regimes. The core idea is a flexible chunking mechanism jointly defined over the temporal axis and denoising steps. This design allows the model to (1) perform flexible chunking according to different device budgets, (2) perform bidirectional inference across chunks for global structure planning, while generating frames autoregressively within each chunk for efficient and fine-grained synthesis, and (3) perform any-order, any-timestep autoregressive generation without the strict causal constraint. Extensive experiments on multiple video generation benchmarks demonstrate that Flex-Forcing achieves consistently better video quality, long-video stability than strong baselines with a rigid inference schedule, while offering faster inference.
Xinyin Ma, Julius Berner, Chao Liu +3
Electrical and Computer Engineering, National University of Singapore, Singapore · NVIDIA Research, California, USA
Autoregressive (AR) video diffusion models generate streaming video of arbitrary length by conditioning each new chunk of frames on those already generated. However, they suffer from exposure bias, where a model trained on clean context must continue from its own imperfect outputs, and errors compound until quality collapses within seconds (drift). Existing methods either require cluster-scale base retraining, apply training-free inference corrections with inconsistent results, or train frozen-base correctors for sampling and cache errors. We instead find that drift is largely deterministic and therefore learnable, as holding the noisy state fixed while varying only the history reveals a counterfactual velocity gap that is ~95% systematic across noise seeds. Based on this observation, we introduce Clean Forcing, which trains a 5.9M-parameter LoRA corrector on the frozen base model to regress this gap, with a closed-loop stage that adds DAgger-style aggregation and a drift-contraction objective to expose the corrector to its own rollouts. Because the clean histories can be the base model's own single-shot generations, training needs no real videos, and the merged corrector adds no inference cost. In 50 s text-to-video generation, Clean Forcing reduces Δ-drift from +11.3 to +1.65 without real video and to -0.16 with 40 real clips, improving on the best published result of +3.57 with over 150x less data. Extensive experiments demonstrate that Clean Forcing outperforms prior methods in aesthetic quality and data efficiency with reduced drift for long-horizon video generation.
Wenqing Wang, Joonghyuk Shin, Jonathan Tremblay +2
Northeastern University · Seoul National University · NVIDIA
Streaming video generation holds strong potential for world modeling, where future frames must be inferred online sequentially to form a continuous video stream. However, streaming video diffusion models introduce a fundamental train-inference mismatch: inference follows a specialized denoising order, whereas advanced training strategies typically require diverse noise-level configurations. To address this trade-off between train-inference consistency and training coverage, we reformulate the video diffusion sampling as a frame-indexed stochastic process over noise levels. Within this stochastic process space, we construct a continuous training trajectory along which the sampling schedule progressively evolves from independent sampling to inference-consistent sampling. We further introduce a joint calibration algorithm and a temporal correlative sampling algorithm to ensure trajectory smoothness and cross-frame correlation. Building on these designs, we propose Stream Forcing, a unified training framework for streaming video generation that balances training sufficiency and inference efficiency. Extensive experiments demonstrate that Stream Forcing significantly improves generation quality with a 36.6% FVD improvement on the UCF-101 benchmark. Furthermore, our method facilitates robust zero-shot extrapolation to long-horizon video generation with a 27.9% FVD improvement on the UCF-101 benchmark.
Yueting Zhu, Yuehao Song, Kaicheng Zhang +5
Huazhong University of Science & Technology · Anyverse Dynamics · Horizon Robotics