Helios: Real Real-Time Long Video Generation Model

Shenghai Yuan; Yuanyang Yin; Zongjian Li; Xinwei Huang; Xiao Yang; Li Yuan

cs.CVMar 4, 2026

Helios: Real Real-Time Long Video Generation Model

Authors: Shenghai Yuan, Yuanyang Yin, Zongjian Li, Xinwei Huang, Xiao Yang, Li Yuan

Organizations: Peking University · ByteDance China · Chengdu Anu Intelligence · Canva

Abstract

We introduce Helios, the first 14B video generation model that runs at 19.5 FPS on a single NVIDIA H100 GPU and supports minute-scale generation while matching the quality of a strong baseline. We make breakthroughs along three key dimensions: (1) robustness to long-video drifting without commonly used anti-drifting heuristics such as self-forcing, error-banks, or keyframe sampling; (2) real-time generation without standard acceleration techniques such as KV-cache, sparse/linear attention, or quantization; and (3) training without parallelism or sharding frameworks, enabling image-diffusion-scale batch sizes while fitting up to four 14B models within 80 GB of GPU memory. Specifically, Helios is a 14B autoregressive diffusion model with a unified input representation that natively supports T2V, I2V, and V2V tasks. To mitigate drifting in long-video generation, we characterize typical failure modes and propose simple yet effective training strategies that explicitly simulate drifting during training, while eliminating repetitive motion at its source. For efficiency, we heavily compress the historical and noisy context and reduce the number of sampling steps, yielding computational costs comparable to -- or lower than -- those of 1.3B video generative models. Moreover, we introduce infrastructure-level optimizations that accelerate both inference and training while reducing memory consumption. Extensive experiments demonstrate that Helios consistently outperforms prior methods on both short- and long-video generation. We plan to release the code, base model, and distilled model to support further development by the community.

Figures & tables

Figure 1: End-to-end throughput (FPS) of various video generation models on a single H100. The results are obtained at the same resolution with all official acceleration techniques, including FlashAttention, torch compile, and KV-cache. Helios is substantially faster than models at the same scale and matches the speed of smaller distilled ones.

Figure 2: Benchmark performance of Helios and its counterparts. For both short- and long-video generation, Helios consistently outperforms existing distilled models while achieving performance comparable to that of base models.

Figure 3: Showcases of infinite videos generated by Helios. Despite overhead comparable to that of the 1.3B models [ 90 , 100 , 59 , 60 , 126 ] , Helios still excels in visual quality, text alignment, and motion dynamics.

Figure 4: Architecture of Helios. Helios is an autoregressive video diffusion transformer built with Guidance Attention blocks. It reduces overhead by compressing historical and noisy context through Multi-Term Memory Patchification and Pyramid Unified Predictor Corrector, while unifying T2V, I2V, and V2V tasks via Representation Control.

Figure 5: Visualization of three representative drifting patterns in long-video generation.

Figure 6: Temporal trends of saturation, aesthetic, and RGB statistics for normal videos versus drifting videos. Normal videos are stable, while drifting videos initially follow a similar trajectory but shift abruptly and remain unstable.

Figure 7: Overhead reduction with Multi-Term Memory Patchification. A hierarchical history window uses progressively larger kernels, keeping the token budget constant while extending the context length.

Figure 8: Outline of Pyramid Unified Predictor Corrector. The process consists of three stages: (i) Low-stage focuses on efficiency, (ii) Mid-stage balances quality and efficiency, and (iii) High-stage prioritizes quality.

Figure 9: Pipeline of Adversarial Hierarchical Distillation. The framework is based on DMD [ 105 ] , with improvements such as Pure Teacher Forcing, Staged Backward Simulation, Coarse-to-Fine Learning and Adversarial Post-Training.

Figure 10: Execution of Cache Grad for GAN. We cache discriminator gradients w.r.t. inputs to decouple backpropagation and free intermediate activations early, substantially reducing peak memory.

Configuration	Stage-1-init	Stage-1-post	Stage-2-init	Stage-2-post
Global Batch Size	128	128	256	192
Optimizer	AdamW, $\beta_{1}=0.9$ , $\beta_{2}=0.999$ $\epsilon=1e-08$ , weight_decay=1e-04	AdamW, $\beta_{1}=0.9$ , $\beta_{2}=0.999$ $\epsilon=1e-08$ , weight_decay=1e-04	AdamW, $\beta_{1}=0.9$ , $\beta_{2}=0.999$ $\epsilon=1e-08$ , weight_decay=1e-04	AdamW, $\beta_{1}=0.9$ , $\beta_{2}=0.999$ $\epsilon=1e-08$ , weight_decay=1e-04
Learning Rate	5e-5	3e-5	1e-4	3e-5
Learning Rate Schedule	Constant	Constant	Constant	Constant
Training Steps	5.5k	7.5k	16k	20k
Gradient Clipping	1.0	1.0	1.0	1.0

Table 1: The detailed training hyperparameters of Stage-1 and Stage-2.

Configuration	Stage-3-ode	Stage-3-post
Global Batch Size	128	128
Real-score initialization	-	Helios-Base
Fake-score initialization	-	Helios-Base
Real-score CFG weight	-	3.0
Optimizer ( $G_{\theta}$ & $p_{fake}$ )	AdamW, $\beta_{1}=0.0$ , $\beta_{2}=0.999$ $\epsilon=1e-08$ , weight_decay=1e-03	AdamW, $\beta_{1}=0.0$ , $\beta_{2}=0.999$ $\epsilon=1e-08$ , weight_decay=1e-03
Learning Rate $(G_{\theta})$	2.0e-6	2.0e-6

Table 2: The detailed training hyperparameters of Stage-3.

Model	#Params	Throughput (FPS) $\uparrow$	Total $\uparrow$	Aesthetic $\uparrow$	Dynamic $\uparrow$	Smoothness $\uparrow$	Semantic $\uparrow$	Naturalness $\uparrow$
Bidirectional Models
SANA Video [ 9 ]	2B	1.34	4.60	9	7	9	5	1
CogVideoX [ 103 ]	2B	1.95	5.55	7	10	7	4	5
CogVideoX 1.5 [ 103 ]	5B	1.47	5.15	7	5	8	5	4
Mochi-1 [ 80 ]	10B	0.53	5.35	6	7	9	5	4
HV Video [ 41 ]	13B	0.36	6.00	8	8	9	5	5

Table 3: Quantitative comparisons on 81-frames short videos. Helios achieves a superior speed-quality trade-off over existing methods, which are either computationally prohibitive or yield suboptimal results. “

\uparrow

” higher is better.

Figure 11: Qualitative comparisons on 81-frames short videos (Part-1). Even as a distilled model, Helios matches or even surpasses the base models in terms of visual quality, motion dynamics, and naturalness.

Figure 12: Qualitative comparisons on 81-frames short videos (Part-2). Despite being a distilled model, Helios matches or surpasses the base models in visual fidelity, text alignment, and overall realism.

Model	#Params	Throughput (FPS) $\uparrow$	Total $\uparrow$	Throughput Score $\uparrow$	Total* $\uparrow$	Aesthetic $\uparrow$	Dynamic $\uparrow$	Smoothness $\uparrow$	Semantic $\uparrow$	Naturalness $\uparrow$	Drifting Aesthetic $\uparrow$	Drifting Smoothness $\uparrow$	Drifting Semantic $\uparrow$	Drifting Naturalness $\uparrow$
Autoregressive Models
NOVA [ 18 ]	0.6B	1.18	2.48	1	2.38	1	1	9	1	2	1	9	1	2
Pyramid Flow [ 38 ]	2B	3.11	2.85	1	2.75	6	3	9	3	1	1	8	1	2
MAGI-1 [ 86 ]	4.5B	0.37	4.92	1	4.82	7	3	10	6	2	1	10	6	5
InfinityStar [ 58 ]	8B	3.38	2.63	2	2.43	4	4	8	2	1	1	9	1	1
SkyReelsV2-DF [ 8 ]	1.3B	0.55	3.29	1	3.19	6	6	9	4	1	1	9	2	1

Table 4: Quantitative comparisons on 120, 240, 720 and 1440-frames long videos. Helios consistently exceeds existing real-time long video generation methods. “

\uparrow

” denotes higher is better.

Figure 13: Qualitative comparisons on 120, 240, 720 and 1440-frames long videos (Part-1). It is clear that Helios consistently outperforms the baseline models in terms of realism and naturalness.

Figure 14: Qualitative comparisons on 120, 240, 720 and 1440-frames long videos (Part-2). It is clear that Helios consistently outperforms the baseline models in terms of text alignment and dynamics.

Panel of Figure 15 — Figure 15: Side-by-side human evaluation of Helios versus counterparts. Left: Long Video; Right: Short Video

Model	Total $\uparrow$	Aesthetic $\uparrow\downarrow$	Dynamic $\uparrow\downarrow$	Smoothness $\uparrow\downarrow$	Semantic $\uparrow$	Naturalness $\uparrow$	Drifting Aesthetic $\uparrow$	Drifting Smoothness $\uparrow$	Drifting Semantic $\uparrow$	Drifting Naturalness $\uparrow$
Helios-Base	6.47	8	6	9	6	5	7	10	8	5
w Guidance Attention*	unstable training process
w/o Guidance Attention	6.23	9	4	7	7	5	6	10	8	2
w/o First Frame Anchor	5.51	8	5	8	7	4	3	10	6	2
w/o Frame-Aware Corrupt	4.70	7	4	8	6	4	2	10	3	1
Helios-Distilled	6.34	8	6	10	5	5	7	10	7	7

Table 5: Quantitative ablation of key components. Changing any component leads to a significant degradation in quality while exacerbating temporal drifting, thereby impairing Helios ’s ability to generate stable long videos. w Guidance Attention* indicates the additional use of a causal mask for self-attention. w Staged Backward Simulation* means incorporating multi-scale

x_{0}^{k}

into real/fake-score estimators during training.

Figure 16: Qualitative ablation on Guidance Attention with Helios -Base. Introducing causal mask hinders the ability to learn temporal coherence, causing each generated section to appear independent. Conversely, removing Guidance Attention results in excessive semantic accumulation over time (e.g., a progressively enlarged bird crest). * indicates the additional usage of causal mask to self-attention to prevent noisy context from affecting historical context.

Figure 17: Qualitative ablation on First Frame Anchor and Frame-Aware Corrupt with Helios -Base. Removing the First Frame Anchor not only introduces drifting but also causes the subject to deviate from the one in the initial frame. Removing Frame-Aware Corruption leads to noticeable drifting even in videos as short as 240 frames.

Figure 18: Qualitative ablation on Stage Backward Simulation and Pure Teacher Forcing with Helios -Distilled. Feeding multi-scale $x_{0}^{k}$ into the fake-score estimator $p_{fake}$ causes the model to converge toward incorrect directions. Pure Teacher Forcing achieves comparable robustness against long-video drifting as Self-Forcing [ 34 ] . * indicates that multi-scale $x_{0}^{k}$ , rather than only the full-resolution $x_{0}^{K}$ , is provided to the real/fake-score estimator.

Figure 19: Qualitative ablation on Coarse-to-Fine Learning and Adversarial Post-Training with Helios -Distilled. Removing Coarse-to-Fine Learning prevents the model from converging, with particularly unacceptable quality in the first generated section. Removing Adversarial Post-Training leads to degradation in visual quality.

Figure 20: Qualitative ablation on Decouple DMD [ 54 ] and Reward-weighted Regression [ 60 ] with Helios -Distilled. The former may hinder convergence and cause grayish outputs, whereas the latter may intensify video flickering.

Figure 21: Text-to-Video showcases. Part of the prompts are sourced from [ 34 ] .

Figure 22: Image-to-Video showcases. Part of the images are sourced from [ 35 , 45 ] .

Figure 23: Interactive-to-Video showcases. Part of the prompts are sourced from [ 100 , 91 ] .

	Inference Time (s)	Training Time (s)
Wan-2.1-T2V-14B [ 90 ]	98.68	398.03
w Flash Normalization	89.91	360.77
w Flash RoPE	93.39	378.77
w Flash Normalization and Flash RoPE	84.41	340.38

Table 6: Quantitative ablation on Flash Normalization and Flash RoPE. We report the total runtime of the DiT component in Helios, measured over 50 forward passes and 50 forward-backward passes on a single NVIDIA H100 GPU.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Metric ( $k$ )	Type	$\bm{\tau_{0}}$	$\bm{\tau_{1}}$	$\bm{\tau_{2}}$	$\bm{\tau_{3}}$	$\bm{\tau_{4}}$	$\bm{\tau_{5}}$	$\bm{\tau_{6}}$	$\bm{\tau_{7}}$	$\bm{\tau_{8}}$
Aesthetic	Higher Better	0.70	0.65	0.60	0.55	0.50	0.45	0.40	0.35	0.30
Motion Amplitude	Higher Better	0.45	0.40	0.35	0.30	0.25	0.20	0.15	0.10	0.05
Motion Smoothness	Higher Better	0.99	0.98	0.97	0.96	0.95	0.94	0.93	0.92	0.91
Semantic	Higher Better	0.30	0.29	0.28	0.27	0.26	0.25	0.24	0.23	0.22
Naturalness	Higher Better	0.65	0.60	0.55	0.50	0.45	0.40	0.30	0.25	0.20
Drifting Aesthetic	Lower Better	0.01	0.02	0.03	0.04	0.05	0.06	0.07	0.08	0.09

Appendix

Table 7: Empirical Thresholds (

\mathcal{T}_{k}

) for Metric Discretization. For the Aesthetic metric, the final two thresholds are marked with an asterisk (*) to indicate a potential anomaly in the empirical data or source configuration.

Efficient Video Generation

See all 25 papers

Jul 29, 2026

Visko Orbis 1.0: A Live Model for Real-Time Interactive Long Video Generation

We present Visko Orbis 1.0, a Live Model for real-time, interactive long video generation. Users can change the prompt at any moment during generation, and the update becomes visible in real time. Visko Orbis 1.0 supports long-form text-to-video, image-to-video, and video continuation, with multilingual prompts and prompt switching while generation is in progress. A bounded multi-scale memory preserves subjects, scenes, and style across chunks, sustaining hour-scale rollouts without evident quality or color drift. The generator is factorized causally in time, matching the causal structure of physical dynamics, and is aligned with a latent world-model reward for predictive consistency. Built on a distilled chunk-wise streaming generator and a streaming video upscaler, Visko Orbis 1.0 delivers 4K video generation at 24 FPS in real time, using an optimized GPU serving engine. In quantitative evaluations, Visko Orbis 1.0 achieves the best DOVER aesthetic and technical scores and the best VideoAlign visual and motion quality, and leads three physical-plausibility protocols (VideoPhy-2, Physics-IQ, and VBench-2.0 Physics); in long-form Arena comparisons, it obtains the highest overall-preference and temporal-stability ratings among all the state-of-the-art real-time interactive video generation systems.

Xiangbo Gao, Siyuan Yang, Ping He +12

Controllable Video Generation Physical Consistency in Video Generation Real-Time Interactive Video Generation Long-Horizon Video Generation Memory-Augmented Video Generation Streaming Video Generation+2

Ask PDF

Aug 6, 2026

Diff-VF: Training-free High-quality Long Video Generation via Diffusion Model

Recently, diffusion models have made great progress in video generation. However, most existing video diffusion models are trained with short videos, and degrade when extrapolated to long videos, struggling to maintain long-range temporal coherence while retaining diverse motions. To generate consistent, high-quality and dynamic long videos, we propose Diff-VF, a training-free, plug-and-play and model-agnostic framework that converts existing short-video diffusion backbones into long-video generators without modifying or fine-tuning the base model. Diff-VF couples three complementary strategies: Hybrid Noise Initialization (HNI) to constrain global semantics, Weighted Window Sampling (WWS) to remove inter-window discontinuities, and Temporal Extended Sampling (TES) to establish long-range dependencies with a timestep-varying fusion. We further extend Diff-VF to long-video enhancement via Skip Residual Guidance that balances fidelity and realism through timestep-dependent guidance. VBench-Long evaluation results show that Diff-VF achieves a more favorable balance between temporal coherence and motion diversity than base models and recent training-free long video generation baselines, including FreeNoise, FreeLong, and RIFLEx, while maintaining competitive frame-wise quality. Experiments on two base models demonstrate the applicability to video diffusion models with different spatial-temporal modeling strategies. Extensive ablations validate the contribution of each component and hyperparameters.

Haoning Yang, Xinyuan Chen, Yaohui Wang +1

Shanghai Jiao Tong University, Shanghai, China · Shanghai Artificial Intelligence Laboratory, Shanghai, China

Diffusion Model Guidance Video Diffusion Models Long Video Generation Temporal Consistency in Video Generation

Ask PDF

Sep 29, 2026

LongLive-Plug: Once-for-All Distillation for Video Generation

Video diffusion models are increasingly developed into specialized models for diverse downstream tasks, and this development often includes a distillation stage, for example to accelerate sampling or to improve long-video generation. This stage is typically repeated for every specialized model. We introduce LongLive-Plug, a once-for-all distillation framework that learns reusable capabilities as LoRAs on a base model for training-free, plug-and-play deployment to compatible downstream models. These capabilities include single-pass classifier-free guidance, few-step sampling, and long-context error correction for autoregressive generation. The adapters remain reusable even when downstream models add conditioning branches, expand output channels. Despite training at a fixed guidance scale, our dedicated CFG LoRA provides text guidance control through its inference weight. Combining it with a few-step LoRA simultaneously preserves few-step generation and CFG controllability on downstream tasks. We verify training-free deployment on 54 downstream models across three backbone families and eight task categories, including world modeling, robotics, editing, and multimodal generation. The approach may support additional compatible models. Each capability can thus be distilled once per backbone family and reused without per-target retraining.

Shuai Yang, Luozhou Wang, Wei Huang +9

NVIDIA

Video Diffusion Models Diffusion Model Distillation Diffusion Model Inference Acceleration

Ask PDF

Abstract

Efficient Video Generation

Explore similar work

Visko Orbis 1.0: A Live Model for Real-Time Interactive Long Video Generation

Diff-VF: Training-free High-quality Long Video Generation via Diffusion Model

LongLive-Plug: Once-for-All Distillation for Video Generation