Helios: Real Real-Time Long Video Generation Model

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.

• The paper introduces a 14B-parameter autoregressive video diffusion transformer that achieves real-time minute-scale video generation at 19.5 FPS on a single NVIDIA H100 GPU.
• It employs a unified architecture with Guidance Attention, Multi-Term Memory Patchification, and Pyramid Predictor Corrector to mitigate long-video drifting and ensure robust temporal fidelity.
• The model leverages adversarial hierarchical distillation to reduce sampling steps from 50 to 3, outperforming baselines in visual quality, motion realism, and efficiency.

Helios: A 14B Real-Time Long Video Generation Model

Introduction and Contextual Motivation

Helios represents a significant advance in generative video models, introducing the first 14B-parameter autoregressive video diffusion transformer capable of real-time, minute-scale generation at 19.5 FPS on a single NVIDIA H100 GPU, with quality matching strong baselines at much lower computational overhead. The model addresses three primary bottlenecks: long-video drifting (where the generated content progressively diverges from the intended distribution), real-time generation (without reliance on standard acceleration techniques such as KV-cache, causal masking, or sparse attention), and efficient training (eschewing parallelism/sharding and supporting image-diffusion-scale batch sizes).

The architecture is explicitly designed for optimal scaling on long sequences, robust temporal fidelity, and universal task-format (supporting T2V, I2V, V2V), meeting the emergent requirements of interactive generative engines and world simulators.

Figure 1: End-to-end throughput (FPS) of Helios compared to prior models, highlighting superior speed at scale.

Architectural Design and Unified Framework

Helios is constructed as a video diffusion transformer employing the Guidance Attention mechanism, Unified History Injection, Representation Control, Multi-Term Memory Patchification, and a Pyramid Unified Predictor Corrector. The model seamlessly unifies T2V, I2V, and V2V modalities through its input format, enabling zero-shot task transition.

Guidance Attention ensures that historical (already generated) context is strictly separated from noisy (to-be-generated) context, enabling robust temporal coherence while avoiding semantic and visual drift. Representation Control allows flexible and efficient task switching—text-to-video if the history is zero, image-to-video if the last frame is nonzero, and video-to-video otherwise.

Multi-Term Memory Patchification compresses historical context hierarchically, preserving essential local and global cues while maintaining a constant token budget as context length increases. This approach radically improves scalability, throughput, and memory efficiency.

Pyramid Unified Predictor Corrector reduces redundancy in the noisy context by progressive multi-scale flow matching, operating from low-resolution global structure (high noise) to full-resolution refinement (low noise). The sampling trajectory integrates ODE-based velocity fields across scales, directly addressing the token and compute bottleneck.

Figure 2: Architecture of Helios, showcasing Guidance Attention blocks, Multi-Term Memory Patchification, Pyramid Predictor Corrector, and unified representation.

Drifting Mitigation and Sampling Efficiency

Helios introduces "Easy Anti-Drifting" strategies targeting canonical drifting modes: position shift, color shift, and restoration shift. Through Relative RoPE, the temporal indices are normalized regardless of video length, alleviating repetitive motion and cyclic resets induced by positional encoding. First Frame Anchor stabilizes color and appearance by always retaining the first historical frame, serving as a distribution anchor throughout autoregressive extrapolation. Frame-Aware Corrupt simulates realistic history drift during training, perturbing historical frames with a range of exposure, noise, and downsampling artifacts, thereby hardening the model against accumulation errors.

A salient feature is that Helios achieves this robustness without self-forcing, error-banks, keyframe/inverted sampling, and without rollouts during training—in stark contrast to existing autoregressive video models.

Hierarchical Distillation and Memory Optimization

For efficient distillation, Helios adopts Adversarial Hierarchical Distillation—a DMD-based pipeline employing pure teacher forcing with only the autoregressive teacher (Helios-Base), staged backward simulation, coarse-to-fine learning, and adversarial post-training via GAN objectives. This reduces sampling steps from 50 to 3 without sacrificing quality, and outperforms conventional rollout-based training in drift mitigation.

Figure 3: Adversarial Hierarchical Distillation pipeline based on DMD, enabling rapid and stable few-step generation.

Memory and throughput are maximized via sharded EMA, asynchronous VRAM freeing, and "Cache Grad for GAN", which caches discriminator gradients to decouple backpropagation and free intermediate activations early. With Flash Normalization and Flash RoPE (Triton kernel fusion), the transformer backbone is further accelerated.

Figure 4: Execution of Cache Grad for GAN, demonstrating efficient peak memory reduction via early activation release.

Empirical Results and Comparison

Across HeliosBench and comparative benchmarks, Helios consistently matches or outperforms existing distilled and base models in visual quality, semantic alignment, motion dynamics, and naturalness, while achieving an unprecedented speed/quality tradeoff. For short (81-frame) clips, Helios-Distilled yields scores of 6.00 and 19.53 FPS, outpacing FastVideo and TurboDiffusion by 2–3$\times$ in speed and challenging Wan 14B and HV Video in quality.

Figure 5: Benchmark performance of Helios and its counterparts, demonstrating superior short- and long-video generation across all metrics.

For long videos (120–1440 frames), Helios demonstrates state-of-the-art Naturalness and Dynamic scores, with lowest drifting across all categories (Aesthetic, Semantic, and Naturalness). Drifting compensation is achieved without conventional anti-drift heuristics, representing a substantial efficiency improvement.

Figure 6: Infinite video showcases generated by Helios, achieving high visual quality and strong temporal coherence with minimal overhead.

Qualitative comparison further demonstrates that Helios-Distilled, through adversarial post-training and hierarchical distillation, surpasses the upper bound imposed by the teacher while avoiding flickering and identity drift.

Figure 7: Qualitative comparisons on 81-frames short videos—Helios matches or exceeds base models in visual quality and motion realism.

Figure 8: Qualitative comparisons on 120, 240, 720, and 1440-frame long videos—Helios maintains realism and text alignment beyond current baselines.

Side-by-side human evaluations corroborate these findings; Helios is consistently preferred across both short and long video settings.

Figure 9: Human evaluation results—Helios preferred over counterparts for both short and long videos.

Ablation Studies

Critical ablations reveal the necessity of Guidance Attention, First Frame Anchor, Frame-Aware Corrupt, hierarchical patchification, and adversarial post-training. Disabling any leads to pronounced temporal drifting, flickering, degradation in naturalness and semantics, or optimization instability. Notably, causal masking (as in CausVid/Self-Forcing) reduces temporal coherence and makes each section independent. Coarse-to-fine learning ensures stable convergence, while adversarial post-training pushes perceptual fidelity beyond pure distillation.

Figure 10: Guidance Attention ablation—causal mask inhibits temporal coherence, while its removal causes semantic drift.

Figure 11: First Frame Anchor and Frame-Aware Corrupt ablation—removal induces rapid onset of color and identity drift.

Application and Scalability

The unified representation enables seamless support for T2V, I2V, and V2V tasks, as well as interactive generation via prompt interpolation, providing fine-grained control for real-time creativity and world-model applications. The supporting infrastructure allows training at image-diffusion-scale batch sizes and inference at scale, making Helios suitable for game engines, interactive systems, and streaming content generators.

Figure 12: Text-to-video showcases—diverse, spatially and temporally consistent generations.

Figure 13: Image-to-video showcases—robust expansion from still to dynamic content.

Figure 14: Interactive-to-video showcases—smooth transitions and real-time prompt adaptation.

Practical and Theoretical Implications

Helios demonstrates that large-scale autoregressive diffusion transformers for video can be made both computationally tractable and temporally consistent without costly rollouts or anti-drift heuristics. The technical contributions in context compression, token budgeting, hierarchical distillation, and architectural unification substantially advance the practical deployment of long-form video models.

Theoretically, the removal of train-as-infer rollouts decouples drift robustness from rollout length, opening new directions for scalable autoregressive world modeling. The methods for compressing both historical and noisy contexts should generalize to other generative sequence modalities, including audio and 3D worlds. The adversarial distillation pipeline, especially when coupled with coarse-to-fine curriculum and memory optimization, sets a new standard for efficient few-step generation.

Conclusion

Helios introduces a suite of innovations allowing real-time, minute-scale, long-video generation from a unified 14B transformer, achieving unprecedented speed, efficiency, and quality on single GPU hardware. The framework eliminates reliance on standard anti-drift and acceleration techniques, instead providing architectural, training, and inference strategies that are robust and scalable. Empirical, qualitative, and human evaluations confirm Helios’s superiority across tasks, durations, and metrics. The practical implications for video foundation models, world simulators, and interactive generative media are substantial, and its architecture is poised to influence future developments in generative AI.

Future Directions

Though Helios achieves strong results at moderate resolutions, perceptually aligned metrics, higher-resolution scaling, and reinforcement learning for temporal smoothness remain open research avenues. The architecture’s compatibility with broader generative modalities and its implications for long-form content moderation warrant further exploration.