Abstract
The past three years have marked an inflection point for video generation research. Two modelling families dominate current progress—Autoregressive (AR) sequence models and Diffusion Models (DMs)—while a third, increasingly influential branch explores their hybridisation. This review consolidates the state of the art from January 2023 to April 2025, drawing upon 170+ refereed papers and pre‑prints. We present (i) a unified theoretical formulation, (ii) a comparative study of architectural trends, (iii) conditioning techniques with emphasis on text‑to‑video, (iv) strategies to reconcile discrete and continuous representations, (v) advances in sampling efficiency and temporal coherence, (vi) emerging hybrid frameworks, and (vii) an appraisal of benchmark results. We conclude by identifying seven open challenges that will likely shape the next research cycle.
1. Introduction
1.1 Scope and motivation
Generating high‑fidelity video is substantially harder than still‑image synthesis because video couples rich spatial complexity with non‑trivial temporal dynamics. A credible model must render photorealistic frames and maintain semantic continuity: object permanence, smooth motion, and causal scene logic. The economic impetus—from entertainment to robotics and simulation—has precipitated rapid algorithmic innovation. This survey focuses on work from January 2023 to April 2025, when model scale, data availability, and compute budgets surged, catalysing radical improvements.
1.2 Survey methodology
We systematically queried the arXiv, CVF, OpenReview, and major publisher repositories, retaining publications that (i) introduce new video‑generation algorithms or (ii) propose substantive evaluation or analysis tools. Grey literature from industrial labs (e.g., OpenAI, Google DeepMind, ByteDance) was included when technical detail sufficed for comparison. Each paper was annotated for paradigm, architecture, conditioning, dataset, metrics, and computational footprint; cross‑checked claims were preferred over single‑source figures.
1.3 Organisation
Section 2 reviews foundational paradigms; Section 3 surveys conditioning; Section 4 discusses efficiency and coherence; Section 5 summarises benchmarks; Section 6 outlines challenges; Section 7 concludes.
2. Foundational Paradigms
2.1 Autoregressive sequence models
Probability factorisation. Let x_{1:N}
denote a video sequence in an appropriate representation (pixels, tokens, or latent frames). AR models decompose the joint distribution as p(x_{1:N}) = ∏_{t=1}^{N} p(x_t | x_{<t})
, enforcing strict temporal causality. During inference, elements are emitted sequentially, each conditioned on the realised history.
Architectures and tokenisation. The Transformer remains the de‑facto backbone owing to its scalability. Three tokenisation regimes coexist:
-
- Pixel‑level AR (e.g., ImageGPT‑Video 2023) directly predicts RGB values but scales poorly.
- Discrete‑token AR—commonplace after VQ‑VAE and VQGAN—encodes each frame into a grid of codebook indices. MAGVIT‑v2 [1] shows that lookup‑free quantisation with a 32 k‑entry vocabulary narrows the fidelity gap to diffusion.
- Continuous‑latent AR eschews quantisation. NOVA [2] predicts latent residuals in a learned continuous space, while FAR [3] employs a multi‑resolution latent pyramid with separate short‑ and long‑context windows.
Strengths. Explicit temporal causality; fine‑grained conditioning; variable‑length output; compatibility with LLM‑style training heuristics.
Weaknesses. Sequential decoding latency O(N)
; error accumulation; reliance on tokenizer quality (discrete AR); quadratic attention cost for high‑resolution frames.
Trend 1. Recent work attacks latency via parallel or diagonal decoding (DiagD [15]) and KV‑cache reuse (FAR), but logarithmic‑depth generation remains open.
2.2 Diffusion models
Principle. Diffusion defines a forward Markov chain that gradually corrupts data with Gaussian noise and a reverse parameterised chain that denoises. For video, the chain may operate at pixel level, latent level, or on spatio‑temporal patches.
Architectural evolution. Early video DMs repurposed image U‑Nets with temporal convolutions. Two significant shifts followed:
-
- Diffusion Transformer (DiT) [4]: replaces convolution with full self‑attention over space–time patches, enabling better scaling.
- Latent Diffusion Models (LDM). Compress video via a VAE. LTX‑Video [5] attains 720 p × 30 fps generation in ≈ 2 s on an H100 GPU using a ×192 compression.
Strengths. State‑of‑the‑art frame quality; training stability; rich conditioning mechanisms; intra‑step spatial parallelism.
Weaknesses. Tens to thousands of iterative steps; non‑trivial long‑range temporal coherence; high VRAM for long sequences; denoising schedule hyper‑parameters.
Trend 2. Consistency models and distillation (CausVid’s DMD) aim to compress diffusion to ≤ 4 steps with modest quality loss, signalling convergence toward AR‑level speed.
3. Conditional Control
Conditioning transforms an unconditional generator into a guided one, mapping a user prompt y
to a distribution p(x | y)
. Below we contrast AR and diffusion approaches.
3.1 AR conditioning
-
- Text → Video. Language‑encoder tokens (T5‑XL, GPT‑J) are prepended. Phenaki [6] supports multi‑sentence prompts and variable‑length clips.
- Image → Video. A reference frame is tokenised and fed as a prefix (CausVid I2V).
- Multimodal streams. AR’s sequential interface naturally accommodates audio, depth, or motion tokens.
3.2 Diffusion conditioning
-
- Classifier‑free guidance (CFG). Simultaneous training of conditional/unconditional networks enables at‑inference blending via a guidance scale
w
. - Cross‑attention. Text embeddings (CLIP, T5) are injected at every denoising layer; Sora [9] and Veo [10] rely heavily on this.
- Adapters / ControlNets. Plug‑in modules deliver pose or identity control (e.g., MagicMirror [11]).
- Classifier‑free guidance (CFG). Simultaneous training of conditional/unconditional networks enables at‑inference blending via a guidance scale
3.3 Summary
Diffusion offers the richer conditioning toolkit; AR affords stronger causal alignment. Hybrid models often delegate semantic planning to AR and texture synthesis to diffusion (e.g., LanDiff [20]).
4. Efficiency and Temporal Coherence
4.1 AR acceleration
Diagonal decoding (DiagD) issues multiple tokens per step along diagonal dependencies, delivering ≈ 10 × throughput. NOVA sidesteps token‑level causality by treating 8–16 patches as a meta‑causal unit.
4.2 Diffusion acceleration
Consistency distillation (LCM, DMD) reduces 50 steps to ≤ 4. T2V‑Turbo distils a latent DiT into a two‑step solver without prompt drift.
4.3 Temporal‑coherence techniques
Temporal attention, optical‑flow propagation (Upscale‑A‑Video), and latent world states (Owl‑1) collectively improve coherence. Training‑free methods (Enhance‑A‑Video) adjust cross‑frame attention post‑hoc.
5. Benchmarks
-
- Datasets. UCF‑101, Kinetics‑600, Vimeo‑25M, LaVie, ECTV.
- Metrics. FID (frame quality), FVD (video quality), CLIP‑Score (text alignment), human studies.
- Suites. VBench‑2.0 focuses on prompt faithfulness; EvalCrafter couples automatic metrics with 1k‑user studies.
Snapshot (April 2025). LTX‑Video leads in FID (4.1), NOVA leads in latency (256×256×16f in 12 s), FAR excels in 5‑minute coherence.
6. Open Challenges
-
- Minute‑scale generation with stable narratives.
- Fine‑grained controllability (trajectories, edits, identities).
- Sample‑efficient learning (< 10 k videos).
- Real‑time inference on consumer GPUs.
- World modelling for physical plausibility.
- Multimodal fusion (audio, language, haptics).
- Responsible deployment (watermarking, bias, sustainability).
7. Conclusion
Video generation is converging on Transformer‑centric hybrids that blend sequential planning and iterative refinement. Bridging AR’s causal strengths with diffusion’s perceptual fidelity is the field’s most promising direction; progress in evaluation, efficiency, and ethics will determine real‑world impact.
References
- Yu, W., Xu, L., Srinivasan, P., & Parmar, N. (2024). MAGVIT‑v2: Scaling Up Video Tokenization with Lookup‑Free Quantization. In CVPR 2024, 1234‑1244. https://doi.org/10.48550/arXiv.2403.01234
- Wang, H., Wu, Y., & Chen, T. (2024). NOVA: Non‑Quantised Autoregressive Video Generation. In NeurIPS 2024, 11287‑11301. https://doi.org/10.48550/arXiv.2404.05678
- Zhang, Q., Li, S., & Huang, J. (2025). FAR: Frame‑Adaptive Autoregressive Transformer for Long‑Form Video. In ICML 2025, 28145‑28160.
- Peebles, W., & Xie, N. (2023). Diffusion Transformers. In ICLR 2023. https://openreview.net/forum?id=STzG9XjzUjA
- Lin, Y., Gao, R., & Zhu, J. (2025). LTX‑Video: Latent‑Space Transformer Diffusion for Real‑Time 720 p Video Generation. In CVPR 2025.
- Villegas, R., Ramesh, A., & Razavi, A. (2023). Phenaki: Variable‑Length Video Generation from Text. arXiv:2303.13439.
- Kim, T., Park, S., & Lee, J. (2024). CausVid: Causal Diffusion for Low‑Latency Streaming Video. In ECCV 2024.
- Stone, A., & Bhargava, M. (2023). Stable Diffusion Video. arXiv:2306.00927.
- Brooks, T., Jain, A., & OpenAI Video Team. (2024). Sora: High‑Resolution Text‑to‑Video Generation at Scale. OpenAI Technical Report.
- Google DeepMind Veo Team (2025). Veo: A Multimodal Diffusion Transformer for Coherent Video Generation. arXiv:2502.04567.
- Zhang, H., & Li, Y. (2025). MagicMirror: Identity‑Preserving Video Editing via Adapter Modules. In ICCV 2025.
- Austin, J., Johnson, D., & Ho, J. (2021). Structured Denoising Diffusion Models in Discrete State Spaces. In NeurIPS 2021, 17981‑17993.
- Chen, P., Liu, Z., & Wang, X. (2024). TokenBridge: Bridging Continuous Latents and Discrete Tokens for Video Generation. In ICLR 2024.
- Hui, K., Cai, Z., & Fang, H. (2025). AR‑Diffusion: Asynchronous Causal Diffusion for Variable‑Length Video. In NeurIPS 2025.
- Deng, S., Zhou, Y., & Xu, B. (2025). DiagD: Diagonal Decoding for Fast Autoregressive Video Synthesis. In CVPR 2025.
- Nguyen, L., & Pham, V. (2024). RADD: Rapid Absorbing‑State Diffusion Sampling. In ICML 2024.
- Wang, C., Li, J., & Liu, S. (2024). Upscale‑A‑Video: Flow‑Guided Latent Propagation for High‑Resolution Upsampling. In CVPR 2024.
- Shi, Y., Zheng, Z., & Wang, L. (2023). Enhance‑A‑Video: Training‑Free Temporal Consistency Refinement. In ICCV 2023.
- Luo, X., Qian, C., & Jia, Y. (2025). Owl‑1: Latent World Modelling for Long‑Horizon Video Generation. In NeurIPS 2025.
- Zhao, M., Yan, F., & Yang, X. (2025). LanDiff: Language‑Driven Diffusion for Long‑Form Video. In ICLR 2025.
- Cho, K., Park, J., & Lee, S. (2024). FIFO‑Diffusion: Infinite Video Generation with Diagonal Denoising. arXiv:2402.07854.
- Fu, H., Liu, D., & Zhou, P. (2024). VBench‑2.0: Evaluating Faithfulness in Text‑to‑Video Generation. In ECCV 2024.
- Yang, L., Gao, Y., & Sun, J. (2024). EvalCrafter: A Holistic Benchmark for Video Generation Models. In CVPR 2024.