InfinityStar: Unified Spacetime AutoRegressive Modeling for Visual Generation

TL;DR

InfinityStar (ByteDance / FoundationVision, NeurIPS 2025 Oral) is a purely discrete, bitwise spacetime autoregressive framework that unifies text-to-image, text-to-video, image-to-video, and long interactive video synthesis under a single architecture, extending the VAR (next-scale prediction) and Infinity (bitwise AR image) lineage from images to video by jointly factoring spatial and temporal dependencies. The model is the first discrete autoregressive video generator to produce industrial-level 720p video and scores 83.74 on VBench — outperforming all prior autoregressive video models and surpassing the diffusion-based HunyuanVideo (83.24). With the same compression-rate tokenizer, it generates a 5-second 720p video roughly 10× faster than leading diffusion baselines. Code, weights, training scripts and 720p / 480p / 480p-interactive inference are all released under MIT.

Key claims

A single purely-discrete autoregressive architecture jointly captures spatial and temporal dependencies and natively supports text-to-image, text-to-video, image-to-video, and long interactive video synthesis without task-specific heads [Abstract, §1].
InfinityStar reaches 83.74 on VBench, a new state-of-the-art among autoregressive video generators and above diffusion-based HunyuanVideo (83.24) [Abstract, §1].
At matched visual-tokenizer compression rate, InfinityStar achieves roughly 10× lower inference latency than leading diffusion video models for a 5-second 720p generation [Abstract, §1].
It is the first discrete autoregressive video model to produce industrial-level 720p video [Abstract].
Long interactive video synthesis is supported via straightforward temporal autoregression over the spacetime token sequence — no separate streaming-tuning recipe is required, the inference script infer_interact_480p.py accepts a reference video plus multiple prompts and continues the video interactively [README, §Inference].
The released 720p model is trained for 5-second clips; an intermediate 480p checkpoint supports variable-length 5s / 10s generation and is the recommended path for image-to-video and video-to-video continuation [README, §Inference].
Training uses FlexAttention (torch ≥ 2.5.1) to make the bitwise spacetime autoregressive objective tractable at scale [README, §Installation].

Method

InfinityStar inherits the bitwise next-scale prediction paradigm of Infinity (arXiv 2412.04431) — itself an evolution of VAR’s next-scale prediction (arXiv 2404.02905) — and extends it from a purely spatial scale-pyramid to a spacetime pyramid that simultaneously refines spatial scales and unrolls along time. Each generation step predicts the next set of discrete bitwise tokens conditioned on all previously generated spacetime tokens; the same transformer trunk handles all four task modes by varying the conditioning prefix (text only for T2V, text + first-frame tokens for I2V, text + previous-clip tokens for interactive continuation). The model is trained on large-scale video corpora at multiple resolutions, with the 720p checkpoint specialized to 5-second clips and an intermediate 480p checkpoint trained to handle longer (5s / 10s) durations and V2V continuation. Inference is sped up by the discrete-AR formulation itself — no iterative denoising loop — plus FlexAttention kernels for the dense spatial-temporal attention. The repo releases three inference paths: 720p T2V/I2V (infer_video_720p.py), 480p variable-length T2V/I2V/V2V (infer_video_480p.py), and 480p long interactive (infer_interact_480p.py).

Results

VBench Total: 83.74 vs. HunyuanVideo 83.24 — first AR video model to surpass HunyuanVideo on the standard benchmark [Abstract, §1].
Inference speed: ~10× faster than leading diffusion video models on 5s × 720p generation at matched tokenizer compression [Abstract, §1].
Image-generation track also reports SOTA among bitwise-AR image models on the InfinityStar image-gen benchmark figure [README, “Benchmark” section]; specific numbers not extracted here.
Human-evaluation comparison vs. HunyuanVideo reported in the InfinityStar video-gen humaneval figure [README]; specific win-rates not extracted.
Quantitative claims here are limited to what is stated in the abstract and README; the wiki has not extracted Table-level breakdowns from the PDF (fetch timed out at filing time).

Why it’s interesting

InfinityStar is the cleanest argument yet that discrete autoregressive video generation can compete with diffusion at frontier scale — it both beats HunyuanVideo on VBench and runs 10× faster, removing the standard “AR is slower than diffusion at video” objection that has held up since SORA. It sits in direct opposition to the diffusion-based streaming-AR lineage tracked under Autoregressive Video Generation (Context Forcing: Consistent Autoregressive Video Generation with Long Context, MALT Diffusion: Memory-Augmented Latent Transformers for Any-Length Video Generation, Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation, Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation) — those papers all start from a diffusion teacher and distill it into an AR student, paying the asymmetric-distillation tax that Causal Forcing identifies; InfinityStar simply trains the AR objective end-to-end on discrete bitwise tokens and skips the diffusion teacher entirely. It also continues the VAR → Infinity → InfinityStar arc of bitwise next-scale prediction, complementing the continuous-AR-with-FM-head wing of Unified Multimodal Models (NextStep-1: Toward Autoregressive Image Generation with Continuous Tokens at Scale, Marrying Autoregressive Transformer and Diffusion with Multi-Reference Autoregression) by showing the discrete-token wing can scale to video at competitive quality. The unification of T2I, T2V, I2V, and interactive continuation in one architecture is also notable: most filed unified-MM systems unify understanding + generation across modalities, whereas InfinityStar unifies generation tasks within the visual modality, with the long-interactive mode functioning as a low-latency world-model-style sampler.