infinitystar.org

What It Is

InfinityStar is a high-resolution image and dynamic video generation framework developed by FoundationVision. The source text describes it as a ByteDance-related project and notes that it was accepted as a NeurIPS 2025 Oral paper. It uses a unified spatiotemporal autoregressive architecture, modeling spatial appearance and temporal motion within the same model. It targets tasks such as text-to-image, text-to-video, image-to-video, long-video generation, and video continuation.

Core Capabilities

The model has 8B parameters, with a checkpoint of around 35GB, and uses Flan-T5-XL as its text encoder. Unlike common diffusion-based video models, InfinityStar adopts a discrete autoregressive approach, treating visual content as a token sequence and predicting it step by step. The capabilities listed in the source include 720p 5-second video generation, 480p 5–10-second variable-length video generation, image-to-video, and video continuation. It reports a VBench score of 83.74 and claims that 720p 5-second generation is about 10 times faster than leading diffusion-based methods.

Pricing and Open Source

The source does not disclose commercial pricing, API billing, or subscription plans. The project emphasizes open source, including training code, inference code, 480p/720p model checkpoints, a web demo, and documentation. Its terms state that it uses the MIT License and can be used for personal, educational, or commercial purposes. That said, open source does not mean low-barrier: the full model is around 35GB, and both inference and training require substantial GPU resources.

Pros and Cons

Its strengths include a unified architecture, broad task coverage, strong short-video generation speed, solid benchmark performance, and open-source availability that benefits research reproduction and further development. The drawbacks are also clear: 720p output is limited to 5 seconds, the 480p model is not specifically optimized for text-to-video generation, it depends on FlexAttention in PyTorch 2.5.1+, and training costs are high. The source also does not provide common commercial product features such as a stable cloud API, enterprise SLA, team collaboration, or copyright review.

Who It’s For and Access from China

InfinityStar is better suited to AI video researchers, engineering-capable development teams, and film, animation, or content teams running proof-of-concept projects. It is less suitable for ordinary creators who just want an out-of-the-box tool. The source does not state how well it can be accessed from China, so connectivity to the domain, model repository, and demo needs to be tested in practice. Payment methods are also not disclosed. If access or compute resources are limited, users may want to compare it with alternatives such as HunyuanVideo, which can be deployed locally or may fit better within the domestic Chinese ecosystem.