开源权重模型，14B参数就能生成1小时连贯视频，还带1.3B小版本跑在单显卡上，长视频生成终于不崩了。
LingBot-World-Infinity 是一个开权重世界模型，参数规模14B，另有1.3B变体可在单消费级GPU运行。该模型通过让自身在训练时主动进入错误状态，再学习如何恢复，从而生成长达1小时以上且跨场景连贯的视频，克服了传统逐帧模型误差累积的问题。论文详细描述了其自我纠错训练方法，并已公开GitHub仓库和模型权重。
This model can generate coherent 1+ hour videos ac…
This model can generate coherent 1+ hour videos across multiple scenes without skipping a beat.
I read their paper so you don't have to. Here is how it works:
First, this is an open-weight model. You can find the GitHub repo and link to the model weights below.
Most video models generate their videos one frame at a time. They look at previous frames to decide what comes next.
Unfortunately, if there's a mistake in one frame, every subsequent frame will compound the error. It gets ugly really quickly.
@robbyant_brain's new world model, LingBot-World-Infinity, works differently. Here are the main highlights:
• The model practices using its own mistakes. Since training only shows the model clean footage, it learns to generate videos that look perfect but never learns to recover when errors show up. Instead, this model runs on its own output for long stretches, letting it drift into the same flawed states it hits in real use.
• Then, the model gets shown the way back. From each of those flawed states, a slower and more careful version of the model demonstrates what a clean continuation looks like. The fast model adjusts its frames to match. It practices the recovery over and over, always starting from its own mess.
• The result is a model that steers itself back to center. Wherever it drifts, it knows the way back, because that recovery is exactly what it trained on. This is what lets it hold together for an hour instead of melting after a minute.
Alongside the 14B primary model, the paper describes a 1.3B variant designed to run on a single consumer GPU.
See links below: