听听 Krea 怎么处理训练数据和降低视频生成成本。
Krea 故意在训练集中移除了 AI 图像，因为蒸馏数据会导致模型走捷径。团队运行了 30 到 40 个内部分类器处理数十亿张图像，并自动替换温度超过 78 度的 GPU。专为小模型设计的框架让开源用户在 90 天内渲染了 130 万个视频。连续视频生成的成本降至 10 美元 3 小时，而旧方式约为 10 美元每分钟。
Two halves of one problem. What you feed the model, and what you do with what comes out: - Krea thr...
Two halves of one problem. What you feed the model, and what you do with what comes out: - Krea threw AI images out of the training set on purpose, because distilled data is sticky and the shortcut costs you the whole point. Roughly 30 to 40 in house classifiers run over billions of images - GPU utilization is a lie. It reads 100 percent while the tensor cores sit idle. Any GPU running above 78 degrees gets pulled rather than debugged - 1.3 million videos rendered by open source users in 90 days, from a framework deliberately designed so a small model could author it - An avatar generating continuously for eight hours with no reset, and a sixteen hour deployment being built. It costs about what serving a voice model costs - 10 dollars now buys three hours of continuously generated video. The old way was a slot machine at roughly 10 dollars a minute Watch the Generative Media track: youtube.com/watch?v=-tviRd… 💬 0 🔄 0 ❤️ 0 👀 317 ⚡