Stable diffusion dreams of psychedelic faces
每日信息看板 · 2026-03-03
2022-08-19T18:32:25+00:00
Published
AI 总结
YouTube 视频展示用 Stable Diffusion 对“psychedelic faces”提示词进行噪声向量球面插值生成连贯过渡画面，说明扩散模型可通过潜空间插值制作长时序生成视频并给出代码参考。
- 使用提示词“psychedelic faces”，通过 Stable Diffusion 从噪声向量采样生成图像
- 将随机噪声向量进行平滑（球面）插值，并沿插值路径逐帧渲染形成视频
- 制作耗时约 8 小时，使用单张 A100 GPU 过夜生成
- 提供用于生成该视频的代码 Gist 链接与 Stable Diffusion 官方博客链接
- 背景音乐为 JVNA 的《Stars》
#YouTube #视频/演讲 #A100 GPU
内容摘录
Prompt: "psychedelic faces"
Stable diffusion takes a noise vector as input and samples an image. To create this video I smoothly (spherically) interpolate between randomly chosen noise vectors and render frames along the way.
This video was produced by one A100 GPU taking about 10 tabs and dreaming about the prompt overnight (~8 hours). While I slept and dreamt about other things.
Music: Stars by JVNA
Links:
- Stable diffusion: https://stability.ai/blog
- Code used to make this video: https://gist.github.com/karpathy/00103b0037c5aaea32fe1da1af553355
- My twitter: https://twitter.com/karpathy