5分钟视频infer

[ifzhang]

请问一下如果infer 5分钟的视频是不是准备好prompt之后，需要把num_frames设置成7200？

[Jahnsonblack]

您好，不需要把 num_frames 设置成 7200。

在当前官方推理脚本中，num_frames 表示单个 shot 的帧数，不是整段视频的总帧数。5 分钟视频是通过 multi-shot 方式生成的：准备多个 shot prompts，每个 shot 生成一段短视频，最后再拼接成完整视频。

默认配置为 num_frames=241, fps=25，每个 shot 大约 9.6 秒。因此 5 分钟视频通常需要约 30 个 shot prompts。

如果需要不同 shot 使用不同长度，也可以修改推理代码，让每个 shot 单独指定自己的 num_frames。

[walkingwithGod2017]

您好，不需要把 num_frames 设置成 7200。

在当前官方推理脚本中，num_frames 表示单个 shot 的帧数，不是整段视频的总帧数。5 分钟视频是通过 multi-shot 方式生成的：准备多个 shot prompts，每个 shot 生成一段短视频，最后再拼接成完整视频。

默认配置为 num_frames=241, fps=25，每个 shot 大约 9.6 秒。因此 5 分钟视频通常需要约 30 个 shot prompts。

如果需要不同 shot 使用不同长度，也可以修改推理代码，让每个 shot 单独指定自己的 num_frames。

Thank you for your research! I was wondering if it's possible to specify a particular character's appearance, for example, by loading a character LoRA? Also, when generating longer videos, will VRAM and system RAM usage increase? Looking forward to your reply!

[Jahnsonblack]

Hi, thanks for your interest!

Currently, JoyAI-Echo does not support loading external character LoRAs directly.

For character consistency, the current official pipeline mainly relies on the paired audio-video memory mechanism. In this sense, after the first shot is generated, later shots can be viewed as memory-to-video generation conditioned on previous visual/audio memory. If you only want to use visual memory, the audio memory can also be replaced with empty/silent audio in a customized pipeline.

Regarding VRAM/RAM usage, it will not grow indefinitely with the final video length. The resource usage is mainly related to the per-shot generation setting and the memory bank size. In our official code, the default maximum memory size is 7, so old memory entries are trimmed instead of accumulating forever.

Thanks again!