工具

影片生成

影片生成和具備影片能力的模型工作流程。

影片生成

這是什麼

影片生成和具備影片能力的模型工作流程。

白話解釋

外包動畫製作與外包靜態圖像的原理相同,只是用影片工作室代替了插畫家。Hermes 交出書面簡報——並可選擇性地提供起始圖片或參考片段——然後取回實際的影片素材。由於它本身沒有搭載任何影片模型,使用者必須先選擇並啟用由哪個外部工作室來完成這項工作。

為什麼重要

文字轉影片和圖片(或參考)轉影片的路徑都透過 fal 和 xai 等插件提供者運行,註冊在 video_gen 之下,擁有 3 個自己的宣告。它與單獨的 video 工具集不同,後者處理 video_analyze 並需要具備影片能力的模型,而不是生成新素材。生成成本和延遲使其明確屬於「刻意啟用,而非預設啟用」的類別,與圖片生成相同。

How it works

video_generate ships with no built-in provider at all — it’s a thin dispatcher that forwards to whichever plugin is enabled under plugins/video_gen/ (fal or xai), selected via video_gen.provider in config (tools/video_generation_tool.py). One schema covers text-to-video, image-to-video, and reference-to-video through shared fields — prompt, image_url, reference_image_urls, duration, aspect_ratio, resolution, negative_prompt, audio, seed — and the tool itself only checks that a prompt was supplied, leaving each provider to clamp or ignore whatever fields it doesn’t support. When xai is the active provider, two extra tools unlock: xai_video_edit and xai_video_extend (tools/xai_video_tools.py), which take the public HTTPS MP4 URL from an earlier clip and either restyle or lengthen that exact footage instead of generating something new from scratch.

A concrete example

A marketer wants a short vertical product teaser, then decides it should run a bit longer.

  1. video_generate with a text prompt and aspect_ratio=‘9:16’, provider set to xai, produces the first clip.
  2. They feed that clip’s returned public URL into xai_video_extend, along with a new prompt and a target duration. The second call returns one continuous clip, not a splice.