flux-3-video

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

FLUX 3 Video Router

FLUX 3 视频请求路由

Route a request to the smallest set of specialists that can complete it. When the request already names one narrow job, use that specialist directly.

User intent	Skill
Diagnose or repair a brief before generation	`flux-3-prompt-doctor`
Write a text-to-video prompt for a new shot	`flux-3-cinematic-inserts`
Build from supplied images or video	`flux-3-keyframes-continuation`
Direct dialogue, voiceover, ambience, effects, or music	`flux-3-audio-dialogue`
Submit, poll, enhance, download, or technically review	`flux-3-generate`

Default sequence, skipping completed stages:

text

Prompt Doctor (until READY; owns the shot-structure decision)
  → Cinematic Inserts OR Keyframes & Continuation (one, chosen by whether
    supplied media must constrain the result)
  → Audio & Dialogue (only when sound matters)
  → Generate (only when execution is requested)

Draft before committing anything unproven: a full render takes minutes, and the chosen draft replays at full quality.

A handoff between skills carries: outcome, the route (

mode

plus its media field:

t2v

i2v

keyframes

v2v

start_video

, or

draft_enhance

draft_cache

), the one required action, source roles, invariants, delivery (duration, aspect, resolution, shot structure), camera, audio, text strategy, and risks. It is complete when the receiving skill can act without inventing a consequential choice.

Deterministic requirements (exact typography, frame-accurate sync, final mixing) need an explicit post-production plan, not a generation attempt.

Reference: docs.bfl.ai

将请求路由至能够完成任务的最小专业人员集合。如果请求已经指定了某一项具体工作，请直接对接该专业人员。

用户意图	技能
生成前诊断或修复提示词	`flux-3-prompt-doctor`
为新镜头编写文本转视频提示词	`flux-3-cinematic-inserts`
基于提供的图像或视频进行创作	`flux-3-keyframes-continuation`
指导对话、旁白、氛围、音效或音乐	`flux-3-audio-dialogue`
提交、轮询、优化、下载或技术审核	`flux-3-generate`

默认流程（跳过已完成阶段）：

text

提示词专家（直到状态为READY；负责镜头结构决策）
  → 电影化镜头生成 或 关键帧续创（二选一，根据是否需要用提供的媒体约束结果来选择）
  → 音频与对话处理（仅当需要音效时执行）
  → 生成执行（仅当请求执行操作时执行）

在提交任何未经验证的内容前先生成草稿：完整渲染需要数分钟时间，而生成的草稿可以以全质量回放。

技能之间的交接需要包含：结果、路由信息（

mode

及其媒体字段：

t2v

、

i2v

keyframes

、

v2v

start_video

或

draft_enhance

draft_cache

）、所需执行的单一操作、源角色、不变参数、交付要求（时长、宽高比、分辨率、镜头结构）、镜头、音频、文本策略以及风险点。当接收技能无需自行做出重要决策即可执行操作时，交接才算完成。

确定性需求（如精确排版、帧级同步、最终混音）需要明确的后期制作计划，而非直接尝试生成。

参考链接：docs.bfl.ai