flux-3-keyframes-continuation
Compare original and translation side by side
🇺🇸
Original
English🇨🇳
Translation
ChineseFLUX 3 Keyframes and Continuation
FLUX 3 关键帧与续拍
Every supplied source gets one declared job, or is dropped. Two modes condition on
media, and each carries exactly one media field:
| Required relationship | Request |
|---|---|
| These exact pixels open the clip (the stable path) | |
| Exact opening and closing frames, filled in between | |
| An ordered storyboard the model connects | |
| Frames pinned at exact moments | |
| Keep going from where this clip ends | |
Keyframes appear on screen, pixel-exact. There is no field that carries a subject
without showing the source image; a source used only as inspiration is described in the
prompt instead. Media is a public URL or inline base64; keyframe images are at least
256x256.
每个提供的源素材对应一个已声明的任务，否则会被丢弃。根据媒体类型分为两种模式，每种模式仅包含一个媒体字段：
| 必需关联关系 | 请求参数 |
|---|---|
| 这些精确像素作为视频片段的开头（稳定路径） | |
| 指定精确的开头和结尾帧，模型自动填充中间内容 | |
| 模型按顺序连接故事板中的帧 | |
| 在指定时间点固定帧画面 | |
| 从现有视频片段的结尾继续生成 | |
关键帧会像素级精确地呈现在画面中。不存在仅作为灵感参考而不显示源图片的字段；若仅将源素材作为灵感，需在提示词中描述。媒体素材可为公开URL或内嵌base64格式；关键帧图片分辨率至少为256x256。
Prompting each shape
针对不同模式的提示技巧
- One keyframe: the image opens the clip. Prompt what moves; redescribing visible pixels invites re-imagination.
- Two keyframes: opening and closing are pinned, so describe the stages of the change in order and hold the camera still. Keep the two ends related enough for a plausible path; large jumps invite drift.
- Storyboard: frames become ordered waypoints. Keep identity, palette, and viewpoint coherent unless a change is intentional.
- Timestamped pins: a mid-clip pin is a deadline; say what the remaining seconds are for, or they trail off. Pin only what matters; extra anchors stiffen motion.
- Continuation: write from the ending. Say what happens next and which momentum, framing, subjects, and sound carry across the boundary; do not recap completed action.
text
The camera continues a slow push-in as the closed flower opens naturally into the final
bloom. The stem remains fixed in place, the same soft morning light holds across the
shot, and a light breeze moves the surrounding leaves without changing the composition.text
Continue from the rider's existing forward momentum. The camera keeps the same low
trailing angle as the horse crests the hill and enters dense fog; hoofbeats and wind
carry across the boundary without a reset.- 单关键帧：图片作为视频开头。提示词需描述画面中的动态变化；若重新描述可见像素，可能会触发模型的重新创作。
- 双关键帧：开头和结尾帧已固定，需按顺序描述变化阶段并保持镜头稳定。确保首尾内容关联性足够强以生成合理过渡；跨度太大可能导致画面偏移。
- 故事板模式：帧画面作为有序的路径点。除非有意改变，否则需保持主体、色调和视角的一致性。
- 时间戳固定帧：中间帧的固定是硬性要求；需明确剩余秒数的用途，否则画面可能会拖沓。仅固定关键内容；过多锚点会使动作变得僵硬。
- 续拍模式：基于现有视频结尾进行创作。说明接下来的情节，以及需要延续的运动趋势、镜头构图、主体和音效；无需回顾已完成的内容。
text
The camera continues a slow push-in as the closed flower opens naturally into the final
bloom. The stem remains fixed in place, the same soft morning light holds across the
shot, and a light breeze moves the surrounding leaves without changing the composition.text
Continue from the rider's existing forward momentum. The camera keeps the same low
trailing angle as the horse crests the hill and enters dense fog; hoofbeats and wind
carry across the boundary without a reset.Chains
多段生成衔接
For clips longer than one generation: preserve a clean ending and record the subject's
direction, speed, and camera relationship; continue from the actual ending artifact;
restate only invariants and the next action; review each seam before extending; change
one continuity variable at a time when repairing drift. A video input at higher
resolution may cap duration below the general maximum, which changes segment planning.
Return the conditioning plan (which mode and shape, and why), the invariants that must
survive, the final prompt, and any pin map or transition risks.
Field limits and current constraints: API reference
对于超过单轮生成时长的视频：需保留清晰的结尾画面，记录主体的运动方向、速度和镜头关系；基于实际生成的结尾画面继续创作；仅重申不变元素和下一个动作；在延长前检查每一处衔接；修复画面偏移时每次仅调整一个连续性变量。高分辨率的视频输入可能会将时长限制在通用最大值以下，这会改变分段规划。
返回条件规划（包括模式、类型及原因）、必须保留的不变元素、最终提示词，以及任何固定帧映射或过渡风险。
字段限制与当前约束：API参考文档