gabrielmoreira FLUX 3 Keyframes and Continuation
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) | mode: "i2v", keyframes: one image |
| Exact opening and closing frames, filled in between | mode: "i2v", keyframes: two images |
| An ordered storyboard the model connects | mode: "i2v", keyframes: 3-10 images plus an integer duration |
| Frames pinned at exact moments | mode: "i2v", keyframes: [seconds, image] pairs, ascending, integer duration |
| Keep going from where this clip ends | mode: "v2v", start_video (input max 15 s / 50 MB; output max 15 s) |
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.
Prompting each shape
- 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.
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.
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
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