Grok Imagine·von xAIVideo edit

Grok Imagine Video Edit

xAI Grok Imagine Video — edits an MP4 with natural-language instructions. Output retains source duration (max 8.7s); billed per second of the input video.

Im Arbeitsbereich öffnenxai/grok-imagine-video/edit-video

Parameter

Dieses Modell nimmt nur den Prompt und seine Eingaben; es gibt keine einstellbaren Parameter.

Eingaben

Dateien, die dieses Modell zusätzlich zum Prompt entgegennimmt.

EingabeAkzeptiertMax. Dateien
Source video
video_url
video1Erforderlich

Prompt-Leitfaden für Grok Imagine

TARGET MODEL: Grok Imagine Video (xAI Aurora — autoregressive; picture AND audio co-generate in the same pass). The prose is the whole instrument: no seed, no negative_prompt, no audio toggle.

Regeln, die gelten müssen

  • The formula, in timeline order: Subject + Action/Motion + Camera + Environment/Lighting + Style + Audio. Front-load the critical action — earlier words weigh more. 30-60 words optimal.
  • AUDIO IS ALWAYS ON with no silence flag: if the prompt doesn't specify sound, the clip gets silent or random audio — so ALWAYS write the audio in. Music, SFX timed to motion, ambient/room tone, lip-synced dialogue.
  • Dialogue: quote the line with a delivery-cue prefix — a quiet whisper: "We made it." · urgent shout: "Stop him!" — and keep lines SHORT (audio is the weakest layer; long lines turn to gibberish). Optional sound-design clause: AUDIO: soft room tone, faint kettle hiss.
  • ONE primary action and ONE named camera move per clip — conflicting moves (zoom+pan) break temporal coherence.
  • State motion magnitude explicitly — the model cannot infer it ("passing" → "passing quickly"). Strong verbs with intensity adverbs ("sprints", "surges", "pitches forward with tremendous force").
  • Negation is ignored — phrase every exclusion as the desired positive state ("sharp focus throughout", never "no blur").
  • One aesthetic per clip — never mix (no anime+photoreal). Lighting by source + direction.

Was am besten funktioniert

  • Camera vocabulary that lands: locked/static (strong default) · slow push-in · slow dolly-in · tracking shot alongside · handheld follow from behind · orbit/360 · slow crane pullback · rack focus · aerial push-in · pull-back · time-lapse. "Cinematic" or "dynamic camera" name nothing executable — name the move or lock the frame.
  • Avoid known failure modes: big body dynamics (jerky), extreme close-ups and macro hand work (distort), long photoreal holds (waxy morph). Favor short beats and implied/reaction motion.
  • Craft pattern for scenes with speech: scene → dialogue in quotes with its delivery cue → optional AUDIO: line.

Aus LUVIs eigenem Handbuch für diese Familie; dieselben Regeln wendet der Rewriter im Arbeitsbereich und die MCP-Prompt-Engine an.

Weitere Modelle dieser Familie