Models / Luma / Video to video

Luma video to video prompting guide

How to prompt Luma in video-to-video mode. Every point links its source; community tips are marked. Your agent gets the same guide through the Atlas MCP (get_guide).

Prompt structure

edit instruction→what to keep→strength band→keyframe plan

• Ray 3.2 prompt 1–6,000 characters: be specific about subject, motion, camera movement, lighting and pacing. source ↗

What the text should describe

• type 'video_edit': source (≤18 s) + prompt describing the transformation ('Transform the scene into moonlit 35mm film footage'). Start with video.edit.auto_controls true. source ↗
• Pick a strength band: adhere_1–3 keeps motion/composition, flex_1–3 balanced (flex_2 common default), reimagine_1–3 lets the prompt drive. Per-signal controls: pose, depth, normals, trajectory, face. source ↗
• Keyframes are 'the single biggest lever on output quality': pin guide images at source-frame indexes (video.edit.keyframes + keyframe_indexes, up to 64) for anything beyond a quick restyle. source ↗
• Workflow: edit a few stills from the clip in an image model, then feed them back as Ray3.2 Modify keyframes with a short prompt for surgical edits inside a finished video. community source ↗

What not to describe

• Rejected on video_edit: image_ref, loop, end_frame; aspect_ratio is ignored (derived from source). source ↗

Duration and limits

• 5 s or 10 s; 360p draft to 1080p; HDR needs 720p/1080p and 5 s; loop not allowed with 10 s, HDR or end_frame; 9:16 to 21:9. source ↗

Multi-shot

• Multi-keyframe i2v pins 1–64 images at exact frame indexes (24 fps); cannot combine with start/end frames or loop. source ↗

Negative prompt

Supported via .

Real prompts, credited

Transform the scene into moonlit 35mm film footage

Next

Sources

• docs.agents.lumalabs.ai/guides/videos/editing/
• docs.agents.lumalabs.ai/guides/videos/generation/
• x.com/alexa_nft/status/2065433360315437180