Models / Kling / Video to video

Kling video to video prompting guide

How to prompt Kling 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

video binding→edit instruction→ref bindings→what to keep→audio keep

• Write prompts as directions to a scene, not object lists. Multi-shot (up to 6 shots): label each shot with framing, subject and motion. source ↗

What the text should describe

• Edit an uploaded clip: prompt references the source as @Video1 and states the change; swap in looks with @Image1.. (style/appearance) and @Element1.. (characters/objects). Example: 'Change environment to be fully snow as @Image1. Replace animal with @Element1'. source ↗
• Source video 3–15 s, .mp4/.mov, 720–3840 px, max 200 MB; max 4 elements + reference images combined when a video is given; keep_audio (default true) keeps the source soundtrack. source ↗
• Higgsfield exposes the same model as 'Kling O3 Video edit' (duration derived from the source, 3–15.5 s; top-level prompt ≤2,500 chars) and 'Video reference' (guide a new shot from footage). source ↗
• Creator pattern: 'put yourself into a clip you upload, keep every move, relight it to golden hour' — source clip + your photo as the element; name what to keep (motion) and the one change (lighting). community source ↗

What not to describe

• Don't re-describe the whole source shot; describe only the change and what must stay (motion, framing). Duration is derived from the source on Higgsfield — don't send a separate duration. source ↗

Camera vocabulary

• Shot language: profile shot, macro close-up, tracking shot, POV, shot-reverse-shot. Moves: slow push-in, dolly left, crane up, handheld, locked-off, orbit; lens feel 35mm/85mm. community source ↗
• Describe camera behaviour over time: tracking, following, freezing when the subject pauses, panning, moving in sync. source ↗

Responds well to

• Define core subjects at the start and keep their descriptions identical across shots. source ↗

Duration and limits

• Kling V3/O3 on fal: 3–15 s, up to 1080p; aspect 16:9, 9:16, 1:1. source ↗
• Longer durations risk drift (fal skill). community source ↗

Multi-shot

• multi_prompt list or shot_type 'intelligent'/'customize'; either prompt or multi_prompt, not both. source ↗

Avoid and failure modes

• Prefer direct declarative prompts; skip prestige adjectives; when negative_prompt is exposed list specific things, not vague ones. community source ↗

Negative prompt

Supported via negative_prompt (default: “blur, distort, and low quality”). source ↗

Real prompts, credited

Change environment to be fully snow as @Image1. Replace animal with @Element1

Next

Sources

• blog.fal.ai/kling-3-0-prompting-guide/
• docs.higgsfield.ai/docs/models/kling-o3
• fal.ai/models/fal-ai/kling-video/o3/pro/reference-to-video/api
• fal.ai/models/fal-ai/kling-video/o3/pro/video-to-video/edit/api
• fal.ai/models/fal-ai/kling-video/v3/pro/image-to-video/api
• github.com/fal-ai-community/skills/blob/HEAD/skills/fal-prompting/references/kling.md
• x.com/Bolly_ai/status/2106882429365502029