Kling·de KuaishouImagen a videoNuevo

Kling Video 4.0 Image-to-Video

Kling 4.0 keyframe video: one continuous 3-30 second shot from a first frame, through up to 8 middle keyframes in order, to an optional last frame. The prompt is optional; keyframes are referred to as @image_1, @image_2… Aspect ratio follows the keyframes.

Abrir en el espacio de trabajo →Preguntar a LuviBotkwaivgi/kling-v4.0/image-to-video

Parámetros

Parámetro	Tipo	Por defecto	Rango u opciones
Duration (s) `duration` Video length in whole seconds, 3 to 30. Billed per second.	select	5	3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30
Resolution `resolution` Output resolution. Billed per second at the selected tier.	select	1080p	720p (720p), 1080p (1080p)

Entradas

Archivos que este modelo recibe además del prompt.

Entrada	Acepta	Máx. de archivos
First Frame `image`	image	1	Obligatoria
Middle Keyframes `middle_images`	image	8	Opcional
Last Frame `last_image`	image	1	Opcional

Reglas de LUVI Prompt Engine para Kling

TARGET MODEL: Kling 4.0 (Kling AI / Kuaishou — native multimodal video with stereo audio generated in the same pass, 3–30 s). It acts as a director: a short prompt is expanded with its own storyline, blocking, camera and even a line of dialogue; a clear prompt is followed faithfully. Multi-shot is ON by default with no switch. No negative_prompt, no seed, no cfg, no weights — the prose is the whole instrument.

Reglas que deben cumplirse

DESCRIBE WHAT THE CAMERA SEES. Kling's own principle: "Use descriptive language and minimize abstract terms and negative instructions." Abstract labels ("premium look", "vintage feel", "cinematic", "cinematic feel", "dynamic camera", "high quality", "masterpiece", "8k") do nothing useful: when the user's own words already give the concrete detail behind a label, keep the detail and delete the label; when they do not, delete the label — never invent the detail yourself. Never add such labels.
A WRITTEN TIME RANGE IS A SHOT. "0–3 s: … 3–7 s: … 7 s–end: …" produces a hard cut exactly at each boundary, even when the range describes a continuous move. Use ranges ONLY when the user wants separate shots or beats that must land at given times. To keep one continuous take across several beats, write them as one flowing sentence in order and add "One continuous shot, no cuts." You do NOT know the clip length (it is a parameter you cannot see): keep the user's own time marks exactly, and never invent second marks — when the user lists shots without times, keep them as ordered shots ("Shot 1: … Shot 2: …"), never convert them to seconds.
PROSE THAT MOVES FROM PLACE TO PLACE INVITES CUTS. Several actions in several locations without ranges come back as a montage. If the user wants one take, keep the action continuous and the camera following it.
KEEP A RECURRING PERSON CONSISTENT: use the user's description of a person once, word for word, and refer to them by the same label in every later shot (never a synonym). Kling changes an undescribed person's face and age across cuts — but do not invent a description; if the user gave none, keep the user's label. When the user names a count, add a count guard in Kling's form: "EXACTLY 2 CHARACTERS — NO DUPLICATES."
AUDIO IS ALWAYS GENERATED. Write the soundtrack the scene needs: the user's dialogue/voiceover + ambient sounds and effects that the described place and actions make. NEVER add music, a song or a score the user did not ask for. When subtitles or background music are not wanted, end with "No subtitles. No background music." — these are the two exclusions Kling says are always worth writing.
DIALOGUE: visible speaker + action first, then "speaks in <language>, <manner>:" and the line in double quotes, in that language's own script: She speaks in natural, clear, slightly slow Spanish: "¿Dónde está la estación?" Describe age, tone, speed and emotion for a specific voice. One speaker per turn, no overlaps, short lines for the seconds available; time several turns with ranges. Keep the user's line word for word.
ON-SCREEN TEXT goes in double quotes, short, ideally one headline line; keep the user's copy character for character, diacritics included. Subtitles the user DOES want: state position, number of lines, style and timing.
EXCLUSIONS: describe the scene positively ("He is shown in profile", not "don't show his face") — one positive choice, never "from behind or in profile". Keep only exclusions Kling itself prescribes: subtitles/background music when unwanted, what a reference must not carry over, "no photorealism" for animation styles, or one specific thing the brief would otherwise invite. Never stack generic negative lists.
DURATION, RESOLUTION AND ASPECT ARE PARAMETERS. Never write "16:9", "4K", "1080p" or "15 seconds long" into the prompt as settings. Ranges inside the prompt are fine.
No weights, no "(term:1.3)", no "--ar", no "seed:", no "Negative prompt:" label — delete them, keep the words they wrapped.

Lo que mejor funciona

Kling's template, blocks dropped when they do not apply: [Overall Settings] Scene and texture: …; visual style: …; camera tone: … · [Sequence of Events] 0–X s: … · [Audio] … — or plain prose for a single take.
Two registers. A ONE-LINE IDEA ("An old man ranting about AI.") is a deliberate request for Kling's own direction: return it essentially as written — fix grammar, delete dead labels, at most add "No subtitles. No background music." — and never write a plot, a dialogue line, wardrobe or a room around it. A BRIEF the user already structured (headers like SCENE CONTEXT, ACTION TIMING, AUDIO, or time ranges) is already in Kling's form: keep its header labels verbatim ("SCENE CONTEXT:" stays "SCENE CONTEXT:"), its ranges exactly as written ("9 s–end" stays "9 s–end"), its wording and order; change only what breaks a MUST rule. Longer is not better.
Camera: when the user names a move, write it with what it follows ("the camera tracks beside her at her pace"); a vague one ("dynamic camera") becomes the one move the user's action implies (following the moving subject) or is deleted. One dominant move per shot. Never invent amounts or percentages.
Beat density: about three clean beats in 10 s, two beats plus one line in 8 s; a 5 s clip with three actions becomes a cut-up montage. For 20–30 s keep motion slow to moderate and give the final seconds a written action.
When the user names a drink or food, name its vessel too if the user's words imply it ("sips her Turkish coffee" → "sips from a small Turkish-coffee cup") — Kling otherwise picks a generic cup.

De las reglas de LUVI Prompt Engine para esta familia: las mismas que aplican el reescritor del espacio de trabajo y el conector MCP de LUVI.