Kling 3.0

Overview

Kling 3.0 is a video generation model from Kuaishou, and what HiAPI offers is its Motion Control route. Developers pass in a character reference image and a motion reference video, and the model makes the character in the image reproduce the motion and expressions from the video — the character's appearance comes from the image, the movement from the video, with the option to keep the reference video's original sound. It suits developers who need precise control over character performance rather than generating motion from text descriptions alone, for use cases such as short-drama performance replication, digital human voiceover, and marketing clips, billed by output seconds.

Highlights

• Motion driven directly by the reference video — motion and expressions come from the reference video while the character's appearance comes from the reference image, giving finer control than text prompts alone.
• Two orientation modes cover different needs — video-following mode suits complex motion, up to 30 seconds; image-following mode suits camera movement, up to 10 seconds.
• Optional retention of the reference video's original sound — keep_original_sound is enabled by default, so the output can reuse the original video's audio, saving post-production syncing.
• Per-second billing, tiered by resolution — costs are settled by output seconds, tiered across 720p and 1080p, making costs predictable.

Capabilities

Character motion replication: Drive the character in the character reference image with a motion reference video, and the model reproduces the motion and expressions from the video; developers get character performance footage directly without describing it frame by frame.

Prompt assists with scene and camera: Use prompts to add scene, style, and camera descriptions, up to 2500 characters; motion and expressions are still determined by the reference video, and the prompt only guides atmosphere and camera.

Original sound toggle: keep_original_sound is enabled by default, so the output reuses the reference video's audio; disable it when you need to re-dub, for easier downstream audio post-production.

Orientation mode switching: character_orientation can be set to video or image; the former makes the output follow the reference video and suits complex motion with a maximum of 30 seconds, while the latter follows the reference image and suits camera movement with a maximum of 10 seconds.

Resolution and billing basis: Output can be 720p or 1080p, billed by output seconds with higher resolutions costing more; developers can trade off final quality against budget.

Use cases

• Dance and performance motion replication — Input a character image and a dance reference video to produce a clip of that character reproducing the same motion, for short-video content production and character re-creation.
• Short-drama character performance and camera movement replication — Use a live-action performance video as motion reference so that in-drama character assets reproduce the performance and camera movement, producing performance clips ready for the editing pipeline.
• Digital human voiceover and lip sync — Input a character image and a voiceover reference video to produce a character voiceover clip with the original sound retained, for digital human broadcast-style content.
• Marketing and product showcase clips — Design character motion with a reference video and produce product showcase assets with character animation, for ad delivery and social media distribution.
• Re-interpretation of existing footage with original sound — Keep the reference video's original sound and replace the character's appearance to produce reusable derivative assets, cutting the work of dubbing and syncing.

Limitations and selection advice

This route requires both a character reference image and a motion reference video; it cannot generate video from text alone. Asset constraints are fairly strict: images must be jpeg/png/jpg, no larger than 10MB, both sides greater than 340px, aspect ratio between 2:5 and 5:2, with clear head, shoulders, and torso; videos must be mp4/mov, no larger than 100MB, 3 to 30 seconds, and a single continuous shot. Output duration is limited by the orientation mode — image-following maxes out at 10 seconds and video-following at 30 seconds. This route supports at most 1080p, so if you need higher resolution or pure text-to-video generation, choose another model or route.