Kling V4 Reference-to-Video Model

Overview

Kling V4 Reference by KlingAI generates short videos from your written direction plus visual references. You anchor characters, products, and scenes with reference images or reference videos, then describe motion, camera moves, and story beats in text. You can also lock specific moments with start, middle, and end keyframes, or edit an existing clip by telling the model what to change and what must stay the same. This workflow helps you keep identity and style consistent across motion, which solves the common “character drift” problem in video generation.

What you can build

• Product hero videos that keep the product shape, label text, and materials consistent across motion
• Character-driven shorts that reuse the same person, outfit, and style across multiple scenes
• Style transfers for existing footage, like turning a clip into ink-and-wash animation while keeping framing and action
• Keyframe-led motion shots where you predefine the opening, key beats, and ending frame
• Marketing variants where you hold the subject constant but change background, lighting, or time of day

Inputs

• A detailed video description in plain text. Include actions, facial expression, camera movement, and timing cues. If you upload references, you can reference them in the text using the tag format shown in the workspace (for example, @image_1 / @video_1 or @Image1 / @Video1).
• Optional reference images to define characters, objects, locations, or a target look. Plan for a total cap of 10 images across all image inputs, including any keyframes you add.
• A target clip length you choose in seconds. This model is designed for short clips up to 30 seconds.
• An output resolution setting (720p or 1080p).
• Optional keyframe images: - A first frame image to lock the opening look.
  • Middle keyframes (up to 8 images) in timeline order to lock specific moments.
  • A last frame image to lock the ending look.
• Optional reference videos to guide motion, pacing, or camera behavior. Use MP4 or MOV clips that are 3–30 seconds each, with a combined upload total capped at 30 seconds. For best results, use clips that are at least 700 px on the short side and between about 24–60 fps.
• Optional base video for editing. Provide a single MP4 or MOV clip (3–30 seconds) and write your edit instruction so the model knows what must remain unchanged.
• An aspect ratio choice. “Adaptive/Auto” can follow your references when you provide them.
• A watermark toggle, if you need a marked output.

Outputs

The model returns a generated video file in MP4 format. The clip reflects your prompt instructions and attempts to preserve the identities, objects, and style you provided in references and keyframes. Outputs are typically provided as a downloadable file with basic file metadata such as name, MIME type, and size.

Recommended settings

• If you care most about identity consistency, use 2–4 strong reference images per subject and explicitly state what must stay fixed (face, hair, outfit, logos).
• If you need specific story beats, add keyframes and describe what should happen between them. Keep keyframes in strict timeline order.
• For motion guidance, use 1–2 short reference videos that clearly show the camera move you want. Avoid mixing multiple conflicting camera styles in the same request.
• Use 1080p when fine text, small logos, or thin patterns must stay legible during motion.

Limitations

• You can’t exceed the reference caps. Expect a hard limit on total images (including keyframes) and a hard limit on reference videos (count and total seconds).
• Reference videos and base videos must meet minimum quality constraints. Low-resolution, heavily compressed, or very low-frame-rate clips can cause artifacts and unstable motion.
• Very busy prompts can reduce control. If you don’t assign clear roles to each reference, the model may guess and produce mixed results.
• Keyframes improve control but can also cause “snapping” transitions if frames differ too much in lighting, pose, or camera angle.
• Non-structured or low-quality inputs increase risk. Blurry images, screenshots with UI overlays, scanned images, or videos with heavy motion blur often lead to identity drift or copied-looking frames.

Safety & compliance

KlingAI’s terms prohibit using the service to upload or generate unlawful content, including pornographic or pedophilic material, and content that harms minors.

If your deployment supports provenance features, generated media may include provenance signals such as embedded credentials and/or invisible watermarks. Keep visible watermarking enabled when you need clear disclosure for social platforms or client review.