Kling 1.6 Image-to-Video Generation: Start-and-End-Frame, Motion Controls, and Platform Access

Kling 1.6 gives you something no other image-to-video model at its price band offers: start and end frame control. You supply a first image and a last image. The model generates the motion between them. That target image steers the motion, rather than leaving the final frame to drift wherever the model's statistical guess takes it. For a solo filmmaker stitching a shot transition or a social media creator matching a specific last frame to the next clip, that control is the difference between a usable output and a re-roll.

Upload two images. The first is your start frame, which the model treats as the only real frame. The second is your end frame, a target for the last frame of the generated clip. The Kling AI documentation (2026) specifies that the end frame is not stitched into the output. It is a conditioning signal. The model tries to make its final moment resemble that end frame image.

When It Works, and When It Doesn't

When it works, you get a smooth transition from source to target, as if a camera tracked the scene across time. When it fails, a seam artefact appears at the point where the generated motion would have to snap to the end frame. The seam shows as a jump in texture, a sudden colour shift, or a break in an object's shape. The seam is not a bug. It is the model telling you that the gap between your two images was too wide for its temporal attention to bridge. Keep your end frame visually close to the start frame. Similar lighting, same subject pose, small change in camera angle. The seam artefact disappears.

Motion strength is a 0 to 10 slider inside the Motion Brush tool, documented in the Kling API documentation (2026). A value of 0 pins the subject in place. A value of 10 produces the largest motion the model can generate from your source image. The trade-off is direct: higher motion strength increases the risk of subject drift. Your person or object gradually morphs into something else over the clip. The model's temporal attention loses the conditioning signal from the source image as the motion magnitude rises.

Setting the Slider for Your Shot

For a product shot that must stay recognisable, stay at 3 or below. For surreal transitions where the subject changing is the point, go to 8 or above. The motion strength parameter controls how much pixel displacement the model predicts. It does not control the direction of that displacement. Direction is set by the Camera Movement controls, which give you pan, tilt, zoom, roll, tracking, and dolly independently of the strength slider.

Access Kling 1.6 through the web application at klingai.com, through the mobile app for iOS and Android, or through a REST API. Local generation is not available. The model is proprietary and cloud only, per the Kling terms of service (2026). The API documentation (2026) publishes inference time per second of video, but that figure excludes queue wait time. During peak usage, server queue wait can multiply total turnaround by 2 to 10 times. A five-second clip that takes half a minute of inference may take several minutes from submission to download when the shared infrastructure is under load.

Timing Your Render

Queue priority depends on your subscription tier, but the terms of service do not publish the specific prioritisation algorithm. If you need predictable turnaround, run tests outside your market's prime hours. The window from 02:00 to 06:00 UTC is the quietest for a China-hosted platform.

Kling 1.6 outputs a native 384p resolution, confirmed by the Kling model card (2025). The model card also states that the effective resolution after upscaling to 1080p shows visible softening, roughly one resolution tier below the claim. The claimed maximum clip length is 10 seconds. Motion coherence degrades noticeably after 4 to 6 seconds. The usable clip length is closer to 5 seconds for a subject that must stay recognisable. At 5 seconds, the model performs well. Most subjects are still recognisable. At 10 seconds, the identity shift failure mode appears: a face may change ethnicity, age, or gender as the model interpolates toward the distribution of faces in its training data. The clip duration range in the API documentation (2026) lists 5 seconds and 10 seconds as the only options for image-to-video in standard mode and in Kling v2 Pro mode. Audio generation is not supported in image-to-video mode, stated in the Kling API documentation (2026).

Your source image must be PNG, JPG, or WEBP format, with a maximum file size of 10 MB and a maximum long edge of 4096 pixels, per the Kling API documentation (2026). Supported aspect ratios are 16:9, 9:16, 1:1, 4:3, and 3:4. Forcing an unsupported ratio produces aspect ratio distortion: subjects stretch or compress unnaturally because the preprocessing pipeline applies naive resizing.

The Structural Limits of a Single Reference Frame

The model's lack of a persistent 3D world model means background sliding is common. The background moves independently of the foreground subject, as if on a separate layer. Flicker, rapid frame-to-frame changes in brightness or colour, appears when the latent space decoding is inconsistent across frames, especially in clips with fast motion. Limbs warp, multiply, or detach during motion because the model has no skeletal model of the person. These failure modes are not bugs. They are structural consequences of predicting pixel displacement from a single reference frame.

Kling is subject to China's generative AI regulations, which impose content controls and watermarking requirements that differ from EU and US rules. The Kling content policy (2026) bans adult or sexual content, violence or gore, political content, celebrity deepfakes, and copyrighted character generation. The same restrictions apply to input images. A benign source image can trigger a content moderation false positive, flagged by classifiers that are overly broad on skin tones or certain object shapes.

Watermarks and Platform Rules

Free tier outputs carry a visible Kling AI watermark, stated in the Kling terms of service (2026). The EU AI Act, Article 50, requires providers to label outputs as AI-generated. YouTube and TikTok require disclosure of AI-generated content. If you upload a Kling-generated clip to a platform that uses its own detection system, a mismatch between Kling's metadata watermark and the platform's detection method can lead to a watermark rejection. The clip is flagged even though the model's watermark is present.

Consent for Real People

If you are animating a real person, you need their explicit consent. The model will generate motion from a single photo, but doing so without permission breaches the content policy and, in many jurisdictions, privacy law. Get written consent before you upload anyone's image.

The Kling terms of service (2026) permit commercial use for paid subscribers and assign output ownership to the user. The free tier is for non-commercial use only. You grant Kuaishou a worldwide, royalty-free, sublicensable licence to host and process your input and output for service operation, but the provider claims no ownership of your output. The US Copyright Office policy statement (2023) holds that works produced solely by AI without sufficient human authorship are not copyrightable, though human selection and arrangement may be protected. If you are a legal or compliance person clearing AI-generated video for broadcast, document the human creative input that went into the prompt and image selection.

The solo filmmaker or video artist who needs b-roll, establishing shots, or surreal transitions without a crew will get the most out of Kling 1.6's start and end frame control. The social media content creator producing high-volume short-form video will value the predictable output at 5 seconds and the region mask control for animating only a selected area of the product shot or meme still. The educator or researcher studying generative tools will find the documented failure modes, especially the seam artefact and subject drift, to be teachable examples of the limits of latent diffusion architectures.

Who Should Look Elsewhere

The open-source tinkerer running models locally should skip Kling entirely: no local generation, no open weights, cloud only. The person who wants video from a text prompt alone should use a text-to-video comparison resource; this model is for image input. The person shopping for an all-in-one video editor with generative features built in should look at video editing software, not a generation model.

The One Thing That Goes Wrong

The single thing that most often goes wrong with Kling 1.6 is expecting the 10-second clip to hold source fidelity, when the usable range is closer to 5 seconds. Set your expectations to 5 seconds. Use the end frame feature to lock down the last frame. Keep motion strength at or below 3 for subjects that must stay recognisable.