How to Plan an Image-to-Video Shot Using Start and End Frame Control
Which image-to-video models accept both a start and end frame, and how to plan the two frames so the interpolated clip looks like one continuous shot.
Which Models Accept Both a Start and an End Frame?
Five image-to-video models support a start frame and an end frame for a single generated clip as of 2025-12-16: Kling 1.6 (Professional mode), Luma Dream Machine (paid tiers), Seedance 1.0, Wan 2.1 (I2V-14B), and LTX-Video 0.9.5 (I2V). Kling 1.6 is the primary example from the dossier. Luma Dream Machine offers keyframe looping via its End Frame feature. Models that provide an image-to-video start and end frame control workflow include Kling 1.6 and Seedance 1.0.
Models that do not support end-frame control: Runway Gen-4, Sora, Veo 3, Hailuo AI, Pika 2.2, and Stable Video Diffusion. This matters because it eliminates the seam artefact, the visible glitch that appears when you concatenate two separately generated clips. Without end-frame support, any shot that requires a specific final pose, a precise background, or a join that hides the cut forces you to splice outputs that were never designed to connect.
Why End-Frame Control Kills the Seam Artefact
Clip concatenation is the standard workaround when a model cannot hold a scene together beyond its maximum native clip length. You generate clip A, generate clip B, and use FFmpeg to stitch them. The problem is that clip B has no memory of clip A. The figure's position, the lighting, and the background have all drifted. The join produces a seam artefact, a frame where the scene visibly jumps, the person shifts by several pixels, or the colour grade changes abruptly.
End-frame control prevents this. You supply the first frame (your source image) and the last frame (a target image you prepare). The model interpolates the motion between them. The result is a single clip with no join. Because the model sees both anchor points, it maintains source-image fidelity at clip end far better than a model that only sees the start frame.
How to Plan a Shot with Both Frames: The Matching Workflow
What Must Match Between Your Two Frames
You control what the model interpolates by controlling what matches between your start frame and your end frame. If the interpolation is to look natural, the following must be consistent:
- Figure position. The main person or object should occupy roughly the same area of the frame in both images. A large lateral jump forces the model to guess a motion path that may not exist in its training data.
- Lighting direction and colour temperature. A warm-golden start frame and a cool-blue end frame will produce a colour shift that reads as a technical error, not as a creative transition.
- Colour grade. Match the overall saturation, contrast, and tint. The model is not a colourist. A grade mismatch in the two frames will produce a grade shift mid-clip.
- Background structure. If the background changes significantly (a door that is open in frame A and closed in frame B), the model will morph the background through an intermediate state that rarely looks plausible.
Preparing and Uploading the End Frame
Prepare the end frame in an image editor. Resize it to the model's required input resolution, 1080p for Kling 1.6, 1360×752 for Luma Dream Machine. Use the same crop, the same aspect ratio, and the same person or object. Upload the start frame as the image input and the end frame in the designated End Frame field. The model handles the interpolation.
Kling 1.6 Start-and-End-Frame Workflow
Kling 1.6 exposes the end-frame feature in its Professional mode. You upload a start image (JPEG, PNG, or WebP; under 10 MB). In the End Frame field you upload the target image. Set the clip length to 5 seconds, the only option in Professional mode. The native resolution is 1080p. You can also select a camera movement from six presets: Horizontal, Vertical, Pan, Tilt, Roll, or Zoom. The model does not expose a motion strength parameter; the camera preset determines the type of motion. The end frame gives you control over the final composition.
A common failure mode is figure drift. If the person or object in your end frame is not recognisably the same entity as the start frame, the interpolation will morph into something neither frame intended. Kling 1.6 does not generate audio. Before you generate a clip of a real person, confirm you have that person's consent for the use you intend.
Luma Dream Machine Keyframe Loop Extend Workflow
Setting Up the End Frame
Luma Dream Machine calls its end-frame feature the End Frame option, available on paid tiers. The free tier is limited to 5-second clips; paid tiers allow 10-second clips. The native resolution is 1360×752 at 5 seconds or 752×1360 in portrait mode. You upload a start image and an end image. The model does not expose a motion strength slider. Camera control is limited to six presets: Pan Left, Pan Right, Tilt Up, Tilt Down, Zoom In, and Zoom Out.
Speed, Fidelity, and the Keyframe Loop
Luma Dream Machine is fast. Inference time lands under 30 seconds for a 5-second clip, but the trade-off is that source-image fidelity at clip end degrades noticeably in longer clips. The end frame mitigates this: the model knows where it must end, so it does not drift as far. Luma does not generate audio. The keyframe loop is a specific use case: you set the start and end frame to the same image, and the model generates a loop that cycles the motion without a visible join. Before you generate a clip of a real person, confirm you have that person's consent for the use you intend.
Other Models That Support End Frames, and How They Compare
Three More Models with the Feature
Seedance 1.0 (ByteDance/Jimeng) supports end frames. Clip length is 2 to 4 seconds, native resolution 1080p. No camera control is exposed. Access the model through the Jimeng AI web app in China. Wan 2.1 I2V-14B supports end frames via a dedicated End Frame feature. Clip length is 5 seconds at 720p. The model is open-weights under an Apache 2.0 licence and can run locally on consumer GPUs. LTX-Video 0.9.5 I2V supports end-frame conditioning. Clip length is 2 to 6 seconds depending on the frame count setting (49-121 frames at 24 fps). Default resolution is 768×512. It is also open-weights under the Lightricks OpenRAIL-M licence and runs in ComfyUI.
Models Without End-Frame Support
Models that do not support end frames include Runway Gen-4, Sora, Veo 3, Hailuo AI, Pika 2.2, and Stable Video Diffusion. For these models, any clip longer than the native maximum requires concatenation, which introduces the seam artefact. Keep clips to the shortest native length that serves the shot. Accept the seam as a post-production problem to be masked with a cut or a transition effect.
The Single Thing That Most Often Goes Wrong
Newcomers prepare a beautiful start frame and a clean end frame but forget to match the aspect ratio. A start frame at 16:9 and an end frame at 4:3 will force the model to stretch or crop the person or object. Check that both frames are the same aspect ratio and that the figure occupies the same proportional area. The model cannot fix a mismatch it never learned to handle.
Common Questions
Which image-to-video models support start and end frames as of 2025-12-16?
Kling 1.6 (Professional mode), Luma Dream Machine (paid tiers), Seedance 1.0, Wan 2.1 I2V-14B, and LTX-Video 0.9.5 I2V. Runway Gen-4, Sora, Veo 3, Hailuo AI, Pika 2.2, and Stable Video Diffusion do not.
What is the difference between Kling 1.6 and the original Kling for end-frame support?
Kling 1.6 introduced the End Frame feature in Professional mode. The original Kling release did not offer it. The two are distinct versions, not interchangeable.
Does Luma Dream Machine support keyframe looping?
Yes. You set the same image as both the start and end frame. The model generates a loop that cycles the motion without a visible seam artefact.
What is a seam artefact in image-to-video generation?
A visible jump or glitch that appears when two separately generated clips are concatenated. It occurs because clip B has no memory of clip A's final frame. End-frame control prevents it by giving the model both anchor points in a single generation.
How do I prepare an end frame for Kling 1.6?
Use an image editor. Match the figure's position, lighting direction, colour grade, and background structure of the start frame. Resize to 1080p. Upload as JPEG, PNG, or WebP under 10 MB.
Can I use end-frame control with Runway Gen-4?
No. Runway Gen-4 does not support end-frame control as of 2025-12-16. You can use its Director Mode camera controls and Motion Brush region masks, but the clip must be generated from a single start frame.
What happens if my start and end frames do not match well?
The model will morph the figure through an intermediate state that rarely looks plausible. You will get drift, identity shift, or background sliding. The interpolation assumes the two frames depict the same scene, so any inconsistency forces the model to guess.