End-frame AI video: which Sume models take a last frame

Seedance 2.5 and the Seedance 2.0 rows, Wan 3.0, Kling 3, MiniMax H3, MiniMax H3 Max and Gemini Omni Flash 1.1 accept a last frame as well as a first frame on Sume. Grok Imagine Video 1.5 takes a first frame only, and Genjutsu and H3 Max Recast take neither, since they work from a source video.

How to send two frames

On /v1/videos, use frame_images with a frame_type of first_frame or last_frame. On the Video Router the fields are image_url and end_image_url. An end frame needs a first frame; it cannot stand alone.

If you send both frame_images and input_references, the frame images take precedence and the request is treated as image-to-video.

Good uses

The model decides the motion between the two frames, so unusually different frames give odd interpolation.

• Seamless loops: use the same image as first and last frame, then check the join.
• Before and after reveals: the first frame is the problem, the last is the result.
• Chaining segments: take the last frame of clip one and use it as the first frame of clip two.

Mixing with references

Do not assume frames and references mix. The docs say frame_images take precedence when both are sent, and MiniMax keeps first and last frame and reference roles apart. Check the model row, then test on one cheap request.