LTX keyframe interpolation is open-source only; Sume takes last_frame

LTX's keyframe interpolation is open-source only: the post says KeyframeInterpolationPipeline has no API endpoint. If you want a clip that starts on one known image and ends on another through a hosted API, Sume's frame_images takes a first_frame and a last_frame entry on models whose catalog row reports support.

LTX's statement is from its post How to make longer AI videos, read 2026-10-01. Sume's side is from Video generation and the Video Router docs.

What does LTX say about keyframe interpolation?

The post presents it as the method for getting from a known first image to a known last image, interpolating each gap as its own segment, and each segment must still satisfy the 8k+1 frame rule. It states that this one is open-source only. The hosted LTX API offers extend and retake endpoints for continuation, not this.

How do I send a first and last frame on Sume?

Use frame_images: each entry has a frame_type of first_frame or last_frame. Entries trigger image-to-video; input_references is the separate reference-to-video route, and if both are sent frame_images takes precedence. Check the catalog field supported_frame_images first, which reports which frame_type values a model accepts.

Which catalog models list an end frame?

From the Video Router table read 2026-10-01:

Is that the same as LTX interpolation?

Not necessarily. The docs call it image-to-video with a start and end image, and do not claim the exact segment-by-segment interpolation LTX describes. Test one pair of frames before building a chain, and use the model's own duration limits. More on one model: MiniMax H3 first and last frame. For which field wins when you mix inputs: frame_images and input_references together.