Which AI video models take an end frame or reference clips on Sume

Seedance 2.5, Seedance 2.0, Wan 3.0, MiniMax H3 and H3 Max take text-to-video, image-to-video, an end frame and reference images, clips and audio, while Kling Video v3 Pro takes no references and Grok Imagine Video 1.5 is image-to-video only. Among these, Gemini Omni Flash 1.1 is the one that edits an existing video, and it takes no reference audio. Special-purpose models such as h3-max-recast and higgsfield-genjutsu also take a source video.

Capability flags from the catalog

Limits on references

Each limit is quoted from the model's constraints in the catalog, and each applies per job.

• H3: up to 9 reference images, 3 clips of 2 to 15 seconds and 3 audio files, 12 files in all.
• Omni: up to 10 reference images and 3 reference clips of 3 seconds each, plus an edit mode that takes a prompt and a video_url.
• Wan 3.0: up to 10 reference images, 5 clips (15 seconds total) and 5 audio files (15 seconds total).

How to use the matrix

Start from the feature, not the brand. If you need to hold a character across shots with reference images, rule out Kling and Grok first. If you need to change an existing clip, Omni is the general choice among these, and the special-purpose edit models are the other route. If you need the last frame to land on a given image, use any model with the end-frame flag. Then pick among the survivors on price and duration limits.

A caution on limits

The flags say a feature exists, not that every combination is allowed. H3, for example, caps the total at 12 reference files and does not allow audio as the only reference, and Omni treats one reference image with no first or end frame as reference-to-video. Read the constraints list for the model before you build a form around it. Duration and resolution limits differ by model as well, so a feature that exists on two models may fit your clip on only one of them.

For migration work, the safest approach is to list the features your current clips use, tick them against this table, and test one clip per surviving model before you commit.