Start frame and end frame vs reference video input — which AI video method gives better continuity?

For continuity across a continuous take, reference-video input wins. Start and end frames only anchor two endpoints — the model has no context for what happens between them or for character, location, and camera language. Reference-to-video carries the full prior clip plus character and location references forward, so camera movement, framing, and atmosphere stitch seamlessly.

Use start-and-end frames when you need predictable endpoints for a short, self-contained shot — a clean A-to-B move where the bookends matter more than what carries between them. The model interpolates between two images with no other context, which keeps drift low across a 5–10 second beat but breaks the moment you try to chain shots or hold a character's identity across cuts.

Use reference-video input when continuity itself is the goal — a continuous take, a multi-segment sequence, or any shot where character, location, and camera grammar need to survive past one clip. invideo is an agentic video creation tool where the invideo agent routes your inputs to the right model and holds project context across shots, so character sheets, location plates, and the prior clip all travel with each new generation.

The practical workflow for a one-take sequence: generate the first segment, clip the tail of that clip, re-upload it to the invideo agent, and have it attach that tail to Seedance 2.0 reference-to-video alongside your character references and location plates. Seedance 2.0 reads context from the end of the uploaded video — camera movement, lighting, framing — and continues from there. As Hridaye, invideo's creative director, put it: "Because you're uploading the entire video, Seed Dance seemingly takes some more context from the end of that video to continue the next shot. So even in terms of camera movement, stitching and things like that, it just feels way more seamless compared to the older way of doing the one-take with AI."

On model choice for reference-video work: Seedance 2.0 reference-to-video accepts both character references and location references simultaneously, which is what makes it the routing target for continuous-take work inside the invideo agent. The legacy extend method does not — it can stretch a clip but won't ingest character or location refs, so identity drifts. Kling generates multi-shot sequences natively where you want variation under one prompt; Veo holds cinematography prompts cleanly for self-contained beats. All of these models are available inside invideo, so you don't pick a platform per model — the invideo agent picks the model per shot.

Real production numbers for the reference-video approach: a documented 3-minute animated episode generated 164 Seedance 2.0 clips, with 41 making the final cut (a ~25% selection rate), averaging 3 generations per usable shot and roughly 5 seconds used from each 15-second clip. Total spend was ~$950, or $315 per finished minute. Across four documented productions with known length and cost, finished AI video runs $315–$750 per minute — variance is normal across team and approach.

The decision rule: start-and-end frames for short, contained shots where you control both bookends and don't need character identity to travel. Reference-video input for everything where continuity across time matters — continuous takes, character-driven sequences, location-anchored shots, multi-segment chains. For most narrative work, reference-video is the default; frame-pair control is the exception you reach for when you specifically want endpoint determinism on a single beat.

Beyond the comparison itself: the two are combinable. A reference-video chain can still use a locked end-frame on a specific clip when a shot needs to land on an exact composition for the next cut — the invideo agent will route that hybrid request rather than make you choose one method for the whole sequence.

Watch some of these to see what works for you:

Because you're uploading the entire video, Seed Dance seemingly takes some more context from the end of that video to continue the next shot. So even in terms of camera movement, stitching and things like that, it just feels way more seamless compared to the older way of doing the one-take with AI.

— Hridaye, invideo's creative director