Prompts are terrible at endings. You type up a shot, the opening frame looks fine, and then the action just wanders off to wherever the model felt like going. First Frame, Last Frame control attacks that exact problem. You hand the model two approved stills — the start, the end — and it only has to build the road between them. For me, that single change turned short clips from happy accidents into shots I could actually cut. Huge difference.
I use Seedance 2.5 day to day for story work and Veo when I need that photoreal, native-audio hero shot. Both live downstream of a storyboard-first workflow, so keyframing was never a new idea for me — my boards already are first and last frames. At the time of writing, first-and-last-frame generation is available in Kling, Luma Dream Machine, Veo 3.1 through Flow and the Gemini API, and Runway in its image-to-video workflow. Names and limits shift between versions, so I check the actual panel rather than trusting some two-month-old tutorial.
Why Endpoints Beat Prompts For Control
A text prompt describes a shot. It does not lock one. Write “woman places lantern on wooden table, warm light” three different ways and you get three tables, three lanterns, and three slightly different women. Add a second shot and the drift compounds. By shot four you are basically casting strangers.
Endpoints solve a different class of problem entirely. When you supply a first frame and a last frame, three things get decided before the model spends any of your credits.
- Composition at both ends. Shot size, subject position, and horizon are no longer suggestions.
- Identity at both ends. Face, wardrobe, product shape, and location design are anchored to images you already approved.
- The cut point. Your last frame is someone else’s first frame. The outgoing clip and the incoming clip already share a frame, so the edit stops looking like a slideshow.
That last point — the cut point — is the one everyone undervalues. A keyframed last frame is an edit decision wearing a generation setting as a disguise. If shot one ends on the exact frame shot two begins from, or on a clean, deliberate new angle you built on purpose, the transition in Premiere Pro becomes a plain hard cut. Not a cover-up job with dissolves and B-roll.
What endpoints do not fix is everything in between. The model still invents the journey. Hands can still smear. Extra fingers can still show up. A simple turn can become a full dance if you let it. Keyframes narrow the problem. They do not eliminate it.
Building A Last Frame Worth Landing On
Most failed keyframe work I see has exactly one cause: the last frame was built in a hurry. The creator polished frame one, threw frame two together as an afterthought, then wondered why the model struggled to connect them. Treat the last frame with the same care as the first. The model has to live inside both.
What I actually do is build both endpoints from the same source material before any video tool opens. Same character sheet. Same location reference. Same image-generation session — ideally the same batch — so the light and lens feel related. If frame one comes from a Monday-morning mindset and frame two from a completely different prompt family, the video model inherits the mismatch. Guess who pays for that.
Four Checks Before You Upload Either Frame
- Same light direction. If frame one is lit from camera-left and frame two from camera-right, the model has to flip the sun mid-clip. Sometimes it does. Often it smears the face trying.
- Same scale logic. A wide first frame and an extreme close-up last frame ask for a huge camera push in a short clip. That is a valid shot — but write it as one. Do not pretend it is a gentle drift.
- Hands in a plausible state. If frame one shows empty hands and frame two shows a full coffee cup, the cup has to materialise somewhere. Give it a table, a tray, or another hand to arrive from. Objects that appear from nothing are where viewers spot the trick.
- A clean background at the end. The last frame is the frame your viewer sits on when the action settles, and it is the frame the next shot cuts from. Busy, half-morphed backgrounds draw the eye exactly where you do not want it.
A useful test I run: put both stills side by side and ask whether a real camera and a real actor could travel from one to the other in the time you are giving them. If the honest answer is “only with a cut,” then you do not have a keyframe problem. You have two shots pretending to be one. Split them. I would rather generate two clean four-second shots than one nine-second generation that spends six of those seconds apologising for an impossible move.
One more habit that pays off: export your last frame at the same aspect ratio and resolution you intend to generate at. Crop or rotate late and you force the model to improvise framing on top of everything else. Decide the frame in the still. Let the video model spend its effort on motion — that is its actual job.
The Two-Frame Prompt: Describing Only The Journey
Here is where most people waste the whole feature. They upload two strong frames and then write a long prompt re-describing everything already visible in them — outfit, room, hair, lighting — and contradict the images in three small ways without noticing. The model then has to choose between its eyes and your words. It splits the difference. And the difference looks like morphing.
With endpoints supplied, I keep prompts short on purpose. The stills carry identity, wardrobe, location, and light. The prompt carries only what a still cannot show:
- The single main action — what changes between frame one and frame two.
- The camera behaviour — locked off, slow push, pan left, handheld drift.
- The pace — settles, snaps, eases in, holds at the end.
- Anything that must not happen — keep negative lines tight and specific, not a shopping list of fears.
A working example from this style of workflow. Frame one: character seated at a desk, product box closed in front of her, camera at medium shot. Frame two: same desk, box open, product angled toward camera, her hands resting either side of it. My prompt is not a paragraph about the room. It is more like: “She lifts the lid with both hands and tilts the product toward camera. Camera holds, very slight push in. Motion settles and holds on the open box.” That is it. The room, the clothes, and the light are already in the frames where they belong.
One detail matters more than length: verb choice. Models read verbs literally. “Moves” gives you wandering. “Places,” “slides,” “lifts,” “sets down” give you endpoints inside the journey — small, checkable actions the model can stage. If your two frames show a meaningful change of pose, break it into the smallest physical steps a person would take and name the main one. Leave the micro-steps to the model. Micromanaging every finger is how you get stiff, robotic motion.
I also write the hold into the prompt when I need it. If the last second needs to sit still — for a cut, a caption, or a product hero moment — I say the action completes and holds. Without that line, many models keep drifting, because drifting is their default. You are giving the clip an ending. That is the entire reason you supplied a last frame.
When Keyframing Fights You: Impossible Journeys
Sometimes both frames are good, the prompt is clean, and the result is still a mess. Nine times out of ten the journey was impossible in a way that was not obvious as stills. These are the three patterns I check first.
| Symptom In The Middle Frames | Usual Cause | Fix That Actually Works |
|---|---|---|
| Face melts, then reforms | Head turns too far between frames, or light direction flips | Build an intermediate frame, or reduce the turn to what one clip can stage honestly |
| Object pops in or vanishes | Prop exists in only one endpoint with no source in frame | Restage the still so the prop enters from an edge, a hand, or a surface in both frames |
| Limbs bend the wrong way | Pose change requires crossing the body or reversing a joint mid-move | Change the action, not the wording — pick a path a real body would take |
| Whole scene breathes or warps | Endpoints were generated in different styles or lens feels | Regenerate both endpoints together from the same references before touching video |
The intermediate-frame move deserves a word of its own, because it is the cheapest save I know. If A-to-C keeps breaking, generate or build frame B — roughly halfway in pose, position, or camera — and run A-to-B and B-to-C as two shorter generations. Yes, you spend an extra generation. You also stop burning four generations on an A-to-C that was never going to land. In a storyboard-first pipeline this is painless, because B usually already exists: it is the panel between the two panels you chose.
What I skip: trying to bully an impossible journey with longer prompts. Longer text does not add physics. If a character needs to face camera in frame one and be turned away, three metres left, holding a different object in frame two, that is not a prompt problem. That is a missing cut.
Storyboard-First Keyframing: Boards Are Already Endpoints
This is why the feature clicked for me immediately. I was already working this way before I ever used the setting. A storyboard panel is a promise about how a shot starts. The next panel is a promise about how it ends. Keyframing just lets me hand those promises to the model as pictures instead of translating them into prose and hoping.
The practical version looks like this on a real short. I board the sequence first — for a 30-second piece, usually five to seven panels. Each panel gets approved as a still before any video is generated: character checked against the sheet, location checked, product checked. Then generation is almost mechanical. Panel one and panel two become the endpoints for shot one where the move is continuous. Where the board calls for a true cut — new angle, new shot size — I generate each shot separately and let the cut be a cut.
That distinction matters, so let me be blunt: keyframing is not a way to avoid editing. It directs continuous moves. The moment your story needs a real cut, make the cut. Creators get into trouble when they stretch one keyframed generation across what should have been two shots, then spend an afternoon in Premiere hiding the morph in the middle.
Character work benefits the most. A face that has to survive a move — turning toward a window, leaning in, reacting — holds far better when both ends were approved as stills first. I check faces against the reference sheet, at full size, before generation. If the endpoint still is even slightly off-model, the video will be worse. Keyframes amplify whatever you feed them, including your mistakes. For the deeper method, the guide on character consistency in AI video covers the sheet system I keep coming back to.
Product shots are the other easy win. A hero product that must end perfectly framed — label readable, angle deliberate — should end on a frame you built and approved, not on wherever the model felt like stopping. Build the hero frame first, then build the start frame that can travel to it. Working backwards from the hero moment sounds backwards. In production it is just planning.
Frequently Asked Questions
Do I Need Both A First And A Last Frame, Or Is One Enough?
One is enough to start. A single starting frame in image-to-video mode already gives you most of the identity and composition benefit. Add the last frame when the ending actually matters — a product landing, a matched cut, a pose the next shot depends on. If the shot is loose B-roll, save yourself the setup time.
My Endpoints Look Right But The Middle Still Warps. What Should I Change First?
Check the journey, not the prompt. Reduce the pose change, check light direction in both stills, and confirm both endpoints came from the same reference set. If all three look fine, insert an intermediate frame and split the move into two shorter generations.
Should The Prompt Describe The Scene Again For Safety?
No. Describe only the action, the camera, the pace. Re-describing visible detail invites contradictions the model then has to resolve mid-clip — which is where a lot of morphing is born.
Short version of this whole method: approve the start, approve the end, write a prompt that minds its own business about everything except the journey. Do that and keyframing stops being a feature you experiment with. It becomes how your shots reliably begin and end where you decided they would. If you have not boarded your next sequence yet, start there — one approved panel per shot — and let your first keyframed clip be the small continuous move between panels one and two. You will feel the difference in the edit immediately.