Video
First and last frame video generation, explained

Key takeaways
First and last frame generation pins an image to each end of a clip, and the model invents the motion between them.
It's the mode to reach for when a shot has to land somewhere specific, as in a transition, a loop, a product reveal, or a cut that has to match the next clip.
The best frame pairs share a scene, a camera, and lighting. The simplest way to get that is to make the end frame by editing the start frame.
Write the prompt about motion between the frames. The images already describe the scene, so describing it again adds nothing.
Keyframes fix where a shot starts and ends. They don't fix what happens in between, so morphing, odd paths, and broken limbs still need re-rolls or more keyframes.
What is first and last frame video generation?
First and last frame generation is an image-to-video mode where you supply two stills, one for the opening frame and one for the closing frame, and the model generates every frame between them. It's also called start and end frames, keyframe mode, or keyframe interpolation, depending on the tool.
The FLUX 3 Video guide from Black Forest Labs says "You pin both ends of time and the model generates the in-between." Both stills appear on screen exactly as supplied, at the start and the end. Everything else is the model's call.
Our explainer on text-to-video vs. image-to-video covers the single start frame. This piece adds the second one.
How the model fills the gap
A plain image-to-video clip has one fixed point, so the model can take the shot anywhere it likes after the first frame. Adding a last frame gives it a destination. The model now has to plan motion that begins at one picture and arrives at the other, inside a fixed number of frames.
That number depends on the clip length. BFL's API places the closing frame at "duration × 24," which is why it needs a whole-second duration for a start-and-end clip. Google's Veo 3.1 on Vertex AI accepts 4, 6, or 8 seconds for first and last frame clips, and the last frame is optional there, so the same request shape covers plain image-to-video too.
Some newer models go past two frames. Kling's comparison of Kling 4.0 and 3.0 says 3.0 offers start and end frames, while "Kling 4.0 adds Multiple Keyframes, supporting up to 10 keyframe images." Luma's Ray3.2 announcement describes "up to 16 keyframes within a single clip." BFL's FLUX 3 Video takes keyframes at any frame index inside the clip, and the open-weight LTX-2.5 model card lets you condition on an image at a chosen frame index with a strength value, frame 0 being the first.
When to use first and last frames
The mode is slower to set up than a single start frame, because you have to make two images that fit together. It pays for that in four kinds of shot.
Transitions are the classic case. Think of a day exterior that ends at night, a sketch that ends as a finished render, a closed box that ends open. BFL's own example is a staged transformation, and its advice is to "name the stages and their order, and keep the camera still so the transformation is the only motion."
Then there are loops. A clip that ends on its own first frame can repeat without a visible cut, which suits backgrounds, idle character animations, and product spins. Some APIs build this in as a switch. Luma's API docs include a loop setting, for example. On models without one, the usual approach is to put the same image in both slots, though identical ends give the model little reason to move, so the prompt has to ask for the motion clearly.
Controlled motion comes next. A plain image-to-video clip might pan left when you wanted it to push in. If the end frame shows the tighter framing, the camera has to get there. The same goes for a character who needs to finish facing the camera, or a product that needs to end on its logo.
Last, stitching. If you want clip B to pick up exactly where clip A ends, make the boundary image first and use it as clip A's last frame and clip B's first frame. The cut then lands on the same picture from both sides. Make that boundary frame as an approved still, though. Grabbing the last frame of a generated clip is the riskier route. Chaining clips off each other's final frames compounds identity drift, because each generated frame carries small errors into the next clip.
How to pick a good frame pair
The pair should look like two moments from the same shot. That means the same location, the same lens and camera height, the same light, and the same subject in the same outfit. Change one thing between the frames (a pose, a position, a time of day) and leave the rest alone.
BFL's guide is direct about how to get there. "The morph pair is one generation plus one edit, not two generations." Generate the start frame, then make the end frame by editing it. If you generate the two ends separately, small differences in the background, framing, or face all become motion the model has to invent. In BFL's words, the video "has to morph the room along with the swan."
A few other checks help.
Match the aspect ratio and resolution. Runway's API notes that for several hosted models, "output aspect ratio follows the first frame" when both keyframes are provided, so a mismatched end frame gets fitted to the first one.
Keep the change physically plausible for the clip length. A person can turn around in four seconds. A room can't rebuild itself convincingly in four seconds.
Keep important details visible in both frames. A hand that's hidden in the first frame and visible in the last has to appear from somewhere, so the model ends up inventing it mid-clip.
Leave the camera alone unless the camera move is the point. Moving the subject and the camera at once doubles what the model has to solve.
How to prompt the in-between
Once both frames exist, the prompt has one job left. BFL's version is "Prompt the motion, not the scene," and its guide adds that "your prompt's job is what happens next." Describe who moves, in what order, how fast, and what the camera does. Don't re-describe the room or the character, since both images already show them.
The ComfyUI guide for the open-weight Wan 2.2 first-and-last-frame workflow gives the matching advice. "Write appropriate prompts according to your first and last frames." A prompt that asks for motion the end frame contradicts, like a walk to the left when the last frame has the subject on the right, gives the model two instructions it can't both follow.
For a transformation, list the steps. "The candle burns down, the flame shrinks, then goes out" gives the model a sequence, where "the candle changes" leaves the order to chance.
What keyframes can't fix
Pinning both ends doesn't control the path between them. In Generative Inbetweening, a 2024 paper by researchers at the University of Washington, Google DeepMind, and UC Berkeley, the authors found that when two keyframes have sparse correspondences, minimal overlap "leads to artifacts in the intermediate frames." They also note that the base video model they built on "struggles with non-rigid, articulated movements," and that a classical interpolation method takes "the shortest path between the end frames," which isn't always the path a real object would take.
Those findings match the failures creators run into.
Morphing: When the two frames differ too much, the model blends one into the other. Faces melt into new faces and objects slide into different objects. Our guide to fixing flicker and morphing covers what to try.
Shortcut paths: A character who should walk around a table can drift through it, because a straight line is the cheapest route between the frames.
Limbs and hands: Articulated motion between two poses is where bodies break, and the pinned ends don't help the middle.
Run-to-run variance: BFL marks end-frame pinning as experimental and says to "expect more variance between runs than the single-keyframe path, and re-roll a miss."
There's also a trade-off with identity tools. Some APIs make you pick between keyframes and reference images. Runway's API reference says of models such as Wan 3 and Seedance 2 that keyframe mode and reference mode "cannot be mixed," and BFL's FLUX 3 Video returns an error if a request carries both. So if a face has to hold across the clip, it has to hold in your two frames, since a separate reference set isn't available in that mode.
When a clip still goes wrong, the fixes are fewer differences between the frames, a shorter or longer duration to match the size of the change, a middle keyframe on models that take one, or another roll.
Which models take a last frame
The table below is a snapshot of first-party documentation as of November 2026, in no particular order. Support changes often, so check the current docs before you plan a shot around a mode.
Model | Keyframe support | Documented in |
|---|---|---|
Veo 3.1 and Veo 3.1 Fast | First frame, optional last frame, 4, 6, or 8 seconds | Vertex AI docs |
Kling 3.0 | Start and end frames | Kling's 4.0 vs. 3.0 comparison |
Kling 4.0 | Up to 10 keyframe images, 3 to 30 seconds | Kling's 4.0 vs. 3.0 comparison |
Wan 2.2 14B (open weights) | First and last frame workflow | ComfyUI docs |
Wan 3 | First and last frames, not mixed with reference images | Runway API reference |
Runway Gen-4.5 | First frame only in the API | Runway API reference |
Luma Ray3.2 | Up to 16 keyframes per clip | Luma's Ray3.2 announcement |
LTX-2.5 (open weights) | Images at chosen frame indexes | LTX-2.5 model card |
FLUX 3 Video | Keyframes at any frame index, end-frame pinning experimental | BFL video guide |
First and last frame video FAQ
What is the difference between first and last frame and image-to-video?
Image-to-video pins one image as the opening frame and lets the model decide where the shot goes. First and last frame pins a second image as the closing frame, so the model has to plan motion that arrives there.
Can I make a clean loop with first and last frames?
Yes. Use the same image as both frames, or use a built-in loop setting where the API has one, such as Luma's. Ask for clear motion in the prompt, because identical ends give the model little reason to move.
Why does my first and last frame clip morph instead of moving?
Usually the two frames differ too much. Different backgrounds, framing, or faces all become changes the model has to blend. Make the end frame as an edit of the start frame, change one thing at a time, and match the duration to the size of the change.
Can I use reference images and keyframes together?
Often not. Runway's API says keyframe and reference modes "cannot be mixed" for several models, and BFL's FLUX 3 Video rejects a request that carries both. Put the identity you need into the two frames themselves.
How do I join two AI clips without a visible jump?
Make the boundary frame as a still first, then use it as the last frame of the first clip and the first frame of the second. Avoid grabbing the final frame of a generated clip as the next start, because errors carry forward from clip to clip.
