Make a 60-Second AI Short Film With the First-Last-Frame Workflow

The sixty-second AI short film is the most under-used format on social platforms in 2026. Longer than a loop, shorter than a reel, structured enough to feel cinematic, small enough to ship in an afternoon. And the generation type that makes it reliably good is the one most creators skip: first-last-frame.

First-last-frame generation lets you specify both bookend frames of a shot and hand the motion interpolation to the model. You get precise narrative control at both ends of every scene, which is exactly what you need when your entire film is only four scenes long. This guide explains what first-last-frame is under the hood, how to prep keyframes for it in Flux 2 Pro and Nano Banana 2, a four-scene hook-build-reveal-tag structure that lands in sixty seconds, and the five pitfalls that ruin most attempts.

What first-last-frame actually does

Most video generation is one-sided. You describe a scene or provide one input image and the model decides where motion ends. You get what you get. Re-rolls are the only recourse, and re-rolls drift.

First-last-frame flips this. You provide two images: the first frame of the shot and the last frame of the shot. The model interpolates the motion between them. The ending pose, the final composition, the closing lighting are all locked by you. The model's creative freedom is restricted to the arc, which is exactly the freedom you want it to have.

In the Versely workflow service this is exposed as the first_last_frame and previous_scene_first_last_frame generation types. The former takes two new keyframes, the latter chains the previous scene's last frame as the first frame of the new shot while you provide only the target end frame. Both produce tighter, more controllable motion than pure text or pure image-to-video.

VEO 3.1 is the best-in-class model for this mode because its motion interpolation handles human faces, subtle camera moves, and lighting transitions more gracefully than any other model in the stack. The VEO 3.1 Fast variant exists specifically for iterating: lower cost per generation, slightly reduced quality, so you can run five variants of a shot, pick the winner, then render the final at full quality.

Keyframe prep: where 70 percent of the quality lives

Every first-last-frame shot is only as good as its two bookend images. Invest here.

Flux 2 Pro handles photoreal keyframes with strong compositional control. Generate the first frame and the last frame separately, not as a pair. Trying to generate a matched pair from a single prompt produces two images that are superficially similar but compositionally misaligned.

Instead: generate the first frame. Lock it. Use Nano Banana 2 to edit that first frame into the last frame. Move the subject three feet to the right. Change the expression from neutral to surprised. Rotate the camera ten degrees. Nano Banana 2's targeted edit capability is what keeps the identity, lighting, and palette consistent between your two bookend frames. Identity consistency between bookends is what makes the interpolation readable as a single continuous shot.

If you need more painterly or illustrated styling, Flux 2 Max is the upgrade and handles the same workflow. For full character edits across multiple scenes, see our piece on character consistency and the I2V fallback chain for why reference anchoring matters.

The hook-build-reveal-tag structure for sixty seconds

A sixty-second film is tight. You have room for four scenes, fifteen seconds each, and that is the whole runway. The proven structure is hook, build, reveal, tag.

Scene	Duration	Purpose	Generation type	Model
1. Hook	15s	Pose a question or tease an image	first_last_frame	VEO 3.1 Fast
2. Build	15s	Escalate, raise stakes	previous_scene_first_last_frame	VEO 3.1
3. Reveal	15s	Pay off the hook	previous_scene_first_last_frame	VEO 3.1
4. Tag	15s	A final beat that recontextualizes	first_last_frame	VEO 3.1 Fast

The hook scene starts cold. A specific image or situation that the viewer wants resolved. "A woman holds a sealed envelope at the edge of a cliff." The first frame is her holding the envelope. The last frame is her hand beginning to open it.

The build scene continues the motion. Its first frame is the hook's last frame, chained automatically by previous_scene_first_last_frame. Its new last frame is the moment of maximum tension. "The envelope is open. A photograph is visible but not yet legible."

The reveal scene pays off. First frame locked to the build's last frame. Last frame shows whatever the photograph contains, now legible to the viewer. This is where your entire film lives or dies.

The tag scene is separate. A cut away. "The envelope drops from the cliff. Camera pulls back to reveal a much wider landscape." First and last frames both new. This is why it uses first_last_frame rather than the chained variant. The tag is a comment on the reveal, not a continuation.

Iterating with the fast variant

VEO 3.1 Fast is where you should live during iteration. Each shot gets three to five fast renders. You pick the winner based on motion quality, not final polish. Then you re-render the winner with full VEO 3.1.

A typical sixty-second film takes:

• 4 final shots at full VEO 3.1 quality.
• 12 to 20 fast variants during iteration.
• 4 to 8 keyframe generations in Flux 2 Pro.
• 2 to 4 Nano Banana 2 edits per scene to produce the bookend pair.

Total compute for a finished film runs well under an hour on modern pipelines. The bottleneck is taste, not GPU time.

The Versely workflow engine handles the iteration loop directly in the AI Video Generator interface. You queue variants, preview them side by side, commit the winner, move on. If you are completely new to text-to-video, start with our beginner primer first and then return here.

Five pitfalls that ruin most attempts

Pitfall 1: Bookend frames that are too similar. If your first and last frames are nearly identical, the model produces a static shot with minor drift. Give the interpolation something to do. Motion should be visible.

Pitfall 2: Bookend frames that are too different. If the subject teleports from a forest to a kitchen between frame one and frame last, the model will produce an incoherent morph. The scene has to be one continuous shot, not a match cut.

Pitfall 3: Prompts that repeat the keyframe description. The keyframes already encode the subject. Your prompt should describe motion, camera language, and atmosphere, not what the character is wearing.

Pitfall 4: Skipping the tag. A three-scene hook-build-reveal film without a tag feels unfinished. The tag is what gives a sixty-second film weight. Do not cut it.

Pitfall 5: Using the full VEO quality for iteration. You will burn compute and time. Iterate on Fast, commit on full. Always.

Sound and polish in sixty seconds

A sixty-second film needs sound or it reads as a tech demo. Three passes cover it.

Narration or dialogue, if any. A single line delivered in the reveal scene is usually enough. Clone your voice once via AI voice cloning or use ElevenLabs for expressive range.

Score. Lyria generates a continuous sixty-second score from an emotional brief. "Sparse piano in the hook, building tension through the build, resolving to a single sustained chord on the reveal, silent for the tag." Keep the mix under minus 18 dB so dialogue sits above.

Captions. Burn them in. Short. One line per scene maximum.

When to stay with first-last-frame and when to switch

First-last-frame is the right tool when you need narrative precision at both ends of a shot. It is the wrong tool when you need long continuous motion, in which case Kling V3 Pro's long-clip capability serves better. It is also the wrong tool for pure establishing shots where you have no strong opinion on the end state, in which case plain text_to_video is fine.

For a comparison of when each generation type makes sense across longer story work, see our guide to long, story-driven AI videos with workflows.

FAQ

Can I use first-last-frame for clips longer than fifteen seconds? VEO 3.1 interpolates up to thirty seconds natively. Beyond that you chain scenes via previous_scene_first_last_frame rather than stretching a single shot.

What happens if my two keyframes have different identities? The model morphs between them. For a surreal short this can be intentional. For a narrative short it reads as a mistake. Use Nano Banana 2 edits to keep identity consistent.

Does the I2V fallback chain apply to first-last-frame? Yes. The chain routes through VEO 3.1 Fast, Vidu Q3, Seedance v1.5 Pro, WAN V2.6, and Kling V2.1 for the interpolation step, preserving both bookend frames across every model.

How many iterations should I run per shot? Three to five fast variants. If the fifth variant is not landing, the bookend frames are wrong, not the model. Go back to Flux 2 Pro.

Can I mix first-last-frame with live-action footage? Yes. Any real photograph can serve as a bookend frame. This is how hybrid AI and live-action shorts are built.

Closing takeaway

The sixty-second AI short film is a format the Versely stack was made for. First-last-frame generation gives you control at both ends of every shot. VEO 3.1 Fast gives you cheap iteration. Flux 2 Pro and Nano Banana 2 give you identity-consistent bookends. A four-scene hook-build-reveal-tag structure gives you the narrative. What you bring is the idea worth sixty seconds of a stranger's attention.