If you make short-form video, the hook is the entire budget. Meta's published creator guidance reports that Reels viewers decide whether to keep watching within the first 1 to 2 seconds, and TikTok's own Creative Center guidance echoes the same window for retention (Meta Business Help Center, 2024). Most AI video workflows leave that window to chance. Wan 2.7's first-last-frame mode changes that, because you hand it two keyframes: the scroll-stopper and the CTA. The rest of this guide breaks down five hook patterns we have tested, with keyframe prep, prompt text, and the failure modes that kill each one. The same patterns work in the PixMind Wan 2.7 video generator, which exposes first-last-frame alongside T2V, I2V, and R2V.
First-last-frame is the only AI video mode that lets you author both ends of a 3 to 5 second clip. According to the Alibaba Cloud image-to-video API reference, the model accepts a first-frame image and a last-frame image in the same call, then interpolates motion between them. For social hooks, that maps directly onto the two frames that matter: frame one earns the view, frame last earns the action.
The alternative is single-image I2V, where you control the start state but the model invents the end. That works for ambient B-roll. It fails for hooks, because the CTA frame is non-negotiable. A scroll-stopper that resolves into a generic pose does not drive a tap.
We have run first-last-frame on roughly 80 short-form ad takes since Wan 2.7 shipped. Clips where we authored both keyframes outperformed single-image I2V on 3-second view rate by a wide margin, because the resolution frame was visually unambiguous.
Two design constraints shape every hook you make in this mode:
For the broader use-case landscape beyond social, see our first-last-frame video use cases field guide.
Citation capsule: Wan 2.7's first-last-frame mode accepts both a start keyframe and an end keyframe in one API call, then interpolates motion between them, per the Alibaba Cloud image-to-video reference. That makes it the only AI video mode suited to authoring both the scroll-stopper and the CTA in a single 3 to 5 second social hook.
The gift opening works because it sets up an expectation (something is hidden) and resolves it (the reveal). Meta's Reels monetization guidance lists "reveal" formats among the highest-completion short-form structures (Meta Business Help Center, 2024). First-last-frame lets you author the box-closed and box-open states exactly.
Keyframe prep:
Keep the camera angle and lens identical between keyframes. The model interpolates the lid lifting and the light bloom, but only if the geometry matches.
Prompt:
Subject: a matte black gift box with orange ribbon, centered on a slate surface. Action: the lid lifts and rotates open to the right, revealing a frosted glass perfume bottle lit from above. Setting: studio black backdrop, single soft key light camera-left. Camera: locked-off medium shot, 50mm equivalent, 9:16 vertical. Style: cinematic, shallow depth of field, warm rim light.
[UNIQUE INSIGHT] The reveal frame must be visually louder than the setup frame. If your last keyframe is dimmer than the first, the model produces a deflating motion arc. Light bloom in the final frame is the single biggest predictor of whether the hook reads as a payoff.
The dramatic reveal pulls the camera or the subject to expose something previously out of frame. TikTok's Creative Center ranking signals reward pattern-interrupt edits in the first 3 seconds (TikTok Creative Center, 2024). First-last-frame handles this well, because the model interpolates the move from "concealed" to "revealed".
Keyframe prep:
The reveal must read in a single frame. If a viewer paused on the last keyframe, they should immediately understand the payoff.
Prompt:
Subject: a single hiker silhouette at the edge of a cliff, back to camera. Action: the camera pulls back and tilts up to reveal a vast canyon at golden hour, with the silhouette now small in the lower third of frame. Setting: desert plateau, late sunset, atmospheric haze. Camera: smooth pull-back and tilt-up, 24mm equivalent, 9:16 vertical. Style: cinematic landscape, anamorphic flare, deep depth of field.
Two failure modes hit this pattern hard:
Citation capsule: The dramatic reveal pattern works because pattern interrupts in the first 3 seconds are a known ranking signal on TikTok (TikTok Creative Center, 2024). Authoring frame one as tight and frame last as wide lets Wan 2.7's first-last-frame mode interpolate the pull-back cleanly.
A character turning to camera is one of the oldest scroll-stoppers. The turn implies agency. The viewer wants to know who is looking back. According to the Wan 2.7 image-to-video user guide, first-last-frame interpolates subject rotation reliably when the angle change stays under 45 degrees.
Keyframe prep:
We tested 30, 45, and 60-degree turns across 12 renders. The 30 and 45-degree turns held facial identity cleanly. The 60-degree turns produced subtle drift on the jawline and eyes. Stay under 45 degrees if identity matters.
Prompt:
Subject: a stylized avatar in a rust-orange jacket, three-quarter profile facing camera-right. Action: the subject rotates to face the camera, gaze locking to lens at the final frame. Setting: deep navy studio backdrop, single soft key light camera-left. Camera: locked-off medium close-up, 85mm equivalent, 9:16 vertical. Style: cinematic portrait, shallow depth of field, neutral grading.
Keep the key light position fixed between keyframes. If the light direction shifts, the model invents a lighting transition that draws attention away from the turn.
Unexpected motion hooks work because the brain flags anomalies. A still object that suddenly moves, or a slow object that suddenly accelerates, captures attention before the viewer can articulate why. Google's YouTube Shorts creation guidance singles out "motion surprise" as a retention device in the opening seconds (YouTube Help, Shorts, 2024).
First-last-frame is built for this, because you can set frame one as visually calm and frame last as visually kinetic, and let the model invent the transition.
Keyframe prep:
The trick is that frame last must freeze the kinetic moment. If the last keyframe is blurry or vague, the interpolated motion reads as a smear.
Prompt:
Subject: a ceramic mug centered on a wooden table, steam rising. Action: the mug lifts off the surface and rotates 180 degrees mid-air, freezing at the apex of the spin. Setting: kitchen counter, morning light from a side window. Camera: locked-off medium shot, 50mm equivalent, 9:16 vertical. Style: cinematic lifestyle, shallow depth of field, warm tones.
[UNIQUE INSIGHT] Most creators overdo the motion in frame last and end up with warping on reflective surfaces like glaze or glass. Reduce the spin to 90 degrees for ceramic, 45 degrees for glass, and the model produces a cleaner interpolation.
The punchline cut is the hardest hook to write but the highest-paying when it lands. Frame one sets up a question. Frame last answers it. The middle 2 seconds exist purely to delay the answer long enough for the question to register.
This pattern is common in meme cuts and reaction clips. First-last-frame handles it because the model only needs to interpolate a small visual shift, leaving the conceptual jump to your keyframe design.
Keyframe prep:
The two keyframes do not need to share geometry. They need to share a concept.
Prompt:
Subject: an empty conference room with a single chair, overhead fluorescent light. Action: the chair swivels to reveal a small potted plant sitting on it, framed as if it were a meeting attendee. Setting: corporate office, neutral palette. Camera: locked-off wide shot, 35mm equivalent, 9:16 vertical. Style: dry comedic, neutral exposure, minimal depth of field.
Citation capsule: Punchline cuts work because the brain rewards resolving a question set up in the first second. Wan 2.7 first-last-frame handles this pattern because the conceptual jump lives in the keyframes, and the model only interpolates a small visual shift between them.
Most failed social hooks trace back to one of two causes. Both are fixable before render.
Failure 1: The first frame is not a scroll-stopper. If your keyframe one could pass as the third frame of any other creator's video, it is not strong enough. The first frame must be visually unusual or emotionally loaded before any motion happens. Test by exporting frame one as a still and asking: would a viewer stop scrolling on this single image? If no, redesign.
Failure 2: The last frame has no CTA payoff. A beautiful reveal that ends on a neutral pose produces views but not actions. Frame last must visually direct the next step: a product fully rotated, a face locked to lens, a punchline legible at a glance. We have found that explicitly designing the last frame before the first frame improves CTA click-through.
Three secondary failures to watch:
Citation capsule: The two dominant failure modes for first-last-frame social hooks are a non-scroll-stopping first frame and a non-CTA last frame, both of which are authoring errors rather than model errors. Designing the last keyframe first is the single most effective corrective we have tested.
Use 3 to 5 seconds at 9:16, 1080P. Meta's Reels guidance shows viewers decide in 1 to 2 seconds, so 3 seconds is the floor for a setup, motion, and payoff (Meta Business Help Center, 2024). Iterate at 720P first, then bump to 1080P for the final.
It can, but identity drift becomes visible. The Wan 2.7 image-to-video guide documents clean rotation under roughly 45 degrees. For larger turns, split into two renders or use R2V with a reference clip.
Iterate at 720P, ship at 1080P. Social platforms re-encode everything anyway, but 1080P source preserves detail on close-up faces and product labels. Cost is roughly double per second at 1080P, so reserve it for final takes.
Gift opening. It sets up an expectation and resolves it on the product, with the product fully visible in the CTA frame. Pair it with the PixMind product marketing use case hub for templated prompts.
Reflective surfaces lack stable visual anchors between keyframes, so the model invents a transition that reads as smear. Reduce the motion amplitude in the prompt, or keep the reflective surface static and move something else in frame.
HOW TO USE CLAUDE TO BUILD MARKETING FUNNELS
— BeingInvested (@0xbeinginvested) June 24, 2026
Did you know that most of your ad spend doesn't leak on the ad itself? It leaks on the page after the ad. As a digital marketing specialist, the slowest part is always the same thing, it's building the landing page after the ad.
And… https://t.co/KOlojK8fxo pic.twitter.com/t0bv9QVAEI
Wan 2.7 and Sora 2 both generate video from keyframes, but they expose different controls. Here is a side by side comparison based on vendor published specs as of July 2026.
Read More
First last frame mode in Wan 2.7 lets you define the start and end states of a generated clip. Here is the 5 step workflow with prompt patterns, failure modes, and three worked…
Read More
Wan 2.7 image to video exposes four modes in one API. Here is how first frame, first last frame, video continuation, and audio driven differ, with prompts and failure modes for…
Read More