Make the reel where someone spins, snaps or walks past the lens and is suddenly wearing something else, without a closet, a tripod or twenty retakes. Prepare two stills of the same person in two outfits, then let a video model generate the motion that joins them.
An outfit change reel is a small magic trick: someone snaps, spins or walks past the camera and is suddenly wearing something else. Filming it for real takes a closet, a locked-off tripod and a lot of retakes. With AI you work the other way round. You make two photos of the same person in two outfits, hand them to a video model as the first and last frame, and it generates the motion that carries one look into the other.
The format holds attention because the viewer is waiting for the trick. The cut lands on a beat, the clothes are different, and the clip loops straight back so people watch it again to catch how it happened. Creators use it for wardrobe capsules, travel looks, lookbooks and costume reveals.
The traditional version needs identical framing, an identical pose and a hidden cut, repeated for every look. The AI version moves that effort onto still images, which are far easier to control. You prepare two pictures with the same person in the same pose, and a video model that accepts a start frame and an end frame invents what happens between them. Kling v3 Video and Wan 2.7 I2V both do, and both are among the 502 models in the Picasso IA catalog, next to the try-on and editing models that prepare the stills. New accounts get free credits, enough to test whether a look works before you plan a whole series.
One rule before anything else: use photos of yourself, or a fictional character you created. A transition reel built from another real person's face is not yours to make, and a lookalike is no better.
There is more than one way to join two outfits, and they behave differently. The first two rows are the workhorses. The others are for specific looks or for reels with more than two outfits.
| Approach | What you supply | Good for | Watch out for |
|---|---|---|---|
| First and last frame with Kling v3 Video | Outfit one as the start image, outfit two as the end image | Longer, smoother transitions with a spin or a camera move in the middle | The shape of the clip follows your start image, so crop the stills first |
| First and last frame with Wan 2.7 I2V | Outfit one as the first frame, outfit two as the last frame | Short, snappy changes that fit a quick beat | Clips are short, so one outfit pair per clip |
| Multi-shot with Kling v3 Video | A list of shots, each with its own prompt and duration | Three or more looks in one render | Shot durations must add up to the total duration |
| Start frame only, prompt does the rest | Outfit one plus a written description of outfit two | Quick tests when you have no second photo | The model invents the garment, so it rarely matches a real item |
| Two clips joined in an editor | One clip per outfit pair, ending and starting on the same frame | Long reels and full capsule wardrobes | You do the join, and the match frame has to be identical |
If you can supply both frames, do. Pinning the end frame is what keeps the second outfit looking like the outfit you chose instead of something the model improvised. The turn photos into videos page covers the single-image version of the same idea.
This is the workflow from a blank folder to a clip you can post. It assumes one photo of you and one photo of a garment, nothing more.
A transition needs a reason to exist. In real outfit change reels the swap is hidden behind something: a spin, a jump, a hand across the lens, a step through a doorway. Give the video model that same cover and the change looks intentional instead of like a morph. Without it, the model tends to dissolve one outfit into the other in plain view, which reads as a filter.
Keep the camera still unless the camera move is the point. A locked camera means the background in frame one and frame two match, and the only thing the model has to resolve is the person. Describe the speed too: a quick spin reads very differently from a slow turn, and the end pose in your last frame should be one the motion can plausibly reach.
Tip: write the prompt as if you were directing a person, not describing a picture. Spin, step, jump, freeze on the last pose. Verbs the body can do produce better transitions than adjectives about the outfit.
For three or more looks, Kling v3 Video can script up to six shots in a single generation, each with its own prompt and duration. Give each shot one outfit pair and one cover motion, and check that the shot times add up to the total length. If you would rather keep control of the join, generate each pair as its own clip and cut them together yourself.
Most failed outfit changes are not the video model's fault. They start with two stills that do not match. Face, hair, lighting, background and crop should be identical in both frames, and only the clothes should differ. That is why the try-on route works better than generating two separate pictures from scratch: it edits your original photo instead of inventing a new one.
Think about the final shape before you generate. Kling v3 Video ignores its aspect ratio setting once a start image is supplied, so the clip follows your picture. For a vertical reel, crop both stills to vertical first. Check the lighting too: if outfit one is shot in warm window light and outfit two came out cooler, the video will visibly shift tone halfway through.
If you are building a recurring character instead of using your own photos, the techniques from consistent AI characters apply directly. For the garment side, AI virtual try-on for outfits goes deeper on getting a believable fit from a flat product photo.
This format has real limits, and knowing them saves credits.
It suits short, deliberate transitions and is a poor fit for showing a garment exactly as it will arrive in the mail.
Yes. Upload one photo as the first frame and the other as the last frame in a model like Kling v3 Video or Wan 2.7 I2V, add a one-sentence prompt about the movement, and the model generates everything between them. The two photos should show the same person in the same pose with the same framing, since the video model resolves one picture into the other and mismatched faces or backgrounds show up as a visible jump.
It depends on the length you want. Kling v3 Video supports start and end frames, clips up to 15 seconds and a multi-shot mode, which suits smoother, longer transitions. Wan 2.7 I2V also takes first and last frames and produces short clips up to 5 seconds, which suits a quick snap on a beat. Run the same two stills through both and keep whichever hides the swap better, since results vary with the pictures.
Use a try-on model. prunaai/p-image-try-on takes your photo as the person image plus up to six garment images, and it keeps your face, pose and body shape while changing the clothes. That means the second photo comes out in the same pose as the first. If you do not have a garment photo, an editing model such as google/nano-banana can change the outfit from a written description instead.
Yes, and it avoids the consent question entirely. Generate a character with an image model, then build both outfit stills from that same character picture with an editing or try-on model. The video model only needs two matching frames, not a real face. Consistency is the hard part, so keep the character reference image and reuse it for every outfit.
Usually because the two stills were not matched, or because the second one was generated from scratch instead of edited from the first. Make outfit two by running your original photo through a try-on or editing model so the face is carried over, keep the lighting the same, and avoid extreme head angles. If the face still drifts mid-clip, shorten the motion and re-run, because a long spin gives the model more room to wander.
Yes, in two ways. Kling v3 Video has a multi-shot mode that scripts up to six shots in one generation, each with its own prompt and duration, and the durations must add up to the total. Or you generate one clip per outfit pair, making sure each clip ends on the exact frame the next one starts with, and join them in your video editor. The second route gives you more control over every join.
Only with a friend's clear permission, and never a celebrity or any real person who has not agreed. Clothes swapped onto someone's body inside a realistic video is exactly the kind of media that can mislead people, so keep it to your own photos, people who have consented, or fictional characters. If you are unsure whether a face is yours to use, it is not, so pick a different one.
Partly. Wan 2.7 I2V accepts an audio file, a wav or mp3 between 3 and 30 seconds, and times the motion to it, so a beat can drive the spin. Kling v3 Video can generate native audio with the clip. The more reliable route for a precise beat drop is to generate the transition on its own, then line it up against your music in a video editor where you can nudge it frame by frame.
Vertical, 9:16, is the usual choice for reels, but check each platform's current requirements instead of trusting a number written here. Because Kling v3 Video follows the shape of your start image, crop both stills to vertical before you upload them. Wan 2.7 I2V also works from your pictures, so framing the stills correctly first is the simplest way to avoid cropping the finished clip.
Every generation costs credits, and the amount depends on the model and the settings you choose, such as resolution and duration, so any figure written here would go stale. A try-on image is an image generation, and the video clip is the larger spend. New accounts start with free credits, and the pricing page holds the current plans, which is the only place worth trusting for numbers.
Ready to try the swap with your own photos? Open the Picasso IA toolkit, make your second outfit, and pin both stills as the first and last frame.
Choose your language
