UNDERSTAND THE WORKFLOW
Raindance AI Character Swap vs Photo-to-Video
A new duet scene, a different face, a replaced character, and a copied performance are four different requests. Choose the input and editing method that match the result you want.
The key difference is what you want to preserve. Photo-to-video builds a scene around your portraits. Face swapping changes facial identity in existing footage. Full-character replacement targets the person, including clothing and body appearance. Motion transfer uses a filmed performance to guide a new character’s movement. These capabilities can overlap inside one product; the label alone is not a quality guarantee.
Four methods, four different starting points
For the Dave and Tems Raindance trend, first decide whether you need the recognizable seaside mood or a specific existing performance. A sunset-pier video can resemble the trend while using entirely new shots.
| Method | What you provide | What it aims to do |
|---|---|---|
| New-scene photo-to-video | Portraits + scene instructions | Generate new people-in-a-scene footage; no original performance to preserve |
| Face swap | Existing clip + face reference | Change facial identity, usually retaining the source body, outfit, and scene |
| Full-character replacement | Existing clip + character references | Recast the person more broadly, potentially including hair, body appearance, and clothing |
| Motion transfer | Driving performance video + character reference | Use the performance to guide movement or expression; the scene may be rebuilt |
The table describes intended scope, not guaranteed results. A recasting workflow may also transfer motion, and a model that accepts photos may support video references too. What matters is which inputs and settings the particular tool actually uses.
1. Photo-to-video: make a new scene
In this workflow, photos tell the AI what the people should look like, while a prompt describes the scene and actions. It is useful when you want a personalized seaside duet without filming yourself or supplying an existing performance.
On this site, two separate adult portraits guide a seated person and a singer. The built-in storyboard requests a full-body pier opening, a singing close-up, and a rear view toward the sunset. The model receives no source video or song reference. See the complete two-photo tutorial for the role labels and actual studio steps.
What to expect: new framing, movement, and cuts. The portraits guide likeness rather than locking every frame. In our sample, planned cuts at four and ten seconds arrived at roughly 3.4 and 7.5 seconds. This is not an exact copy of a filmed dance.
2. Face swapping: target facial identity
A face swap starts with footage that already exists and attempts to change who the face resembles. A face-focused edit is different from changing the entire person. A red cape, a suit, the shape of the body, or the original choreography may remain even when the face looks different.
This distinction gives you a useful acceptance test: if your request is “keep this exact outfit and shot, but change the face,” inspect the face throughout the clip. If your request is “turn this performer into my character, including their clothing,” a face-only edit is too narrow a target.
Our site does not currently offer an existing-video face-swap input. We have not published a face-only benchmark here. The historical failure examples instead come from broader character-replacement experiments and should not be treated as a face-swap comparison score.
3. Full-character replacement: recast the performer
A character swap asks for a broader transformation: not just facial identity, but the hair, clothing, and person’s overall appearance. You supply a source clip and reference images, then check whether the intended character remains consistent across distant views, close-ups, and camera cuts.
Higgsfield’s Genjutsu guide describes separate Motion Transfer and Object Swap workflows, working with real or generated footage. Its explanation is a useful reminder to distinguish rebuilding a scene around motion from targeting an element of a source shot. Those are documented product capabilities, not results we have independently verified.
Our own historical recasting tests show why whole-sequence review matters. A Seedance 2.0 Mini render replaced the foreground and close-up characters, but left the old distant performer. A stronger prompt did not fix that distant identity. A separate standard Seedance 2.0 test improved identity replacement but retained a red cape. See the actual frames and settings.
Do not confuse task success with a successful swap. A playable MP4 proves that a render completed. It does not prove that every character, costume, or shot changed as requested. Our historical experiments are not the current public studio workflow.
4. Motion transfer: use a driving performance
Motion transfer uses an existing performance to guide another character’s movement. It is the more relevant category when you care about a particular gesture sequence or expressive performance, rather than only the seaside mood.
KIE’s Kling 2.6 Motion Control documentation specifies an image reference and a motion video as inputs. It allows at most one of each per request. That single-character input pattern should not be assumed to recast an entire two-person, multi-shot duet automatically.
Runway’s Act-Two guide describes a driving performance and a character image or video. It distinguishes body-gesture control with image inputs from facial performance applied to character-video inputs. Its guidance also favors a single subject and uninterrupted shots. These are product-specific requirements; we have not benchmarked Act-Two on Raindance.
Motion preservation, background preservation, and audio preservation are separate questions. A tool can transfer gestures while generating a different setting. It can also retain or add audio without producing exact lyric lip-sync. Read the workflow’s documentation and inspect the output rather than assuming one feature implies all three.
Which method do you need?
- “I want a seaside duet with these two people.” Start with new-scene photo-to-video if new shots and approximate movement are acceptable. Our two-photo studio supports that request.
- “I want this exact clip, but a different face.” Look for a workflow explicitly designed to edit facial identity in existing video. Check every shot and keep expectations for the source outfit and body separate.
- “Replace both performers and their costumes.” You need broader character editing and explicit per-person mapping. Inspect the distant actor as carefully as the close-up; a costume remnant means the edit is incomplete.
- “Use my character, but follow this performance.” Evaluate motion or performance transfer. Start with the documented number of subjects and cuts, and verify timing, anatomy, and audio independently.
Use portraits and footage you have permission to use, follow the provider’s input rules, and label generated scenes appropriately. Changing to an AI-generated source does not establish permission, guarantee acceptance, or prove that recasting will work.
Method comparison FAQ
Is photo-to-video the same as a Raindance face swap?
No. Photo-to-video can create an entirely new scene using photos as appearance references. A face swap targets facial identity in existing footage. Similar-looking outputs do not mean the inputs or editing method were the same.
Can a face swap remove a costume or fix body proportions?
A face-only workflow is not designed to rebuild the whole costume or body. Full-character editing can target those areas, but our historical tests show that clothing remnants and unacceptable proportions can still survive.
Does motion transfer keep the background and song exactly?
That depends on the specific workflow. Some methods rebuild the scene around transferred motion; others aim to edit a subject while preserving the shot. Audio handling is separate and varies by product. Verify both the rendered picture and downloaded audio.
Which method does this Raindance AI website use?
The current studio generates a new fictional sunset-pier scene from two authorized adult portraits and a fixed prompt, using Seedance 2.0 Mini through KIE. It does not accept a source performance video. The song excerpt is added after generation; exact choreography and lyric lip-sync are not included.
Sources and testing scope
Official references checked October 10, 2026: Seedance 2.0 Mini, Higgsfield Genjutsu, Kling 2.6 Motion Control via KIE, and Runway Act-Two. Product descriptions are attributed to their documentation; no comparative quality ranking is claimed.
Our firsthand evidence covers the published two-photo sample and earlier Seedance/Kling recasting attempts. We have not run Higgsfield Genjutsu or Act-Two for this comparison. For the visual evidence and limits of each conclusion, continue to the troubleshooting article.