---

Seedance 2.5 is a video model you can run on Pollo AI. What sets it apart from most models: it pulls from several sources in a single render. Multiple image references of you, multiple room angles, separate audio tracks. That is exactly what you need if you want yourself as an avatar that talks and moves like you.

It is no replacement for content with real substance. But for hooks that stop people from scrolling, it is stupidly good.

What you prepare

Seedance builds the avatar from your references. The cleaner the sources, the better the match. You need four things:

• Face: 4 photos from different angles (front, left, right, slightly above). Individual files, no collage sheet.
• Body: 2 shots of you, half and full body.
• Room: 3 photos of your room from different angles, so the background stays consistent.
• Voice: one clean recording of 10 to 20 seconds, spoken calmly. That is your voice identity.

Your first step: Stand near a window, shoot the 4 face and 2 body photos on your phone, then 3 shots of the room. Record the voice separately, no music in the background.

Step 1: Upload references to Pollo AI

1. Open pollo.ai and pick Seedance 2.5 as the model.
2. Upload the images as individual references and give them speaking names (me_face_1, room_1, and so on). You point to those names in the prompt next.
3. Upload the voice recording as an audio reference.

Step 2: My prompt

The trick is the format: an English wrapper for the instruction, German dialogue lines for the lip match, and the lines kept deliberately short. This is the prompt I built the skit with. Swap the @ names for your uploaded references and the dialogue lines for yours.

GOAL: A hyper-realistic 30-second selfie-style skit. The same man appears as two
versions of himself in a room, having a short banter, then talks to camera.
Indistinguishable from a real phone video, not cinematic.

REFERENCES (individual files, NO character sheets; one image per angle):
@me_face_1..4   my face, four angles, separate images
@me_body_1..2   me, half and full body
@room_1..3      my room, three angles for a consistent background
@voice          a clean recording of my voice (voice identity)
One subject, shown as two instances in frame. Keep under 20 reference images.
Keep voice and room tone as separate audio files.

CONTINUITY: same face and skin texture throughout. Give the two versions one
distinguishing detail (e.g. one wears headphones). Same room and lighting across
all stages.

STAGES with German dialogue (keep lines short so the lip sync holds):
Stage 1 (0 to 6s): Version A sits at a desk, working, not looking up.
Version B leans into frame.
 B: "Ich hab da ein Tool gefunden."  A: "Chris."
 B: "Ich hab da ein Tool gefunden."  A: "Nee."
Stage 2 (6 to 12s): B keeps pushing. A stays annoyed.
 B: "Seedance 2.5 ist raus."  A: "2.5, 2.6, 2.whatever. Erzähl mir was Neues."
Stage 3 (12 to 20s): B lists the specs. A finally looks up, interested.
 B: "30 Sekunden am Stück. 30 Bildreferenzen. 10 Videos, 10 Audiospuren. Ein Render."
 A: "Dann sollten wir wohl generieren."

LOOK: natural phone-camera realism, slightly handheld, available indoor light,
realistic skin texture with pores, no cinematic grade, no beauty smoothing.
Believable as a real selfie video.

CAMERA: handheld selfie framing with subtle natural shake and lens breathing.
No stylized moves.

AUDIO: two speaking voices from @voice, natural room tone. No music during the
skit. Synchronous lip movement to the German lines.

EXCLUSIONS: no on-screen text, no logos, no watermark, no skin over-smoothing,
no extra people.

Step 3: Think in two renders

My video runs longer than a single render, so it comes in two parts:

1. Render 1 is the skit above, 30 seconds.
2. Render 2 is the reveal, you alone, straight into the camera. Pass Render 1 as a reference so face, room and light stay consistent.

The reveal prompt is short: same look, you alone, dialogue in. Mine was the line "This video was fully AI. Seedance 2.5 over Pollo AI. A few photos of me, my room, my voice. Prompt in, generated. So, would you have noticed it is not me?"

Common mistakes

• German lip sync. This is the critical part. Seedance mumbles on long German words. That is why the lines in the prompt are short. If pronunciation still trips up: generate the video, then lay your clean German recording over it in the edit, in sync.
• Two clones instead of two versions. If both instances look identical, it reads cheap. Give one version a detail, mine wore headphones.
• Collage sheets as references. Individual files per angle work better than one sheet with four faces on it.
• Thinking too long. Anything over 30 seconds belongs in a second render with the first as reference, not in one prompt.

Next Steps

Start with a 6-second test: one face, one line, a cheap tier. Once the match holds, build the full skit. And the principle is not just for skits. The same way you make UGC hooks, product talks or an intro for any reel, without ever standing in front of the camera.