We tested Seedance 2.0 Mini, Fast, 2.0 and 2.5 on first and last frames. One lands the ending almost exactly. One drifts. One invented a language.
*Four Seedance models, one question, and a model that invented a language to fill four spare seconds.*
Most AI video tools now let you set two images: a first frame and a last frame. Start here, end there. It sounds like a solved problem. It isn't quite, and it matters more than it looks.
An exact last frame is what lets clips chain together with no visible cut, what makes a clean transition between two shots, and what makes a perfect loop possible. If the model only gets close, every one of those shows a seam.
The belief we kept hearing from creators: the first frame is honored exactly, the last frame is only approximated. Then we generated a short clip that ended exactly on its last frame, which didn't fit that story. So we tested it properly.
These tests cost real credits. We ran them so you don't have to.
## What we tested
Four models from ByteDance's Seedance family: **2.0 Mini, 2.0 Fast, 2.0 and 2.5**. For every clip we supplied a first image and a last image, then compared what came out against what went in.
## How we measured it
We used two standard image-similarity scores, PSNR and SSIM. You don't need the math, only the scale:
- **Around 35 dB or higher:** effectively the same image, allowing for video compression.
- **Around 10 dB:** two unrelated images.
One detail matters. We judged each clip's last frame against how well that same clip matched its first frame, so every clip is its own ruler. That cancels out differences between scenes.
We also went frame by frame to find when each clip reached its ending, and how much it was still moving afterward.
## Finding 1: Seedance 2.5 lands the last frame almost exactly
We gave 2.5 our hardest pair. The first frame was a screenshot of our own Project Studio video editor. The last was a full-screen, live-action group shot. The prompt asked the editor's preview window to pop out of the interface and come to life.
It did. At both 12 and 15 seconds, the final frame scored **35 dB (SSIM 0.95)** against our image. That is compression-level identical. It doesn't snap onto the image at the last moment, either. It glides in over the final fraction of a second.
[](https://media.aetherwavestudio.com/marketing/project-studio-comes-alive-2026-09-23/project-studio-comes-alive-seedance25-12s-720p.mp4)
*[Watch the clip: Seedance 2.5, 12 seconds](https://media.aetherwavestudio.com/marketing/project-studio-comes-alive-2026-09-23/project-studio-comes-alive-seedance25-12s-720p.mp4)*
## Finding 2: The 2.0 models redraw your frames rather than copy them
Every 2.0 model scored between **15 and 20 dB** on both ends. Visually it is the same shot: the same people, poses and props. Measurably it is a redraw, a little darker and softer than what you supplied.
So the common belief didn't hold. On the 2.0 family, the first frame is not more exact than the last. They came back equally faithful.
## Finding 3: Longer isn't always better
Seedance 2.0 Fast locked its last frame at 4, 8 and 12 seconds. At 15 seconds, something changed.
It reached the ending pose at **8.5 seconds, 57% of the way through**, then spent the remaining 6.5 seconds improvising with natural motion. By the final frame it had drifted to a close approximation: the same scene and people, slightly shifted framing, and a smile showing teeth that wasn't in our image.
Arriving early and ending approximately turned out to be the same thing. Our read: the model runs out of story before it runs out of time, so it fills the gap.
2.0 Mini and 2.0 held their endings at every length we tried.
## Finding 4: Give a model spare seconds and it will fill them
In one 15-second 2.5 clip, a beat described the cast "chatting and laughing." The model took that as permission to talk. It added about four seconds of dialogue that sounds like English and isn't. Our transcription came back as *"Hey, you've got to SoC Works on Here Live? Yes."*
We checked whether it might be another language. Forcing the transcriber to Chinese returned the same English-shaped words, which doesn't happen with real Mandarin. The model produced the sound of casual English without committing to any words.
The line we actually scripted, *"And cut. Perfect, people. Take five,"* landed right on time. We now write "no dialogue" on any beat that shouldn't speak.
[](https://media.aetherwavestudio.com/marketing/seedance-last-frame-blog-2026-09-23/seedance25-15s-invented-dialogue-subtitled.mp4)
*[Watch the clip with sound on: the invented dialogue, subtitled](https://media.aetherwavestudio.com/marketing/seedance-last-frame-blog-2026-09-23/seedance25-15s-invented-dialogue-subtitled.mp4)*
## Finding 5: Time-coded prompts work
We wrote prompts as timed beats: *"0-2s: she poses. 2-3.5s: a voice offstage says... 3.5-5s: she walks out of frame."* On both Seedance 2.0 and 2.5, the clips hit those moments almost to the second. We checked frame by frame and against transcripts.
[](https://media.aetherwavestudio.com/marketing/amy-wardrobe-change-2026-09-23/amy-wardrobe-seedance2-12s-720p.mp4)
*[Watch the clip: Amy's wardrobe change, Seedance 2.0, 12 seconds](https://media.aetherwavestudio.com/marketing/amy-wardrobe-change-2026-09-23/amy-wardrobe-seedance2-12s-720p.mp4)*
## What it costs
Frontier video models aren't cheap, and cost is one of the most common things people ask us about. Here is a 12-second clip at 720p on each model, in AetherWave credits:
| Model | Credits per second | 12-second clip |
|---|---|---|
| Seedance 2.5 | 89 | 1,068 |
| Seedance 2.0 | 58 | 696 |
| Seedance 2.0 Fast | 35 | 420 |
| Seedance 2.0 Mini | 12 | 144 |
Setting a first and last frame adds nothing to the price, and audio is included on all four.
So is 2.5 worth more than seven times the price of Mini? That depends on what you need the ending for.
- **Seedance 2.5:** when the ending has to match your image, for seamless chains, transitions and loops. The premium buys precision.
- **Seedance 2.0 Fast:** for short clips up to 12 seconds. It lands the ending there for 61% less than 2.5.
- **Seedance 2.0 Mini:** when "the same shot" is enough and exact pixels aren't. It held its ending at every length we tried, for 87% less than 2.5, though it tends to ease into the final pose on long clips.
- **Seedance 2.0:** when you want a steady ending on longer clips without paying for 2.5.
A note on scope: we measured how well each model lands the frames you give it, and how it behaves at the end of a clip. We didn't run a general picture-quality comparison, so none of this is a claim about which model looks best overall.
Our credit bundles and plans are on the [pricing page](https://aetherwavestudio.com/pricing).
## What we haven't tested yet
We'd rather tell you the limits than have you find them:
- **Each result comes from one clip per setting.** We haven't yet measured how much results vary from run to run.
- **2.5 and 2.0 were tested on different scenes**, so this isn't a controlled head-to-head. The direction is clear; the exact gap isn't.
- **We haven't tried 2.5 at 4 or 8 seconds**, or in vertical 9:16.
- **Loops are next:** the same image as both the first and the last frame.
## Try it yourself
- **Set a first and last frame** in the [video studio](https://aetherwavestudio.com/video-studio).
- **Chain scenes into something longer** in [Project Studio](https://aetherwavestudio.com/creator-studio/project-studio.html).
- **Build a character like Amy and give her lines** in [UGC Studio](https://aetherwavestudio.com/ugc-studio).