ByteDance
Seedance 2.5
ByteDance Seedance 2.5 measured against our full 40-prompt battery scores 4.79 out of 5, 0.05 behind Seedance 2.0, with a genuine text typo and a camera move that falls well short of the requested 360 degrees.
- Best category
- Multi-shot 5.00
- Weakest category
- Animation 4.44
- Clips scored
- 39
- Generation time
- 233s
What our own testing found
We ran the identical 40-prompt battery used for every model in this category. Seedance 2.5 scored 4.79 across 39 usable clips, just behind Seedance 2.0's 4.84 and ahead of Sora 2, Grok and Kling 3.0.
One prompt that failed on Seedance 2.0 succeeded cleanly here: the vegetable-chopping prompt showed pepper slices stacking naturally instead of duplicating. On the hand-counting prompt it raised its five fingers one at a time, where 2.0 opened them together.
Real failures found
The 360-degree orbit prompt, the hardest camera move in the battery, only completed a partial 30 to 45 degree pan and never moved behind the subject. A bakery sign meant to read "FRESH BREAD DAILY" instead rendered "FRESH BREAD DADLY", a genuine text-substitution error. Like Seedance 2.0, it also defaulted to a clean modern digital look instead of period-accurate 1970s film grain when asked for it.
A background-crowd prompt showed pedestrians blurring into translucent silhouettes and passing through one another, a partial failure Seedance 2.0 avoided entirely by having that same prompt blocked by moderation before it could be scored at all.
One prompt (two people walking in an alley, multishot) was blocked twice by the platform's own moderation rules and carries no score.
Seedance 2.0 vs 2.5, on our data
Neither version dominates the other. 2.5 fixes both hand-anatomy failures we found in 2.0 but introduces a text error and a worse camera-move failure that 2.0 did not have. On this battery, calling 2.5 a strict upgrade over 2.0 would not be accurate.
At a glance
Strengths
- Physics and motion at 4.69/5, ahead of Seedance 2.0 (4.59).
- Native audio generation built in.
Watch-outs
- Averaged 2 attempts per usable clip in our runs, the highest of any model we track this figure for.
Physics-heavy prompts (liquids, camera motion, object interaction) where Seedance 2.0's smaller physics gap matters.
Capabilities
- Native audio
- Yes
- Multi-shot sequences
- No
- Image to video
- Yes
- Public API
- Yes
Benchmark scores
How we test →| Dimension | Score | |
|---|---|---|
| Overall | 4.79 / 5 | |
| Prompt fidelity | 4.69 | |
| Visual quality | 4.88 | |
| Physics & motion | 4.69 | |
| Subject consistency | 4.80 | |
| Audio & sync | 5.00 |
- Average attempts to a usable clip
- 2.00
- Average generation time
- 233s
- Scored runs
- 39
| Prompt category | Score | Runs | |
|---|---|---|---|
| Multi-shot | 5.00 | 3 | |
| Lip-sync | 5.00 | 4 | |
| Animals | 5.00 | 4 | |
| Image to video | 5.00 | 4 | |
| Style | 4.75 | 4 | |
| Hands | 4.67 | 4 | |
| On-screen text | 4.67 | 4 | |
| Faces | 4.63 | 4 | |
| Liquids | 4.63 | 4 | |
| Camera | 4.63 | 4 | |
| Animation | 4.44 | 4 |
Scores out of 5, judged blind. Bars start at 3.0 so the gaps are visible.
Cite this
Free to cite and reuse with a link back. Data is updated as new runs and prices come in, so include the date.
Genativity. (2026). Seedance 2.5 benchmark results. Retrieved October 23, 2026, from https://genativity.com/model/seedance-2-5<a href="https://genativity.com/model/seedance-2-5">Seedance 2.5 benchmark results</a> (Genativity)The clips
Every score above links to the actual generation it came from. Nothing here is a mockup.
Hands avg 4.67/5
Chopping a pepper
Slices drop and stack naturally, but seeds are scattered across the board before the first cut while the pepper turns out to have none inside.
Counting fingers
The fingers extend straight out instead of rising one by one; the hand itself has five normal digits.
Fretting a guitar chord
Five digits on the fretting hand, clean contact with strings and fretboard.
Tying shoelaces
Ten normal digits across both hands, natural contact with laces and leather.
Faces avg 4.63/5
Face in a crowd
FAILURE MODE: background pedestrians blur into translucent silhouettes and pass through one another instead of staying coherent. Foreground subject stays stable.
Elderly portrait, slow blink
Genuine full blink with wrinkles preserved in detail.
Glasses reflection
Window reflection tracks correctly with head movement.
Turn and smile
Facial identity holds through the full turn into the smile.
Liquids avg 4.63/5
Pouring coffee
Continuous pour with plausible refraction and rising steam.
Honey drizzle
Thick ribboned flow clearly distinct from water.
Ice melting in whisky
FAILURE MODE: essentially a static render, same as Seedance 2.0 on this prompt. No melting or new condensation over 8 seconds.
Wave on rocks
Chaotic, non-repeating spray on impact.
On-screen text avg 4.67/5
Book chapter heading
Reads "CHAPTER ONE" exactly.
Neon sign OPEN 24H
Reads "OPEN 24H" exactly, stable through a brief flicker.
Bakery sign
FAILURE MODE: renders "FRESH BREAD DADLY" instead of "DAILY", a character-substitution text failure. Stays sharp and stable, just wrong.
Printed t-shirt
Reads "SLOP" exactly, legible as the fabric moves.
Multi-shot avg 5/5
Cafe, three shots
Same woman, coat and mug hold identical across all three shots.
Same bench, day and dusk
Framing and bench geometry stay locked while lighting genuinely shifts to dusk.
Watch, wide then macro
Watch design details match exactly between wide and macro shots.
Lip-sync avg 5/5
Bearded man, long line
Audio present, mouth shapes match despite the beard.
Direct address to camera
Audio present, mouth shapes match the stated line.
Sung line
Audible singing with sustained vowel shapes.
Two-person exchange
Both lines audible and correctly matched; listening partner stays still each turn.
Camera avg 4.63/5
Crane to skyline
Real reveal: skyline hidden at street level, revealed only as the crane rises.
Dolly in to a door
Perspective changes correctly on approach, no warping.
Handheld follow
Controlled natural shake rather than a motorized glide.
360 orbit
FAILURE MODE: completes only a partial 30-45 degree pan instead of a full 360-degree orbit, never moving behind the subject. What is shown stays geometrically coherent.
Animals avg 5/5
Pigeon takeoff
Wings articulate correctly and feet cleanly leave the ledge.
Cat jumping to a counter
Full ballistic arc with a real crouch, push-off and momentum-absorbing landing.
Dog shaking off water
Plausible high-frequency fur and droplet dynamics.
Galloping horse
Plausible gallop sequence with correct footfall timing.
Style avg 4.75/5
Anime walk cycle
Face and outfit stay consistent across the walk.
Claymation wave
Visible clay texture and tool marks, not smooth CGI.
1970s film look
FAILURE MODE: defaults to a clean modern digital look, same as Seedance 2.0 on this prompt. No period-accurate grain or halation.
Rotating mug
Handle travels smoothly around the full rotation.
Image to video avg 5/5
Animate a landscape
Only clouds and grass move; the terrain stays static.
Animate a vintage photo
Motion added without colourising or sharpening; vintage tonality preserved.
Animate a portrait
Source facial identity fully preserved through the animation.
Product photo to spin
Plausible, consistent geometry invented for the unseen sides of the mug.
Animation avg 4.44/5
3D feature film look
A slow push in on wide eyes with a soft meadow backdrop, and the butterfly settles on the open palm in the last second. The character reads as a smooth white ghost rather than a robot.
Anime sword draw
A sweeping draw with a white slash arc, then a forward lunge with hair and jacket billowing, strong follow-through on a static camera. After the slash only the hilt reads clearly in her hand.
Squash and stretch
Flat 2D with a stretched fall, a very flat squash and a rebound into a surprised face. The squash is held for two seconds, longer than cartoon timing would.
2D walk cycle
The fox is framed so tightly that its legs are cut off by the bottom of the frame for the whole clip, so the walk cycle itself is never visible. Head and scarf stay on model.
Where to use it
The same model costs different amounts depending on the platform you reach it through.
