Seedance Lab is our benchmark programme: eight repeatable tests, each with a fixed prompt, fixed settings, a stated reference policy and a scoring rubric, run against every Seedance version we offer. Results are published per run with the generation date, exact settings and the output file. Nothing on this page is a placeholder result; where a test has no published run yet, it says so.
Why a lab instead of a claims table
Most model comparisons on the web repeat vendor claims or cherry-pick one striking clip. Neither tells you what the model does with your prompt on an ordinary day. The Lab exists to produce first-hand, dated evidence that anyone can reproduce: same inputs, same settings, all runs published, including the bad ones.
Methodology
01
Fixed inputs
Each test has a canonical prompt and a reference policy. Runs never reword the prompt per model. Reference assets are Seedancee-owned originals (a Lab character, a Lab product) hosted on our media domain. Runs are generated from Seedancee staff accounts; generations made by site users are never published as Lab runs.
02
Fixed settings
Duration, resolution, aspect ratio and audio are held constant for every run of a test so outputs are comparable across models and dates.
03
Multiple runs
Where the protocol says three runs, all three are published. We report medians and ranges rather than best-of.
04
Human scoring
Every output is watched by a person and scored against the rubric. Where two reviewers disagree, both scores are shown.
05
Dated and versioned
Each run records the model id, provider, submit time and time-to-output. When a model version changes, old runs stay published and new runs are added, so you can see drift.
06
Known limitations
Each run lists what went wrong, because the failure modes are more useful to a prompt writer than the successes.
The eight tests
测试 01
Same prompt across versions
One unchanged text prompt run on every Seedance version to see what actually improved between releases.
暂无已发布的运行
测试内容
Visual fidelity and lighting realism at the same resolution.
Whether the model follows the shot description or substitutes a generic one.
测试流程
Use the canonical prompt verbatim; no per-model rewording.
Text-to-video only, no references.
Run each model three times; publish all three, not the best one.
Record submit time and time-to-output for every job.
标准提示词
A lighthouse keeper in a wool coat climbs a spiral iron staircase at dusk, lantern in hand. Slow handheld follow shot from behind, warm lantern light against cold blue window light, gentle film grain. The keeper reaches the top and pushes open a heavy door as wind hits the frame.
参考素材: None. This test isolates text understanding.
The same original character, defined by reference images, across three different scenes to measure identity drift.
暂无已发布的运行
测试内容
Face and costume consistency across separate generations.
Consistency within a single clip as the character turns and moves.
How many reference images are needed before identity stabilises.
测试流程
Character is an original design created for the Lab (no likeness of a real person, no existing IP).
Three reference images: front, three-quarter, full body. Same set for every run.
Three scene prompts: a market street, a rainy bus stop, a rooftop at night.
Reference-to-video mode on models that support reference images.
标准提示词
The character from the reference images walks through [SCENE], looks over their shoulder toward the camera, then continues walking. Medium shot, natural light, no cuts.
参考素材: Three reference images of the Lab character, identical for every run.
固定参数
时长
6s
分辨率
720p
画面比例
9:16
音频
关
模型
Seedance 2.5, Seedance 2.0, Seedance 2.0 Fast, Seedance 2.0 Mini
评分标准
Identity across scenes (0–4): a reviewer can recognise the same character in all three clips.
Identity within clip (0–3): no face morphing during the head turn.
A product with a simple wordmark, rotated and lit, to see whether text and shape survive motion.
暂无已发布的运行
测试内容
Legibility of a short wordmark throughout the clip.
Shape and proportion preservation during rotation.
Material and reflection plausibility.
测试流程
Product is a Lab-made bottle with the fictional wordmark 'NORTHFALL'. No real brand is used.
Two reference images: front view and 45° view.
Same rotation prompt for every model.
Frames sampled every 0.5 s and checked for text corruption.
标准提示词
The product from the reference images rotates slowly on a matte white turntable, 180 degrees, under soft studio light with a single specular highlight. The label stays sharp and readable. Plain background, no cuts.
参考素材: Two reference images of the Lab product, identical for every run.
固定参数
时长
5s
分辨率
1080p
画面比例
1:1
音频
关
模型
Seedance 2.5, Seedance 2.0, Seedance 2.0 Fast, Seedance 2.0 Mini
评分标准
Wordmark legibility (0–4): count of sampled frames in which every letter is correct.
Shape (0–3): silhouette matches the reference through the rotation.
Lighting (0–3): highlight behaves like a real specular.
A short spoken line and an ambient cue to evaluate generated audio quality and mouth synchronisation.
暂无已发布的运行
测试内容
Whether the requested line is spoken at all, and in the requested tone.
Lip-sync accuracy on a front-facing speaker.
Presence and plausibility of ambient sound described in the prompt.
测试流程
Speaker is the Lab character (original design).
Audio generation enabled; no reference audio in the baseline run.
A second run supplies a Lab-recorded reference voice clip on models that accept reference audio.
Reviewers watch muted first, then with audio, and log timing offsets.
标准提示词
Close-up of the character from the reference images, facing camera in a quiet workshop. They say, clearly and calmly: "We tested it three times. The third one worked." Faint rain on the roof, a clock ticking. No music.
参考素材: Three character reference images; optional one reference audio clip (Lab-recorded).
固定参数
时长
6s
分辨率
720p
画面比例
16:9
音频
开
模型
Seedance 2.5, Seedance 2.0, Seedance 2.0 Fast, Seedance 2.0 Mini
评分标准
Line delivered (0–3): exact words, correct tone.
Lip sync (0–4): offset under roughly 100 ms, no dropped syllables.
A multi-part physical action to expose where motion coherence breaks down.
暂无已发布的运行
测试内容
Limb and object coherence during fast, multi-stage motion.
Cause-and-effect ordering (throw, then catch, not both at once).
Cloth, hair and liquid behaviour under motion.
测试流程
Text-to-video, no references, so every model faces the same problem.
Same prompt, three runs per model.
Reviewers list each failure with a timestamp rather than a single overall impression.
标准提示词
A skateboarder in a loose grey hoodie pushes off, ollies over a low curb, lands, and coasts to a stop while pulling their hood down. Side tracking shot at chest height, overcast daylight, wet asphalt.
A dense prompt with ten checkable details to measure how much of a specific brief each model honours.
暂无已发布的运行
测试内容
Count of explicitly requested details present in the output.
Whether negative instructions ('no text', 'no people') are respected.
Whether colour and count constraints are followed.
测试流程
The prompt lists exactly ten verifiable details; reviewers tick each one.
No references; text-to-video only.
Three runs per model; report the median and the range.
标准提示词
A red bicycle with a wicker basket leans against a yellow door. Three potted ferns on the step. A grey cat sits on the windowsill and yawns once. Morning light from the left, long shadows. The door has a brass number 17. No people, no text overlays, no music. Static camera, slight film grain.
One point per detail (0–10): red bicycle, wicker basket, yellow door, three ferns, grey cat, yawn, left-side light, brass 17, no people, static camera.
Penalty (−1 each): text overlays, added people, music.
Comparison pages (Seedance 2.5 vs 2.0, Seedance vs other models) link to the relevant tests and quote results only from published runs. Model pages show capability facts from the registry; the Lab shows behaviour. When a run reveals a reliable prompting pattern, it is added to the prompt guide with a link back to the run.
Reproducing a test yourself
Copy the canonical prompt from the test card.
Open the workspace, select the model, and set duration, resolution, aspect ratio and audio to the fixed settings.
Add the reference assets if the test uses them (Lab assets are downloadable from published runs).
Generate three times and score with the rubric.
If your results differ materially from ours, tell us; we re-run and, if needed, correct the published run.
Run a test in the workspace
Use the canonical prompt and fixed settings from any test card.
Because we only publish runs we have actually performed, stored and reviewed. The methodology is published first so it cannot be adjusted after seeing results. A run appears under its test only once the output is on our media host, a person has watched it and written the notes and known limitations, and the generation followed the test's fixed prompt, references and settings. Tests without such a run say so.
Which models does the Lab cover?+
Every Seedance model offered on Seedancee: 2.5, 2.0, 2.0 Fast, 2.0 Mini and 1.5 Pro. Reference-based tests skip 1.5 Pro because it does not accept reference inputs.
Will the Lab compare Seedance to other companies' models?+
Where we can run the same prompt and settings on another hosted model, yes, and those runs will be published on the relevant comparison page with dates. We do not quote competitor benchmarks we did not run.
Are the reference assets real people or brands?+
No. The Lab character and the Lab product are original designs created for these tests. No real person's likeness and no existing brand is used.
Can I submit a test?+
Suggest one through the contact page. Good candidates are repeatable, use original assets, and measure something the current eight tests miss.
How often are tests re-run?+
When a model version changes, when the provider reports a capability change, or on a monthly cadence for the flagship model. Each run carries its date.