Seedance 2.5 vs Wan 3.0 (2026): One 25-Second Prompt, Both Full Clips Published

Runs: 2026-08-13 (Seedance 2.5) and 2026-08-25 (Wan 3.0) · Published: 2026-08-29

The short answer

We wrote one brief: a 25-second, single-take fantasy sequence with seven timecoded beats. We ran it, word for word, on the two models on our platform that can hold a shot that long: Seedance 2.5 and Wan 3.0. One take each, no re-rolls, both full clips embedded below.

• The style split is the headline. The prompt demanded Hollywood live-action texture (ARRI camera, anamorphic 35mm, film grain). Seedance 2.5 delivered photorealism that holds up frame by frame. Wan 3.0 rendered the same story as stylized game-CG animation from the first frame to the last.
• Beat adherence favored Seedance 2.5. It hit the seven timecoded beats within a second or two. Wan 3.0 ran about two seconds behind through the middle of the clip and skipped one physical event entirely.
• The cheaper model wins real points too. Wan 3.0 drew the floating-island world in visibly richer detail, kept a story prop (the broken chain) through second 17 that Seedance 2.5's sampled frames do not show. And the identical 25-second 720p run prices at 99 credits versus 383 as of publication.

The brief both models received

The prompt is a Chinese-language, single-shot brief: a sixteen-year-old flyer and a twenty-meter chained storm dragon on a floating market island, seven beats with explicit time ranges (0–3s chain break, 3–6s stalls smashed, 6–10s rooftop chase, 10–12s cliff dive, 12–16s freefall, 16–19s landing on the dragon, 19–25s burst above the clouds into a pull-back wide), plus a style block requesting live-action fantasy texture. Both models received the identical text; only the platform's standard parameters (25 seconds, 720p, 16:9) were set alongside it.

整体风格：好莱坞奇幻大片实拍质感,写实风格CG,电影感构图,ARRI摄影机拍摄,35mm变形宽银幕镜头,浅景深;整体色调:黄昏暖金色为主调,阴影偏青蓝色,金橙与青蓝互补色对比;光线:低角度夕阳侧逆光,轮廓光明显,云层体积光;质感:皮革、鳞片、石材纹理细节丰富,材质有真实反光;胶片颗粒感轻微,柔和光晕,高动态范围,无过度锐化,参考《阿凡达》《驯龙高手》真人版的影像质感。史诗奇幻电影,一镜到底无剪辑,25秒连续镜头。架空世界:漂浮在万米云海之上的浮空岛屿,岛上是石砌市集小镇,岛底垂挂巨大岩石根系与瀑布。主角:16岁东方少女,深色短发,额头推着黄铜圆框飞行风镜,穿棕色皮质飞行短装;巨龙:二十米长青蓝色鳞片风暴龙,颈部有深色鬃鳞和半截断裂铁链,半透明翼膜泛金光。两者形象全程严格保持一致。第0-3秒: 市集中景,巨龙被粗铁链锁在石柱上,周围商贩叫卖、行人穿行;巨龙猛然暴起长啸,铁链崩断,铁环飞崩迸出火星。
第3-6秒: 断裂的锁链横扫市集,砸翻一排水果货摊,果子滚落、帆布棚飞起,人群惊叫四散;镜头快速向右横摇,跟住巨龙冲进石板街道。
第6-10秒: 巨龙拖着半截断链沿狭窄街道狂奔,撞开木桶和吊灯笼;少女逆着逃散人流翻身跃上屋顶,在连绵赤陶瓦面上全速奔跑追赶;镜头与她平行横移跟拍,背景云海与远处浮岛快速掠过。
第10-12秒: 巨龙冲到浮岛断崖边缘,张开双翼纵身跃下悬崖俯冲入云;少女紧随其后从岛缘飞身跃下。
第12-16秒: 镜头紧跟少女垂直向下俯冲,自由落体,画面中浮岛底部倒悬的岩石根系、藤蔓和细瀑布飞速上掠,流云扑面,强烈速度感;少女收紧身体,逐渐逼近下方展翼滑翔的巨龙。
第16-19秒: 少女稳稳落在巨龙背上,双手抓住颈部鬃鳞,巨龙猛然振翅拉起;镜头围绕龙身环绕半圈,两者一同扎入厚厚云层,云雾扫过镜头,能见度骤降。
第19-25秒: 冲出云海的瞬间画面骤亮,金色夕阳铺满天空;镜头缓缓后拉并升高成大全景:少女骑在振翅巨龙背上,前方无尽云原悬浮着大小浮岛,鲸形飞艇在云间缓缓游弋,飞鸟群掠过,逆光剪影,史诗感定格。
全程一个连续镜头,无剪辑无转场,各时间段动作与运镜自然衔接,锁链断裂与货摊翻飞有真实物理细节,云层有体积光,电影级动态光影。

Both clips, unedited

These are the complete outputs, recompressed for the web but otherwise untouched. Frame comparisons and the spec table below were measured on the original files, and both originals are archived unmodified.

Style obedience is the biggest split

The two clips differ most in something more basic than sharpness or motion: whether the model listened to the style block. The prompt spends its entire opening paragraph asking for live-action texture, down to a named camera, a named lens format, film grain and real material reflectance.

Seedance 2.5 obeyed. Its market square reads as a location shoot: stone masonry with believable weathering, extras moving naturally through the square, and a dragon composited the way a VFX house would grade it. Wan 3.0 rendered the same scene as a stylized animated feature, with clean-lined half-timbered houses and a rounder, friendlier dragon. It is a handsome look. It is just a different look from the one the prompt asked for, and no frame of the 25 seconds drifts toward photorealism.

The counterweight: Wan 3.0 drew far more of the world. The brief describes floating islands with root systems and waterfalls hanging from their undersides. Wan 3.0 puts them on screen repeatedly and in detail; Seedance 2.5 keeps them distant and suggestion-sized until the finale. If the deliverable were a worldbuilding pitch rather than a live-action sequence, the scoring would tilt the other way.

Beat by beat against the timecodes

Character and prop consistency

Both models kept the girl stable for all 25 seconds: dark short hair, brass goggles pushed up, brown leather flight gear, across seven camera setups. Character drift over a take this long was routine a model generation ago; neither clip shows it on the girl.

The dragon is where they part. Seedance 2.5's dragon stays the same teal, heavy-bodied creature from the market square to the final wide. Wan 3.0's dragon drifts: teal and cartoon-proportioned at second 1.5, noticeably pale (almost white) by the finale. On a brief that says “both characters strictly consistent throughout,” that is a real miss.

But Wan 3.0 takes one meaningful consistency point back. The prompt hangs half a broken chain on the dragon's neck as a permanent prop. At second 17.5, six beats and several camera moves after the chain snapped, Wan 3.0 still renders it on the dragon's neck. In the frames we sampled from Seedance 2.5, the chain is not visible after the market sequence. We sampled seven frames per clip rather than all 600, so we will only claim “not visible in our samples”; the full clips are above for anyone who wants to adjudicate.

Specs and what each run costs

A note on the frame rates: the prompt asked for cinematic texture, and 24 fps is part of how film reads as film. Seedance 2.5's 24 fps output matches that convention; Wan 3.0's 30 fps, combined with its render style, pushes the result further toward the game-engine look. Both are valid choices; 24 fps is simply the one the brief implied.

On price: at publication pricing on our platform, the identical 25-second 720p job costs 99 credits on Wan 3.0 (under its supplier's current −30% promotion; 141 at list) and 383 credits on Seedance 2.5. In dollar terms, a credit costs between roughly $0.018 and $0.030 depending on plan, which puts the Wan 3.0 run under $3 and the Seedance 2.5 run under $12 at any tier. Whether the cheaper version is usable depends on whether your project can live with the style it chose.

Why FLUX 3 is absent

This was planned as a three-model comparison. FLUX 3 caps clips at 20 seconds, so we rewrote the brief's timecodes to fit. The trimmed prompt was then refused by FLUX 3's content guardrails, so no clip was produced and no comparison is possible. We are reporting the refusal as an observed outcome; we do not know which element triggered it, and moderation behavior differs sharply between models. Our measured refusal-rate data shows the same input passing some models and failing others as a matter of course.

The absence matters because of what currently fills the gap. No FLUX 3 comparison with a published prompt and complete, inspectable outputs exists yet; the numbers in circulation come from the vendor's own preference testing, which its materials label a preliminary evaluation. A run with receipts is the missing piece, and we will extend this test to FLUX 3 if the brief can run unaltered.

Method and limits

• One take per model, no selection. Each model ran this prompt exactly once on our platform: the Seedance 2.5 run on 2026-08-13 (generation 2a810059), the Wan 3.0 run on 2026-08-25 (generation 09025ff9). What you see is the first and only attempt, not the best of several. The runs are twelve days apart, so supplier-side model updates between them cannot be ruled out.
• Same text, same parameters. Identical prompt string, 25 seconds, 720p, 16:9 on both. No model-specific prompt tuning.
• Frames were sampled, not exhaustively reviewed. We pulled one frame from the midpoint of each of the seven scripted beats per clip and scored against the prompt from those plus full-clip viewing. Where a claim depends on a frame we did not sample (the chain question above), we say so.
• A single run has limits. This measures what each model did with one brief on one day. Style adherence and beat timing could shift with re-rolls. We chose depth on a single verifiable artifact over averages built from runs that cannot be inspected. That trade-off is the reverse of vendor win rates and has its own blind spots.
• Style calls are subjective. “Reads as live-action” is our reading. Every frame and both full clips are published so you can check it against the evidence.