Seedance 2.0 Puts ByteDance's AI Soul in Motion

"The Era of Reference Is Here"

After two days with Seedance 2.0, my personal take: the following groups of people are about to lose their jobs.

First, AI gurus who treat "mastering prompts" as gospel and run classes teaching convoluted, counterintuitive, run-on-sentence prompts.

Second, so-called short video creators whose entire production cost is a selfie stick.

Third, startup founders who just got into AI social.

And I'm absolutely here for it.

One interesting update in Seedance 2.0 is the added ability to reference images, videos, and audio.

The reference video feature especially — according to the official docs, it "supports replication of camera language, complex action rhythms, and creative special effects."

This is a genuine need for me.

Previously, if I saw a complex movement, camera technique, or style, I'd have to figure it out myself for half an hour, maybe even pay for a membership in some reverse prompt engineering software, and the final product still wouldn't match.

This torturous workflow spawned an entire industry of prompt gurus who claim that plain-language prompts are fundamentally flawed — you must employ fake causality terms, pseudo-perspective vocabulary, robustness disruption, and JSON frameworks to get AI video that meets expectations (I am not making this up).

I've had it with these people. When will they understand that paying users have no obligation to learn prompt engineering?

The whole point of AI is using simple natural language to accomplish complex tasks like programming and video production. Yet here they are inventing new jargon, further complicating the workflow, offloading the burden of comprehension that large models should bear onto us consumers.

You people are traitors to humanity.

Now, according to Seedance 2.0, you just upload a reference video to Dreamina, and Seedance 2.0 automatically parses the relevant information and replicates it on demand.

Is it really that good? I tested a few cases.

First, replicating action and camera work. I chose something a normal person couldn't describe: that viral Indian "AAA masterpiece" perfect dodge that was all over the internet recently.

The dodge, head touch, spin-and-reset — plus the push-pull and slow-motion camera work — so bizarre that Douyin creators rushed to imitate it.

Since I happened to see Na Ying herself heading to the Milan Winter Olympics, I had her compete for national glory on the ski big air, pulling off a perfect avalanche dodge mid-run, then executing a high-difficulty aerial spin.

Yeah, Gemini told me Quad Cork 2160 is the highest difficulty aerial trick. I actually have no idea what that means.

After a long wait, Na Ying's Winter Olympics highlight reel was generated.

The ice blocks appearing were a bit jarring, but the opening skiing sequence actually looked pretty realistic;

The perfect dodge sequence did replicate the Indian AAA masterpiece's action and camera work quite well;

Even the BGM matched the action — that was genuinely impressive;

The aerial spin, though, was basically a head-tossing, frame-breaking mess. I didn't bother counting how many rotations;

But overall it was mature and watchable. Seedance 2.0 does have some chops when it comes to replicating action and camera work from reference videos.

Next, I wanted to make Everything Everywhere All at Once starring Sister Yu (Yu Jie), paying tribute to Michelle Yeoh's high-intensity traversal across 100+ parallel universes, while testing Seedance 2.0's ability to replicate stylized, multi-shot videos.

Turns out suspected disgraced artists can't make comebacks — it refused to generate no matter what.

Had to settle for second best — still got our Na Ying to star in Everything Everywhere All at Once (Russian Edition):

What can I say, the vibe was there, pretty funny even.

But it failed to replicate the original's ultra-fast editing rhythm and consistent subject positioning, so it was a failure.

Though this is an ultra-high-difficulty task that normally requires hand-crafting hundreds of frames in Nano Banana to stitch together, so I don't blame Seedance 2.0 for not pulling it off.

Let's try something simpler.

When Sora 2 was hot, there was a trend on Douyin of "new Valorant agent" satire videos.

From Chinese and foreign historical figures to domestic and international celebrities, indiscriminate roasting — every creation sparked a World War III.

I had generated one with Sora 2 back then. My impression: despite being a meme video, the physics were severely broken, the Chinese characters weren't in any dictionary, the viewing experience was fragmented.

How would Seedance 2.0 perform?

Well, I cracked myself up, but the same problems persisted.

The above were all single-video references. Actually images and audio can be referenced too. And you can reference up to 12 at once — straight into the era of mass reference.

This helps a lot with consistency, and gives users more confidence to tell relatively complex stories.

But Seedance 2.0 has a problem: ask too much of it and it falls apart.

Though it gives you 12 slots, the vibe seems to be "slot in two or three to show you tried, don't actually send me a whole essay."

For example, I wanted to expand on Na Ying's skiing video, adding Soviet-era and Stranger Things elements, using Hitchcock zoom and bullet time techniques, and having Na Ying speak in her own voice.

I imported all relevant materials into Dreamina, referencing them as needed.

The result? Final competition footage below:

As you can see, character appearance and voice were maintained well, actions and physics were logical;

The Matrix-style bullet time and Stranger Things Max-style levitation were executed decently;

But the opening Hitchcock zoom and the mid-sequence Indian AAA perfect dodge to evade ice blocks had nothing to do with the originals.

Seedance 2.0 has another problem: not only can it not handle too many commands, it can't handle too many people.

The most crowded place in 2026 is Wonton Tavern, where disillusioned middle-aged men gather to dance on tables and display masculine charm.

So I invited Donald Trump — the world's biggest fan of dancing in public — to Wonton Tavern, to perform a county-town business war with the nun vegetarian restaurant owner and the veteran Japanese restaurant owner.

If you get all these references, I suggest you touch grass

Final result:

The overall plot was roughly there, but Kobe became a clone, the nun started teleporting.

Most crucially, referenced characters like Trump seemed to be on a different layer from everyone else.

Not only was his height Attack on Titan-level, but his body edges had that white fringe of bad chroma keying.

Makes me wonder — does Seedance 2.0 reference character images by separately applying AI animation to uploaded images and pasting them onto video? That can't be right.

Needs improvement.

Also, Seedance 2.0 generates so many shots but doesn't let you adjust individual shots — if you want changes, you have to start over entirely. Such a waste of money. Must criticize.

That said, I have to admit that when generating less complex videos, Seedance 2.0 is already completely sufficient.

For example, I wanted to see if it could replace short drama actors, so I made Episode 1 of I'm a Janitor at the White House.

The texture, the movement, the fabric folds, the expressions, the sound effects, the rhythm of breathing.

Honestly, if I posted this on TikTok, who could tell it's not real footage?

People would think the Epstein emails revealed a new show, and the White House would beg me not to post it internationally.

Whether short drama actors lose their jobs I don't know, but in the AI short drama agent space, Sora 2 is about to be out of a job. Heard OiiOii recently got backstabbed by Sora 2 — recommend urgently integrating with the domestic pride Seedance 2.0 to save themselves.

So for the AI video model track of small casts, ordinary actions, and normal plots, Seedance 2.0 probably does kill the competition, as they claim.

But actually none of this matters.

Including above, when I used this supposedly SOTA model to generate some less-than-satisfactory videos, I didn't feel disappointed.

Because I think Seedance 2.0's greatest update this time isn't model capability at all — it's that ByteDance is finally copying Sora 2's product logic.

Now open Dreamina mobile, and in under a minute you can create your own AI avatar.

Then standard workflow: make video, make same-style, co-create... anyone who's used Sora 2 knows the drill.

Even has built-in characters, including Xavier from Love and Deepspace — I reckon there's gonna be bloodshed over this.

I'm not criticizing this copying — I genuinely think it's good.

When Sora 2 first launched, I made this hot take:

Of course, later I realized this thinking was completely wrong — not only did these creators not delete their accounts, Sora 2's retention rate nearly hit 0%.

But now with Seedance 2.0, I think this take can be revived, and expanded.

Because Sora 2 has no content library or distribution channel, stuck in an endless loop of producing and consuming its own slop. Seedance 2.0 is plugged into the Douyin universe.

In Lark (Xiao Yunque), users can directly copy-paste viral Douyin links as material for AI video generation.

My prediction: this feature will definitely be integrated into Dreamina, and definitely into Douyin, becoming as important as the repost button.

Then what? Then content creation truly has no barrier to entry.

For example, there's a category of so-called comedy creators on Douyin with zero original content ability — whatever's trending, they perform it. Gomoku skills went viral? They sing it. Someone shoots a skit? They copy it. Their entire feed is exaggerated expressions and "POV: When you..." titles — pure human photocopiers.

I used to fume seeing these people succeed and get brand deals through cringe skits, while I was too lazy and too proud to chase that bag. Drove me crazy.

Once Seedance 2.0 is integrated into Douyin, when I see this kind of content, I just say "replace the protagonist with me" — done.

And whether changing scenes, costumes, props, dialogue, or plot, the cost is basically zero — just electricity and token fees.

Short video content officially enters the Yiwu era. Whether Douyin becomes a digital slop concentration camp or breeds a new generation of genuinely talented creators through survival of the fittest — that's up to its own fate.

Beyond that, if Seedance 2.0 is truly integrated into Douyin, ByteDance might actually pull off social.

As we all know, ByteDance has always had a social dream — not only sporadically developing obscure social apps, but also stuffing forgotten social features into Douyin: streaks, flame buddies, ultimately just making everyone laugh.

Fundamentally, nobody has a need to chat on Douyin. The app never developed social functions that WeChat couldn't already do. Sometimes when I forward Douyin videos, I download them and send via WeChat.

But what if Seedance 2.0, and Dreamina's current AI avatar feature, were integrated into Douyin?

Then I'd genuinely only be able to seamlessly create co-creation videos with mutuals on Douyin, and post them without pressure, forming new social topics.

In fact many ByteDance executives have started demonstrating on Dreamina — seems mainly Kelly Zhang going around finding people to co-create with.

Used to type words to talk, now we'll make videos to talk — acting together in videos, I act then you act, can bring others in too, turn chat history into a short drama.

Me and Xianyu acting ⬇️

Wait, isn't this just the metaverse? It's 2D, but virtual avatars, immersive experience, open creation, strong social — nothing's missing.

So Seedance 2.0 is ByteDance's metaverse plan — the "Yuanshen" (Genshin/original god) is truly launching.

Some say this is all my fantasy. Not really — if you scroll Douyin daily like me, you'll see this has already started deploying.

Open Douyin's "Creation Inspiration" feature now, and you'll find you can already invite mutuals to co-create — this feature just hasn't been announced to the world yet.

Once this feature is announced, once Seedance 2.0 is blatantly embedded in Douyin, once all Douyin users can co-create videos with friends, family, celebrities, otome game leads and galgame heroines — the AI social track won't need any entrants.

But the more fake resembles real, the more real resembles fake. The more fakes, the cheaper the app's ecosystem becomes.

So is Seedance 2.0's hyper-realism really a good thing for Douyin?

Hard to say.

Don't end up with ByteDance killing ByteDance — that'd be quite the joy for Kuaishou.

(Cover image generated by ChatGPT, text purely human-written.)