I generated a clip I loved—perfect lighting, the character's face finally consistent, the camera move exactly right. Then I tweaked one word in the prompt to fix a tiny background detail, hit generate again, and the whole thing changed. New face. New light. New everything. The keeper was gone, and I couldn't get it back.
If that's happened to you, you've run into the single most misunderstood control in AI video: the seed. Understanding the Seedance 2.0 seed is the difference between gambling on every generation and actually steering the tool—keeping what works while changing only what doesn't.
This guide explains what a seed is in plain English, how to lock one to preserve a look, how to spin out controlled variations for A/B testing, and how to reproduce a video you (or someone else) already made. No jargon you don't need. Just the mental model that turns "roll the dice again" into "change one thing on purpose."
What Is a Seed? (The Plain-English Version)
Every AI video starts from randomness. Before the model draws a single frame, it grabs a big pile of random numbers as its starting point—its "noise." That random starting point is what the seed controls.
Think of the seed as the shuffle number for a deck of cards. Same shuffle number, same order of cards, every time. Change the number, and you get a completely different shuffle. The seed is just an integer—a number like 48213—that tells the model which random starting point to use.
Here's the key consequence, and it's the whole reason seeds matter:
Same seed + same prompt + same settings = the same (or near-identical) video. Change the seed, and you get a fresh random result.
That's it. The seed is the one input that decides whether your next generation is a repeat or a reroll. Most people never touch it, so every generation is a random reroll—which is exactly why their keepers vanish.
Rule of thumb: The prompt describes what you want. The seed decides which specific version of that description you get. You need both to control an output.
How Seeds Usually Work in the Interface
Seedance 2.0 is available through several surfaces—web generators, apps, APIs—and the exact wording of the seed control differs between them. So rather than promise a specific button label, here's the general pattern you'll find almost everywhere, and what to look for.
| What you'll usually see | What it does |
|---|---|
| A seed field (often blank or set to "random") | Leave it empty and the tool picks a new random seed every time |
| A "random" toggle or dice icon | On = new seed each run; off = reuse a fixed number |
| The seed number shown on a finished video | The exact seed that produced that clip—your key to reproducing it |
| A "reuse seed" / copy-seed option | Pulls the last result's seed back into the field |
The important habit isn't memorizing labels—it's knowing that every finished generation has a seed attached to it, even when you never typed one. If you can find and copy that number, you can reproduce or iterate on that exact clip. If the interface hides it, look in the generation details, history, or metadata.
Rule of thumb: The moment you make a clip you like, find its seed and write it down. That number is the only thing standing between "I'll never make that again" and "I can rebuild it on demand."
The Core Technique: Lock the Seed, Vary One Variable
This is the move that separates people who fight the tool from people who direct it.
Once you have a generation you mostly like, you don't want a fresh random result—you want that result with one small change. So you lock the seed (fix it to the number that produced the keeper) and then change exactly one thing in the prompt or settings.
Because the random starting point stays the same, the model keeps most of what you loved—the composition, the face, the overall feel—and applies your single change on top of it. It's the AI-video equivalent of adjusting one slider instead of retaking the whole photo.
A practical loop:
- Generate normally until you get a clip that's 80% there.
- Find and copy that clip's seed. Paste it into the seed field so it's locked.
- Change one variable—swap "red dress" for "blue dress," or "golden hour" for "overcast," or nudge the camera line.
- Regenerate. You'll get the same shot with that one change, not a total reroll.
- Repeat, changing one thing at a time, until it's 100%.
This pairs perfectly with the one-variable discipline from our Seedance 2.0 tips and tricks: change one thing, learn one thing. Locking the seed is what makes that discipline actually work, because it removes the randomness that would otherwise drown out your change.
Rule of thumb: A locked seed turns prompt editing into precision tuning. Without it, every edit is a coin flip that happens to also change the wording.
Generating Controlled Variations (A/B the Right Way)
There are two opposite reasons to touch the seed, and mixing them up is where people get lost.
Goal A — Explore: You don't love your result yet and want genuinely different options. Here you want new seeds. Keep the prompt fixed and let the tool roll a fresh random seed on each run. Every clip is a different "take" of the same idea, and you pick the best starting point.
Goal B — Refine: You have a keeper and want the same shot with a tweak. Here you lock the seed and change the prompt, as above.
The clean A/B test lives inside Goal B. To honestly compare two prompt phrasings—say, "walking slowly" vs. "striding confidently"—you run both on the same locked seed. Now any difference in the output comes from your wording, not from random noise. That's a real experiment. Run them on different random seeds and you're comparing two coin flips, learning nothing.
| Your situation | Seed strategy | Prompt strategy |
|---|---|---|
| Hunting for a good starting composition | New random seed each run | Keep prompt fixed |
| Comparing two prompt wordings fairly | Lock one seed, use it for both | Change only the words you're testing |
| Perfecting a clip you already like | Lock the keeper's seed | Change one variable at a time |
| Making a set of intentionally different takes | Try a few specific seeds, note each | Keep prompt fixed |
Getting your base prompt strong first makes all of this cheaper—a tight prompt plus a locked seed is a scalpel. If your prompt is vague, no amount of seed control saves it. Nail the wording using our Seedance 2.0 prompt guide before you start fine-tuning with seeds.
Rule of thumb: New seed = new idea. Locked seed = same idea, refined. Decide which one you're doing before you hit generate, and you'll stop confusing yourself.
Reproducibility: Rebuilding an AI Video On Demand
"Can I reproduce an AI video?" is a question I see constantly, and the honest answer is: usually yes, if you saved the right inputs.
To reliably recreate a clip, you need the full recipe—not just the seed:
- The exact prompt (word for word, including punctuation and any reference tags)
- The seed number
- The settings (duration, resolution, aspect ratio, any creativity/adherence value)
- The same references (identical uploaded images, video, or audio, if you used any)
- The same model version
Match all of those and you'll get the same clip, or so close it doesn't matter. Miss one—a different model version, a re-uploaded reference file, a settings change—and the result can drift even with the seed locked. Seeds are powerful, but they only reproduce within the same conditions.
A quick technical-depth note on why it's "near-identical" and not always byte-for-byte: generation can involve tiny non-deterministic factors (hardware, batching, floating-point order) depending on how the model is served. For creative work this is invisible—you get the same shot. For anyone expecting a literal pixel-perfect clone every time, know that "reproducible" means "visually the same," not necessarily "mathematically identical."
The practical takeaway is simple: build a swipe file. For every clip worth keeping, save the prompt, the seed, and the settings together in one note. That record is your undo button, your version control, and your shortcut to consistency across a whole project.
Rule of thumb: A seed without its prompt and settings is half a recipe. Save the whole recipe, or you'll only be able to reproduce approximately.
When You Should Ignore the Seed Entirely
Seed control is a power tool, not a rule. Skip it when:
- You're still exploring. Early on, random seeds are a feature—they show you range you wouldn't have thought to ask for.
- You haven't found a keeper yet. There's nothing to lock. Reroll freely until something clicks.
- Your prompt is the real problem. If every seed looks bad, the seed isn't the issue—the wording is. Fix the prompt first.
Seeds earn their keep in the refinement phase, once you have something worth protecting. Before that, let the randomness do its job.
Frequently Asked Questions
What is a seed in Seedance 2.0? It's a number that sets the random starting point for your video. The same seed with the same prompt and settings reproduces the same result; a different seed gives you a fresh random one. It's how you control whether a generation repeats or rerolls.
Where do I find the seed number? Look on the finished generation—most tools display or store the seed used, in the result details, history, or metadata. Even when you never typed one, a seed was assigned. Copy that number to reuse or reproduce the clip.
How do I keep the same look but change one detail? Lock the seed to the number that made your keeper, then change exactly one thing in the prompt or settings. The shared random starting point preserves most of the look while applying your single change.
Can I reproduce an AI video exactly? You can get the same clip if you match the full recipe: the exact prompt, the seed, the settings, the same references, and the same model version. Miss any of those and the result can drift. Expect "visually identical," not always mathematically pixel-perfect.
Does the same seed always give an identical video? Within the same prompt, settings, references, and model version, yes—the same or near-identical. Change any of those inputs and the seed alone won't hold the result steady. Seeds reproduce within fixed conditions, not across changed ones.
Should I use a random seed or a fixed one? Random when you're exploring for a good starting point. Fixed when you've found a keeper and want to refine it or run a fair A/B test between two prompt wordings.
Do different seeds cost more credits? No—the seed itself doesn't change cost. Credits scale with things like duration and resolution, not which seed you pick. See the pricing page for how per-generation cost works.
The Bottom Line
The Seedance 2.0 seed is the quiet control that decides whether you're rolling dice or directing a shot. Master three habits and you've mastered it: use random seeds to explore, lock a seed to refine one variable at a time, and save the full recipe—prompt, seed, and settings—so any keeper can be rebuilt on demand.
Do that and AI video stops feeling like a slot machine. Your good clips stop vanishing. Your edits become precise instead of random. And "reproduce an AI video" turns from a hopeful question into a routine step.
Ready to stop rerolling and start steering? Find a shot you like, copy its seed, and change one thing on purpose:
Start free → Seedance 2.0 AI Video Generator