Seedance 2.0: the practical guide

The one-sentence mental model

Seedance 2.0 is a reference-to-video model: you hand it identity photos, refer to them in the prompt, and it renders the person those photos triangulate — acting, moving, in new scenes. That contract is what separates it from first-frame models, and it's why it holds a face through motion instead of merely animating a photograph.

Identity references, not first frames

Most "image-to-video" endpoints treat your upload as the opening frame: the clip literally starts as your photo and animates away from it — same pose, same framing, same room. Seedance's reference endpoints treat uploads as identity citations: the schema has you reference them in the prompt itself ("@Image1"), and the model composes a brand-new shot starring that person. Practical consequences:

• You can ask for any framing and camera movement — the shot is composed fresh, so a close-up reference still yields a wide shot when the prompt asks for one.
• Several references beat one: a front portrait plus a three-quarter plus a full-body triangulates the person (the full recipe is in our consistency guide).
• The reference decides who; the prompt decides everything else. Spend your words on scene, action, light, and camera.

The photoreal fence: real faces get rejected

The single most expensive thing to learn by surprise: ByteDance's Seedance endpoints refuse reference images of photoreal people — the request fails with a content-policy violation after you've already queued it. We verified it live on 21 August 2026, when the fence was on 2.5 alone and 2.0 took the same faces happily. It is no longer a split: on 3 September 2026 2.0 refused photos it had accepted eleven days earlier, and ByteDance now documents that the 2.0 series does not support direct uploads of reference images containing real-person faces at all. Where that leaves things:

• Photoreal humans → Kling O3 Pro. Both Seedance lanes now refuse them; no policy fence, up to 15-second clips.
• Illustrated and mascot characters → Seedance 2.5. The newer model, clips up to 30 seconds — its policy only bites on real-looking people.

If a tool doesn't route this for you, route it yourself — a Seedance render of a photoreal character isn't a quality risk, it's a guaranteed rejection that still costs you the attempt on some platforms. (Picacho fences this before send: pick either Seedance lane with a photoreal character and a banner offers the one-tap switch to a model that accepts them.)

The outfit trick nobody uses

Because Seedance accepts several reference images, one of them can be a clothing photo — a product shot or flat-lay, no person in it — cited in the prompt as what the character wears. The render then matches the garment pixel-for-pixel: cut, color, print. On models that can't take a clothing image, the outfit travels as a written description instead — colors and logos land, exact stitching doesn't. We live-tested this split in August, before the photoreal fence reached 2.0: for a specific real garment, Seedance 2.0 with the photo riding along was the only reliable route. With a photoreal character that route is closed now — the identity photo is refused whatever else rides along.

Durations and honest economics

Seedance 2.0 renders 5, 10, or 15-second clips at 720p; fal.ai's list price is $0.3024 per second (verified against fal's pricing page, 21 August 2026) — about $1.51 for a 5-second clip. Seedance 2.5 runs $0.4730 per second and reaches 30 seconds. Two budgeting consequences: a failed take costs the same as a good one, so everything that reduces retries (reference sets, validated camera phrasing, policy routing) is directly money; and 15 seconds of 2.0 costs less than 10 seconds of 2.5.

Prompt patterns that hold

• One scene, one action, one camera move. Seedance follows a clean shot brief beautifully and splits the difference on a muddled one — two camera moves in one prompt fight (our camera-movement cheat sheet shows nine that work, as real clips).
• Keep identity words out of the scene prompt. The references own the face; re-describing it in words gives the model two masters. Describe what happens, not who they are.
• State what stays still. "She stands still as the camera orbits" beats "orbit shot" — the model needs to know what is fixed.
• For sequences, continue — don't restart. Seedance accepts a previous clip as reference material, so the next shot inherits the world instead of reinventing it (shipped in Picacho as "Continue this clip", live-tested before release).

When Seedance is the wrong tool

• You want your exact photo animated: that's a first-frame job — use an image-to-video lane, not a reference lane.
• No character at all: pure landscape/product motion doesn't need identity references; cheaper text-to-video models do fine.
• Budget clips in volume: at roughly $0.06/second-class pricing, Kling's budget tier renders five drafts for the price of one Seedance take — draft cheap, finish on Seedance.