Seedance 2.5 on fal: 30-Second Takes, 50 References, and Why It Still Cannot Use Your Face (October 2026)
Seedance 2.5 offers 4 to 30 second generation from up to 50 reference files, reference, editing and extension tasks, and a price near $0.46 per second at 720p. What it is good at, what it costs, and the real-face restriction confirmed by its vendor.
Seedance 2.5 has the most impressive specification in AI video. ByteDance announced it on 31 July 2026, it reached fal in August (including US-hosted endpoints), and it takes up to 30 reference images, 10 reference videos and 10 audio clips per call, generates from 4 to 30 seconds, and has three task modes: reference, editing and extension. It is also the model people ask us about most, and the answer to "can I put my face in it?" has not changed: no. This guide explains what it is good for, what it costs and why that restriction is not a bug you can work around.
What it can do
- Length: a single take of 4 to 30 seconds, the longest in the mainstream category.
- References: up to 30 images, 10 videos and 10 audio clips (50 files in all), mixed in one request.
- Tasks:
reference(generate guided by your files),editing(modify an existing clip) andextension(continue one). - Resolutions: 480p, 720p and 1080p, with a draft mechanism (
draft_id) for previewing before a full render.
On paper that is exactly the shape of a "put me in this video" tool. The catch is on the input side, not the output side.
What it costs
Seedance 2.5 bills by video tokens at $0.0214 per 1,000. Tokens scale with the number of pixels, the frame rate and the duration, which works out near the following per output second (our arithmetic from the formula; fal does not print a per-second price):
| Resolution | About, per second | 10 seconds | 30 seconds |
|---|---|---|---|
| 480p | $0.21 | $2.10 | $6.30 |
| 720p | $0.46 | $4.60 | $13.90 |
That puts it at three to five times the price of the mainstream models for a clip with no recognisable person in it, because with a person in it the request is rejected.
The real-face restriction
ByteDance's BytePlus platform confirmed on 31 August 2026 that references containing real people's faces are not accepted by Seedance 2.5, and that matches what we found on 2.0: the filter sits in ByteDance's inference layer, upstream of every reseller. The US-hosted endpoints on fal run the same model, so hosting it elsewhere changes nothing. We documented the earlier tests, including why published bypasses stop working within weeks and why stylising a face to get past the filter destroys the performance you wanted to copy, in why Seedance blocks real faces. We have not spent money on a new bypass attempt and will not.
Our catalogue keeps all Seedance 2.5 rows marked deprecated for exactly that reason, so nobody re-litigates it every time a new version number appears.
What it is genuinely good for
If no real person's face is in the request, Seedance 2.5 is a capable and unusually controllable generator:
- Long single takes of scenery, products, animals and stylised characters, up to 30 seconds.
- Reference-guided motion: pass a video to guide movement and audio to guide timing, on a subject that is not a real person.
- Extension: continue a clip you generated, which helps with the 15-second ceiling of other models.
For these jobs it competes with FLUX.3 video at $0.085 per second and Wan 3.0 Prime, both far cheaper. The question to ask is whether you need its specific strengths (30 seconds, many references) enough to pay five times as much.
What to use instead for a real face
| Job | Option |
|---|---|
| Put yourself in a new scene | HappyHorse reference-to-video, Kling O3 reference-to-video |
| Copy a dance onto your character | Kling Motion Control |
| Replace a person in an existing clip | HappyHorse video-edit, Recast, Kling O3 edit |
| Make a photo talk | OmniHuman 2.0, Kling Avatar |
A caution about demos
Seedance demos are gorgeous, and many of them contain people. They are either generated characters (no real likeness), approved partners, or footage the model accepted because the face was not recognised as a real person's. None of that is a promise that your photo will work. Test with a clip you can afford to lose, and read the rejection message rather than retrying the same input: it names the policy, and the policy does not change between attempts.
Common questions
Can Seedance 2.5 use my face if I host it myself or via a reseller?
No. The restriction is applied in ByteDance's inference layer, upstream of every host, and BytePlus confirmed it for 2.5 on 31 August 2026.
What is a task mode?
reference generates guided by your files, editing modifies an existing clip and extension continues one.
How is it priced?
By video tokens at $0.0214 per 1,000. Tokens scale with pixels, frame rate and duration, which works out near $0.46 per second at 720p.
What should I use for faceless long takes?
Wan 3.0 Prime for up to 30 seconds, or FLUX.3 video, both far cheaper than Seedance.
Is there any workaround?
None that lasted. Published bypasses were patched within weeks, and stylising a face enough to pass destroys the performance.
How to do this without code: Because Seedance rejects real faces, the autopilot uses models that accept them. In AutoHustle this is the autopilot: paste the Reel link, add one photo of yourself and confirm the plan (about 4 credits per second of finished video, so 60 credits for a full 15-second call; analysing the Reel is free). See How to put your face in any Reel and Recreate an Instagram Reel.
Related reading: Seedream 5.0 Pro on fal: A Cheap Image Editor With Three Tiers, and the Face Question (October 2026) · What Is HappyHorse? Alibaba's Video Model, Explained From the Schema
For developers: the fal endpoints
| Model | fal endpoint |
|---|---|
| Seedance 2.5 reference-to-video (US) | bytedance/seedance-2.5/us/reference-to-video |
| Seedance 2.5 image-to-video (US) | bytedance/seedance-2.5/us/image-to-video |
| Seedance 2.5 text-to-video (US) | bytedance/seedance-2.5/us/text-to-video |
Billing: $0.0214 per 1,000 video tokens; tokens ≈ width × height × fps × seconds / 1024. Reference-to-video tasks: reference, editing, extension. Read on 18 October 2026.