AI Guides › Workbench
By Nigel Guy · 9 min read
The usual way people meet a new video model is a reel of the best clips, a "free unlimited" banner, and a sign-up made on the spot. Then the clips you make look nothing like the reel, because the reel was chosen from many attempts and the banner had conditions, a plan level and an end date in the small print. Seedance 2.5 is a real step forward. It still deserves the same treatment as any other paid tool: check what it does, what it costs, and how long any free offer lasts before you start building with it.
The rule: check the offer before you check the output. Find out which plan, which resolution and which end date apply, then spend a small, fixed budget on one test shot of your own before you commit to anything.
Seedance 2.5 is ByteDance's video model. ByteDance's Seed team released it on 31 July 2026, first in its Chinese apps (Jimeng and Doubao), and other platforms followed. According to ByteDance's own announcement:
| Capability | What ByteDance says |
|---|---|
| Length | Up to 30 seconds of video with sound in one generation (Seedance 2.0 managed 15), and you can extend a clip over several rounds into multi-minute pieces |
| Audio | Generated together with the picture, not added afterwards |
| References | Up to 30 images, 10 video clips and 10 audio clips in a single request |
| Reference types | Normal images plus "clay render" references (plain 3D blocking that sets poses, camera paths and lighting direction), motion references and creative references |
| Editing | Changes tied to timestamps, so you can alter part of a clip without regenerating all of it. Also green-screen background replacement, camera-move edits and edits guided by a reference |
| Realism | Work on skin, eyes, textures, lighting and colour saturation, plus fewer stray subtitles and unwanted background music |
ByteDance also lists where it still falls short: physically believable complex motion, and scenes where several subjects interact. Read that as the place your tests are most likely to fail.
Higgsfield's versions of these numbers differ slightly. Its launch post says "up to 50 references" without the image, video and audio split. It says the model generates at 480p, 720p or 1080p with 4K by upscaling, but Higgsfield's own model page says "native 4K". Go by the launch post and ByteDance's figures until you've seen your own export.
| Route | What it is | Cost at time of writing | Best for | The catch |
|---|---|---|---|---|
| Higgsfield (higgsfield.ai) | A creative suite that offers Seedance 2.5 alongside Kling, Veo, Sora and others, all paid from one credit balance. It adds its own controls: era, genre, lighting, physics, lens and region edit | Seedance 2.5 is listed on Plus (US$47 a month billed annually, about £35, or US$59 monthly, about £44) and Ultra (US$99 a month annually, about £74, or US$129 monthly, about £96). Prices exclude VAT. Its launch post puts a 10-second clip at 30 credits for 480p, 70 for 720p and 120 for 1080p | Trying several models side by side with one login | Higgsfield bills in US dollars. The free plan has 0 credits and no commercial use. Starter only gets Seedance 2.0 Fast and 2.0 Mini |
| Dreamina (dreamina.capcut.com) | CapCut's creative app, from ByteDance's own family | Its Seedance 2.5 page says "free" and "unlimited" and offers "Get started for free". I couldn't confirm the actual free allowance or the paid plan prices | Getting ByteDance's version of the model with the least setup | The page also mentions a 180-second "beta long-video mode". That is Dreamina's claim and isn't in ByteDance's announcement |
| BytePlus ModelArk API | ByteDance's international developer platform. The model ID is dreamina-seedance-2-5-260628 |
You pay per token. The pricing table lists 10.70 per million tokens for 480p and 720p output with no video in the input, and 11.70 for 1080p (US dollars, as far as I can tell). The tutorial suggests keeping a balance above US$30 | Developers building video into their own product | You need a BytePlus account and some coding. Per-token pricing is hard to turn into a per-clip cost until you've run one |
Jimeng and Doubao are where the model launched, but both are built for the Chinese market. Most UK readers can skip them.
Higgsfield's launch post says "unlimited access now runs 33 days on Seedance 2.5", with the exact length and per-clip limits "varying by plan". The page is dated early August 2026, so a 33-day window from then would have closed by early September. Its pricing page still shows an "Unlimited Seedance 2.5" tab, though, so an offer of some kind may still be running. I couldn't confirm the start date, the end date or which plans qualify. Before you pay, open the pricing page and read the tab yourself. Three more conditions from Higgsfield's pricing FAQ are worth knowing:
On every route, read "free" as "free for now" and plan your work as if the offer ends next week.
On Higgsfield: go to Video (higgsfield.ai/ai/video) and choose Seedance 2.5 as the model. Upload your references before you write anything: the faces, outfits, locations and products that must stay the same. Next set the selectors (era, genre, lighting, physics, then camera and lens). Last, write a short prompt that covers only the action and anything the selectors don't. Start at 480p while you're adjusting the prompt and move up to 1080p once the shot works. On Higgsfield's own figures, a 1080p clip costs four times as many credits as a 480p one.
On Dreamina: open the AI Video tool, pick Seedance 2.5, attach your references and write a prompt. Look at the credit or allowance display before your first generation, not after.
On ModelArk: follow BytePlus's Seedance 2.5 tutorial. Send one short test request and compare the tokens it was charged with the pricing table before you write any code that loops.
Whichever route you take, these two prompts work in any general chat assistant to prepare the input before you spend credits.
Prompt 1: turn an idea into a timed shot brief. Fill in the bracketed fields. Leave a field blank if you don't know it yet, and the assistant will ask.
You are a film director's assistant who writes prompts for an AI video model that makes clips of up to 30 seconds with sound.
Goal: turn my idea into one prompt the model can follow in a single generation.
My inputs:
- Idea: [ONE_SENTENCE_IDEA]
- Length in seconds (max 30): [LENGTH]
- Aspect ratio: [16:9 / 9:16 / 1:1]
- References I will attach, numbered: [e.g. Image 1 = presenter's face, Image 2 = shop interior]
- Mood and look: [MOOD]
- Sound: [AMBIENCE / MUSIC STYLE / DIALOGUE, or "none"]
- Things that must not appear: [EXCLUSIONS, e.g. logos, readable text, real songs]
Steps:
1. If the idea, length or references are missing, ask me for them and stop. Do not invent them.
2. Split the clip into timed beats (e.g. 0-4s, 4-10s), one action per beat.
3. Give each beat one camera instruction (move, lens feel, framing).
4. Refer to references only by their number, e.g. "the woman in Image 1".
5. Add one line each for lighting, sound and exclusions.
Output: the finished prompt in one block, then a three-line note on which beat is most likely to fail and why.
Before you answer, check that the beats add up to the stated length, that every reference is used at least once, and that no beat asks for two subjects to touch or swap objects unless I asked for it.
Prompt 2: write a narrow fix instead of a re-roll. Fill in what went wrong and the timestamp where it happens.
You help me fix one problem in an AI-generated video clip without regenerating the whole thing.
Clip summary: [WHAT_THE_CLIP_SHOWS]
The problem: [WHAT_IS_WRONG, e.g. "the label on the bottle is garbled"]
Where it happens: [TIMESTAMP_RANGE]
What must stay exactly the same: [EVERYTHING_ELSE_TO_KEEP]
Steps:
1. If the timestamp or the problem is unclear, ask me one question and stop.
2. Write a single edit instruction that changes only the named problem inside that time range.
3. Say plainly what must not change: characters, camera, lighting, sound.
4. Suggest one fallback if the edit fails, such as regenerating only that beat.
Output: the edit instruction in one block, then the fallback in one sentence.
Check before answering: the instruction mentions one problem only, includes the time range, and does not describe a new scene.