Consistent AI video and image sequences with Seedance 2.5, no code

Seedance 2.5 is a model from ByteDance's Seed team that generates video and sound together. The BytePlus documentation gives the numbers: one request makes a clip of 4 to 30 seconds, at 480p, 720p or 1080p, in 21:9, 16:9, 4:3, 1:1, 3:4 or 9:16. You can extend a video over several rounds. ByteDance says extension keeps the main characters, environments and narrative pacing consistent, and that this gets you to videos several minutes long. You can also edit a finished video at specific timestamps.

For a no-code workflow you use it inside an app. Outside China that means Dreamina, CapCut's AI creation tool, or CapCut itself. CapCut announced on 31 July 2026 that Seedance 2.5 would roll out to subscriber accounts of people over 16 in Europe, Asia, the Middle East and South America. In China, ByteDance launched it in Jimeng AI and Doubao Pro. Developers can call it through BytePlus ModelArk.

In Dreamina, the path in its own guide is: open the AI Video page, choose Seedance 2.5 as the model, switch on Omni reference, and click + to add your references. Check what your account really offers before you plan around it. Dreamina's guide warns that a model can show up in the picker and still need an upgrade for the settings you chose. It also lists 480p and 720p for base generation in the app, with upscaling as a separate tool, so don't count on 1080p there. Prices and credits change, so look at the current plan page.

One rule shapes the whole workflow. The BytePlus documentation states that Seedance 2.5 does not support directly uploading reference images or videos that contain real human faces. That page describes the developer platform, and I found no Dreamina page that says how the app applies the rule. So I plan as if it applies everywhere. The person is described in words. Everything else (the café, the desk, the laptop, the jacket, the colour grade) comes from reference images.

Every planning prompt below has four parts: the goal, the inputs, the constraints and the definition of done. After that, make one change at a time and look at the result before the next prompt. That's good practice with any AI tool. With video it also saves money, because every generation spends credits.

What to do: before you open Dreamina, fix the story on one page. Who the person is (in words), the places, the props, the look, and a shot list with seconds.

Tool: Claude or ChatGPT.

Plan in blocks of up to 30 seconds, because that's one generation. The QuickBilling story has two. Block 1 is the café: opening the laptop, sending the invoice. Block 2 is a home desk: the payment notification, a small smile, the laptop closing. Block 2 will be an extension of block 1, not a separate video.

Dreamina's guide also documents a Long Video mode of 30 to 180 seconds. I stay with 30-second blocks here, because a short block is cheaper to check and to redo when one shot goes wrong.

Goal: a story bible and shot list for a 50-second AI video made with Seedance 2.5 for an invoicing app.
Inputs: a fictional freelancer in Berlin, late twenties, short curly dark hair, thin round glasses, navy work jacket over a white T-shirt. Block 1 (about 25 seconds): a small café, opens a laptop, sends an invoice. Block 2 (about 25 seconds): a home desk in the evening, a payment notification, a small smile, closes the laptop.
Constraints: describe the person only in words, in one fixed paragraph I can paste into every prompt. List the places, props and wardrobe as separate items, because each will get a reference image. Natural light, calm, realistic. No on-screen text.
Definition of done: the character paragraph, a list of reference images to make, and a timestamped shot list per block in whole seconds with no gaps.

What to do: make one image for each place, prop and look on the list. The café table by a window, the home desk, the laptop with the invoice on screen, the jacket on a hanger, one style frame for the colour grade. No people in any of them.

Tools: any image model you like. Dreamina's guide suggests preparing references in its own image generator and names Seedream 5.0, GPT Image 2 and Nano Banana. CapCut says Seedream 5.0 Pro is available to its paying subscribers. Photos you took yourself work too, as long as no face is visible.

Use the same light and colour words in every image prompt, so the references agree with each other. Then number the files in the order the subjects first appear in the story. ByteDance's guide numbers references by upload order (Image 1, Image 2 and so on) and recommends uploading them in order of first appearance.

Don't upload more than you need. Dreamina's guide makes the same point: more references don't give you more control by themselves, and each one should have a stated job.

A small Berlin café table by a large window, morning light from the left, a closed silver laptop and a flat white on the table, wooden chair, plants in the background, soft natural colours, realistic photo, eye level, no people, no text.

What to do: in Dreamina, choose Seedance 2.5, switch on Omni reference and add the block 1 references in story order. Set the aspect ratio and the duration. Then write the prompt in four parts.

Tool: Dreamina (AI Video, Seedance 2.5, Omni reference).

The four parts come from ByteDance's prompt guide. Say which reference is what. Give a one-sentence summary: subject, location, event, style, camera. Write the plot as a timeline. Then add notes for anything that must stay the same through the whole clip.

Bind each reference in words, such as "the café in @Image 1". Don't leave the model to guess. In Dreamina you link an uploaded file by typing @ and picking it. If your workspace labels the files differently from my example, use the labels it shows. For the timeline, the guide says Seedance 2.5 follows whole-second timestamps and asks you to avoid gaps between ranges. It also says "no subtitles" and "no BGM" style instructions work for text and sound.

References: @Image 1 is the café table and window. @Image 2 is the laptop with the invoice on screen. @Image 3 is the navy work jacket. @Image 4 is the colour and light style.
Summary: a freelancer in a small Berlin café sends an invoice from a laptop, calm and realistic, natural morning light, slow handheld camera.
Character: [paste the fixed character paragraph from step 1], wearing the jacket from @Image 3.
0-4s: wide shot of the café in @Image 1, the freelancer sits down at the window table.
4-10s: medium shot, opens the laptop from @Image 2, the invoice is on screen.
10-18s: close-up of hands on the keyboard, a short pause, a click.
18-25s: medium shot, the freelancer leans back and drinks from the cup, looking relieved.
Notes: keep the face, hair, glasses and body proportions identical in every shot. Light from the left as in @Image 1, colours as in @Image 4. Café sounds only, no music, no subtitles, no on-screen text.

What to do: before you go further, fix block 1. When one moment is wrong, edit that moment. The prompt guide supports edits by time range. Its own example changes a man's action between seconds 4 and 6 and leaves the rest of the video alone.

Tool: Dreamina, with the block 1 video added as the video reference. I can't tell you the exact name of the edit entry in your workspace, so look for the editing option next to the generated video or add the video as a reference and use the prompt below.

Name the seconds, the subject and the one change, and where you can, describe it as a change from A to B. That's the guide's advice too. Two limits from the BytePlus documentation matter here. A video you edit has to be between 4 and 30 seconds long, so you edit a block and never the joined film. And the edited video keeps the length of the original.

Why not just regenerate the block? Because you'd be gambling the face and the light you already got right against one small fix.

This is a video editing task. Edit @Video 1 and change only what is asked; keep all other subjects, composition, timing and style unchanged. From 10 to 12 seconds, change the laptop screen from the invoice form to a "Sent" confirmation.

What to do: for block 2, don't generate a new clip. Extend the one you have. ByteDance's guide lists the phrases that mark an extension: "extend forward", "extend backward", "continue", "continue from" and "extend the story". The launch post shows this exact kind of request: continue from a video and keep the characters, scene, style and sound consistent.

Tool: Dreamina, with the corrected block 1 video as the video reference and the home desk image as the only image.

This is the step that does the most for consistency. A new clip has only your words to go on. An extension also has the person it just drew. Paste the same character paragraph and the same notes anyway.

Three practical points from the documentation. An extension keeps the aspect ratio of the video it continues. Reference videos may add up to 30 seconds in total, one more reason to keep each block under that. And an extension is itself 4 to 30 seconds long. If Dreamina hands you only the new part, join the two blocks in CapCut or any other editor.

Continue from @Video 1 and generate the next 25 seconds. Keep the character, wardrobe, visual style and sound design consistent with @Video 1.
The scene moves to the home desk in @Image 1 in the evening.
0-5s: the same freelancer sits at the desk, the laptop open, warm lamp light.
5-12s: a soft notification sound, the freelancer reads the screen.
12-19s: close-up, a small smile.
19-25s: the freelancer closes the laptop and switches off the lamp.
Notes: same face, hair, glasses and jacket as in @Video 1. Calm, realistic. Room sounds only, no music, no subtitles.

What to do: if you need still images for onboarding screens, a carousel or a case study, take them from the finished video. Open it in a video editor or a player that can save a single frame, pause on the frames you want, and export them.

Tools: the downloaded MP4 and any editor or player with a frame export.

This is the route I'd take for "continuous images". Eight separate image generations drift a little each time. Eight frames from one continuous video start from the same person in the same place, so they tend to match far better. Check them anyway, because a face can shift inside a video too.

Generate at the highest resolution your plan offers, and crop the stills. Don't enlarge them. If you need more pixels, Dreamina has a separate upscale tool.

Goal: choose eight stills from my 50-second video for an onboarding carousel.
Inputs: the timestamped shot list below.
Constraints: one still per story beat, no two from the same second, prefer moments where hands and face are still (less motion blur). Square crops for a carousel.
Definition of done: a list of eight timestamps with the beat each one shows and a one-line caption.

What to do: watch the full video with sound. Then put the stills side by side. Check the face, the hands, the glasses and any text on a screen. Keep the labels.

CapCut says these models come with invisible watermarks, C2PA Content Credentials and visible AI labels, and that its policies forbid the unauthorised use of someone's likeness. Leave those marks in place, and add a short note of your own that the video is AI-generated. The EU AI Act's transparency rules (Article 50) apply from 2 August 2026. I'd label it even without them.

Save the story bible, the reference images in their order, the prompts and the final video in one folder. The next episode can then start at the extension step with the same references, and it stays in the same world.

If the story needs titles, captions or an animated product explainer around it, pair this with how to make motion graphics with AI without code. If you want video generation inside your own product, with review steps, cost limits and labelling built in, that's the kind of system I build in my embedded assistants and cost and safety rails work.