How to Make AI Videos with Seedance 2.5 (Using Nano Banana for Keyframes)

Monica — Tue, 13 Oct 2026 12:29:38 +0000

Most "AI video" tutorials stop at type a prompt, press generate. That works for a demo, but the moment you need a specific opening shot, a specific ending, or a clip that matches a brand, prompt-only generation turns into an expensive slot machine.

This post walks through a more controlled workflow:

Design your keyframes as still images with Nano Banana.
Animate between them with Seedance 2.5.
Extend or edit the clip if the shot needs it.
Stitch the final shots together locally with ffmpeg.

I'll use Veemo AI for steps 1–3 because it hosts both models in one studio, so there's one account and one credit balance instead of two separate tools. The ideas carry over to any host that exposes the same controls.

What Seedance 2.5 actually gives you

Before writing prompts, it helps to know which knobs exist. On Veemo's Seedance 2.5 page, the model is exposed through four operations:

Operation	Input	Use it when
Create	Prompt + optional image, video or audio references	You're generating a new shot
Edit	A primary video + prompt (+ optional references)	You want to change existing footage
Extend	A primary video + a description of what happens next	The shot needs to keep going
First & Last Frames	Exactly two images	Both the opening and closing compositions are fixed

The video controls are: duration from 4 to 30 seconds, 480p or 720p output, aspect ratio, an audio toggle, output format, and an optional last-frame return.

Two things worth noting from the docs:

Extension is prompt-guided. Veemo states plainly that it "does not guarantee frame-exact seamless continuation", so plan your edit with that in mind.
720p is the ceiling here. If your delivery needs higher resolution, that's a separate step (or a different model).

Step 0: Set up

You'll need a Veemo account with credits. Paid plans start at $29.90/month and the visible tiers go up to $199.90/month; subscription credits refresh each billing cycle. If you only generate occasionally, the one-time credit packs (Starter, Creator, Professional, Advanced) are worth a look because pack credits don't expire. New users also get 16 free credits, which is enough to generate a video, so you can try the workflow before paying.

Each generation shows its credit cost on the Generate button before you submit, and the number changes with model, duration and resolution. Get into the habit of reading it before you click.

Step 1: Design keyframes with Nano Banana

Seedance's First & Last Frames operation needs two images. You could use photos, but generating them gives you full control over composition and style, and that's where Nano Banana fits.

Veemo positions Nano Banana as its fast image model, built for quick iteration and batch variations. Its prompting guidance is short and useful:

Keep prompts concise.
Focus on mood, color palette and subject instead of very precise spatial instructions.
Use style keywords such as watercolor, cel-shaded or concept art.

Here's a pair of keyframe prompts for a short product-style shot:

FIRST FRAME
A matte-black ceramic coffee mug on a pale oak table, early morning,
soft window light from the left, warm neutral palette, minimal,
product photography, empty space on the right

LAST FRAME
The same matte-black ceramic mug on the same pale oak table, steam rising,
morning sun now brighter, a single croissant beside it, warm neutral palette,
minimal, product photography

Tips that save credits later:

Lock the aspect ratio first. Generate keyframes in the same ratio you'll use for the video.
Change as little as possible between frames. The more the two frames differ, the more the model has to invent in between.
Pick from variations. Nano Banana is cheap per image, so generate several of each frame and choose the closest pair.

If you need sharper detail or consistent characters across frames, Veemo also offers Nano Banana Pro, which its page describes as supporting 1K, 2K and 4K output, character consistency for up to five people, and up to 14 reference images.

Step 2: Write the shot like a spec

Seedance responds better to a structured brief than to a pile of adjectives. A template I like:

SUBJECT:   what must stay the same (shape, color, materials)
ACTION:    what changes, in order, with rough timing
CAMERA:    one camera move, described once
LIGHT:     how the light behaves over the shot
AUDIO:     ambience / effects / silence
EXCLUDE:   things that must not appear

Filled in for our mug:

The matte-black mug keeps its exact shape, handle and finish throughout.
From 0–4s steam begins to rise slowly; from 4–7s sunlight warms across the table.
Camera: a single slow push-in, no cuts.
Light: soft morning light gradually brightening from the left.
Audio: quiet kitchen room tone, a faint ceramic clink at the start.
Exclude: text, logos, hands, extra cups.

Step 3: Generate with Seedance 2.5

Open the Seedance 2.5 video generator and:

Choose First & Last Frames.
Upload the two Nano Banana images.
Paste the brief from Step 2.
Set a duration. Start short (around 8 seconds) while you're testing.
Choose an aspect ratio matching your keyframes, and toggle audio on only if the shot needs it.
Check the credit quote, then generate.

A sensible testing order is: get the motion right first, then worry about audio and resolution. Turning audio off for the first pass removes one variable when you're diagnosing what went wrong.

If you don't care about a fixed ending, use Create instead and pass the first keyframe as an image reference. That gives the model more freedom with the ending.

Step 4: Extend or edit

Two follow-up operations make Seedance useful beyond single shots:

Extend: upload the clip you just made and describe what happens next ("the camera continues to push in until the mug fills the frame"). Since continuation isn't guaranteed to be frame-exact, give yourself a natural cut point or a hold at the end of the first clip.
Edit: upload a clip and describe the change ("replace the croissant with a small bowl of berries"). Keep the request narrow; one change per pass is easier to review.

Step 5: Stitch it locally

Once you've downloaded your shots, ffmpeg can join clips that share codec, resolution and frame rate without re-encoding:

# clips.txt
file 'shot-01.mp4'
file 'shot-02.mp4'
file 'shot-03.mp4'

ffmpeg -f concat -safe 0 -i clips.txt -c copy final.mp4

If the clips differ (say, one is 480p and another 720p), drop -c copy and let ffmpeg re-encode.

Gotchas

Text and logos inside generated video often need fixing in post. Plan to overlay them in your editor rather than prompting for them.
Keep a log. Save the prompt, settings, credit quote and output filename for every attempt. That turns "it worked once" into something you can repeat.
Failed generations: Veemo's refund policy says credits are normally returned when a task fails because of a system or provider error. That's a credit return, not a cash refund.

Wrapping up

The pattern is simple: stills for control, video for motion, an editor for the final cut. Nano Banana makes keyframes cheap enough to iterate on, and Seedance 2.5's First & Last Frames, Extend and Edit operations let you direct a shot instead of hoping for one.

If you try this workflow, I'd love to see what you make. Drop your prompt template tweaks in the comments.

DEV Community: Monica