Seedance2 landed as a capable AI video model with joint audio+video output in HD and with up to 15 seconds output. One caveat, it’s pretty costly to run. Every generation takes credits and time.

More than ever it’s important to make the right decisions before you hit generate to increase your chance of getting the right result. So I built some tools that I use in my own practice and I’ve decided to share them for free.

The prompting folklore vs the official guide

I find that every thread about Seedance2 tells you something different about how to prompt it. Conflicting advice doesn’t just confuse — you risk wasting money and hours testing it.

There is an official prompting guide you can sit down and read. But if you are like me there’s a real difference between reading a manual and remembering to apply the rules every session. That’s what these tools aim to solve.

Decisions before you hit generate

Some of how Seedance behaves is documented, and some of it is counterintuitive. One example: if the camera moves and the subject moves and everything moves at once, you get soup — the model can’t cleanly interpolate competing motion. So you pick one dominant motion. That’s a decision you make on paper, for free, instead of paying a generation to find out.

A diagnosis and solution when a generation fails

A good prompter can coax a decent prompt out of ChatGPT as is. But ChatGPT doesn’t always keep its eye on the ball. It will give you general video AI advice that might not apply to Seedance2 specifically.

The tools I share - one for ChatGPT and one for Claude - are built from the official documentation for Seedance2. Not only do they help me craft my prompts, they remind me of the core principles applied.

The tools do more than write prompts though. If a video fails, I can ask the tool why, and how to fix it.

Planning is everything

The tools help you build the prompt, but a prompt is only as good as the shot you planned. If you haven’t resolved how things move before you build the prompt, no tool will save you — and you’ll waste generations.

So “fewer wasted generations” has two halves: plan the shot, then prompt it. The prompting tool is the second half, the planning pass is the first. I wrote a seperate guide called The Planning Pass (available for free through Gumroad) that is applicable no matter what video model you use.

One tool for ChatGPT, one for Claude

The prompting tool comes in two flavours depending on whether you prefer ChatGPT or Claude. The custom Seedance GPT is here: Seedance2 customGPT

The Claude version comes in the shape of a ‘skill’ which is a set of files you install in Claude. Instructions + download is available for free through Gumroad: Seedance2 skill for Claude

For a long time my work has been making images move — light across a stage, projections behind a band, something drifting on a screen. The tools change every few years; the real question never does: how do you make a moving picture feel like it carries something? So when AI video arrived I wasn’t looking for a shortcut — I wanted to know what new expressions it could enable.

A quick introduction

I’m Vibeke Bertelsen. I live in Copenhagen with my family. By day I work at Vertigo, an artist studio making light, projection and visual worlds for concerts, theatre and as art installations. We have worked with the biggest cultural institutions of Denmark such as The National Gallery of Denmark and the Royal Theatre. Alongside that I make my own work — experiments with digital tools and AI — and right now I’m deep in an AI short film.

I’ve done motion graphics for years, long before any of this existed. (My two favourite subjects at school were art class and computer science, which explains most of how I work.) I don’t use AI on its own — I fold it into the tools I already know, and the good stuff usually comes out of that friction.

Why I care about this

A lot of AI video feels cheap. Not because of the technology — because the maker handed over the decisions that matter.

AI itself isn’t creative. Left alone it drifts toward the average, the already-seen — the strange idea is never going to come from the machine. It has to come from yourself.

That’s the conversation I want here — honestly, including the parts that don’t work.

What this is — and isn’t

If you follow my Instagram (@_udart_), that’s my sketchbook — the work, the experiments, the strange fragments. This is where I share my thinking.

Expect occasional studio letters: process, methods, what I’m figuring out. Reflective rather than hyped — more notes than lessons. It’s for fellow makers who want to use these tools without their work starting to look like everyone else’s.

Say hello

Some things don’t change. Long before I shared any art online, I was sharing small tools and tips. I’ve always treated both the same way — a way of connecting with people. What I love most is when it works: when something I make finds another likeminded soul.

So please get in touch. A question, a disagreement, something you’re making — all welcome. If you’re reading this in your inbox, just hit reply. Or come find me on Instagram, @_udart_.

Glad you’re here. Let’s see what it can do.

— Vibeke

Why I can’t stop looking at botanical plates

I’ve always loved antique botanical plates for their calm authority. They look neutral, but they carry a whole worldview: classification, control, possession through naming.

In this series, I wanted to see what would happen if that visual language stayed intact — but quietly began to speak about modern violence instead.

That’s how this series, The Bitter Fruit of Mankind, started.

The Strategy: Hiding modern violence in plain sight

The project is built on a tension that I can’t stop thinking about. Botanical art is usually about healing, food, and scientific knowledge. But modern power usually shows up as machines, weapons, and systems of force. When you put those two worlds together, the vibe shifts. The image stays calm, but it starts telling a very different story.

I call the approach “visual infiltration.”

It’s not meant to be a parody or just “vandalizing” old art. It’s a quiet intervention—hiding a new meaning inside the original authority. If you just glance at it, it looks like a standard botanical plate. It’s only when you linger that the image starts to betray itself.

In one piece, there’s an aloe plant with a fighter plane woven into the leaves. It’s tucked into the structure so when seen in small size it just looks like botanical detail until your brain finally catches up. In another, I turned an artichoke into something resembling a hand grenade. It’s the geometry of a weapon disguised as plant anatomy.

It’s basically collage logic, but I wanted it to be seamless—like the image was just retranslated into a darker language.

Don’t break the spell: The rules of the project

I started by hunting down illustrations with a really strong “voice”: engraved lines, disciplined shading, and that neutral, scientific look. Then, I set one hard rule for myself:

Don’t break the illusion.

The tweak should feel like it was always there. I wasn’t just “adding an object” on top; I was selectively transforming the image—rewriting one layer while keeping everything else steady.

Here’s how I actually did it:

• Start with structure. I looked for clear stems, leaves, and “negative space” where a new form could hide.

• Design for the second look. I wanted the realization to hit late—a small shock after you think you’ve already seen the whole image.

• Keep the “safe” zones. I intentionally left parts of the original drawing untouched so the “archive” felt real and present.

• Match the texture. The density of the lines, the shading, and the rhythm of the engraving had to stay consistent. If the texture is off, the spell breaks.

When the edit becomes a mutation

The best moments were when the transformation didn’t just feel like an “insert.” The plane didn’t look like it was behind the aloe; it felt like it had grown out of it. That’s when the series moved away from “hidden objects” and toward something more uncomfortable: the realization of how easily the language of nature can be twisted to carry human violence.

The history under the surface: Empire and extraction

Botanical illustrations aren’t actually neutral. They’re tied to the history of exploration, empires, and extraction. They were tools used to name and classify the world so it could be mapped, moved, grown, and owned. These images were how we made nature “legible.”

That history honestly shares a lot of DNA with military imagery. Both are about control—one through naming things, the other through force. When a fighter plane hides inside a medicinal plant, it’s not just a visual trick; it’s two different kinds of power colliding.

The effect only works if the intervention stays subtle enough to be discovered late.

Why Nano Banana was the right tool

I didn’t start with the AI; I started with the concept. I used Nano Banana because it’s great for “selective transformation.” I could rewrite specific parts of an image while keeping the style and the “untouched” areas exactly as they were.

For this series, “transformative remixing” meant:

1. Dropping in a new object (like a grenade or a plane).

2. Keeping the original visual rules (the engraving texture and shading).

3. Preserving the stable areas so the history of the archive didn’t disappear.

If I changed everything, it would just be an “AI-generated image.” By keeping parts of it fixed, it becomes a hybrid: an archive with a glitch inside it.

How to try “quiet intervention” yourself

If you want to play around with this kind of work, here’s what I learned:

• Pick sources with strong “rules.” Look for consistent linework and disciplined shading.

• Decide what to leave alone. Figure out what stays untouched before you start editing.

• Write for gesture and shape. Don’t just ask for a “vibe”—describe how the object sits or what it’s doing within the plant.

• Match the texture. Ensure the edit doesn’t “announce” itself by being too sharp or too blurry.

• Aim for the “second look.” That’s where all the tension is.

The full series

I’ve posted the full Bitter Fruit of Mankind series on Instagram. Part I and Part II

The Spark

I didn’t realize how much MidJourney was doing for me until I tried to recreate the same “effortless aesthetic” somewhere else.

In MidJourney, style is almost a default setting. You drop in a prompt, maybe add a --sref, push --stylize, and the image arrives already wearing a cohesive look. MidJourney is opinionated about aesthetics — and that’s its superpower.

Then I started using Nano Banana Pro more seriously, and the outputs felt like raw footage: not bad, just neutral. The upside is real control over composition and subject matter. The downside is you have to build the aesthetic scaffolding yourself.

That’s when I started thinking of the shift as moving from stylize to structure.

The Observation

The difference, in practice:

• MidJourney is style-forward. It prioritizes a strong look, even if it improvises details.

• Nano Banana is composition-forward. It’s more literal and more controllable — if you tell it clearly what must be true.

So the question becomes: how do you keep the aesthetic coherence you’re used to, while gaining the control Nano Banana offers?

A quick example (where MidJourney drifts, and Nano Banana holds)

Here’s the MidJourney prompt:

MidJourney prompt:
A young man is holding a horse by the harness, dynamic framing, nature details.

The gesture matters here — “holding by the harness” is the core instruction — but MidJourney often prioritizes vibe over literal pose. You get something beautiful, but not necessarily the hands and contact you asked for.

When I translated the same idea into Nano Banana language, it held the pose far more reliably. This is the structured prompt I used (converted):

Converted Nano Banana prompt:
Use images 1–2 as mood and atmosphere reference. A young man holds a horse by the harness with one hand, leaning his head into the horse’s neck while wrapping the other arm around it in a quiet embrace, captured in a single intimate moment. They are positioned amidst a dense field of glowing, fiery orange flowers and intricate botanical silhouettes.

What changed isn’t “more words” — it’s explicit choreography (hands, contact, lean) plus clear scene anchors, with the reference images carrying most of the aesthetic weight.

The Experiment

Once I started testing systematically, three patterns showed up fast:

1. MidJourney shorthand doesn’t translate well.
   MJ tolerates compressed prompts because it supplies taste automatically. Nano Banana responds better to clear, structured natural language.

2. Style is a bundle, not a knob.
   In MidJourney, --stylize can feel like “turn up the art.” In Nano Banana you build the look from smaller parts: references, lighting behavior, texture discipline, lens logic, and what you explicitly avoid.

3. References matter more than adjectives.
   Without references, Nano Banana can look correct-but-generic. With references, it locks into a visual world.

The bridge: translation (why I made a converter)

This is why I stopped thinking of it as “prompting differently” and started thinking of it as translation.

MidJourney prompts are like compressed code. Nano Banana wants a director’s brief.

Crossing that gap manually every time is slow, especially if you’re migrating a library of MidJourney prompts or iterating quickly. So I made a MidJourney → Nano Banana converter: a reusable template that expands MJ shorthand into Nano Banana-friendly structure.

What it’s good for:

• turning short MJ prompts into clear Nano Banana prompts that specify pose + scene logic

• preserving the intent of your original prompt while making it “readable” to Nano Banana

• giving you a consistent starting point, so your time goes into iteration instead of rewriting

Next level: structure as a reusable workflow

Once you have a translation template, the next obvious question is:

Why do this manually every time?

If you already work with node-based editors (Imagine Art, Krea, Runway, etc.), you can turn the translation step into a reusable workflow.

The simplest version is:

1. A text / LLM node that converts a MidJourney prompt into a structured Nano Banana prompt (using the converter template as a system prompt)

2. A Nano Banana node that generates from that structured prompt + your reference image(s)

That’s it.

And the value isn’t speed for the sake of speed. The value is consistency: you can test variations rapidly without constantly rewriting your prompt language from scratch.

The Takeaway

If Nano Banana feels “flat” coming from MidJourney, try this:

• Treat it as translation, not a model weakness.

• Write pose + staging like choreography (hands, gaze, contact, weight).

• Use style references instead of stacking aesthetic adjectives.

• Define the image system (lighting, texture behavior, lens logic) rather than “vibe words.”

• Add avoidances to prevent the look from drifting generic.

Resource

If you want the converter template I use (the exact text structure that turns MidJourney shorthand into Nano Banana-friendly prompts), you can grab it here:

MidJourney → Nano Banana Prompt Converter (free)