sarah wilson

One thing we've been thinking about while working on XMK Seedance is how quickly an AI video workflow becomes complicated once references are involved.

A text prompt is simple. You write an instruction, generate, review, and try again.

But real projects rarely stay that simple.

Someone might start with a product image, add a visual reference for lighting, another reference for camera movement, and then use an existing clip to explain the pacing they want. At that point, the challenge isn't just "generate a better video." It's making all of those inputs understandable and manageable.

That has changed the way we're thinking about the Seedance workflow.

We're experimenting with a more reference-led approach around XMK Seedance（https://www.xmk.com/seedance）, where images, video, audio, and written instructions can work together instead of forcing the user to describe everything through increasingly complicated prompts.

The interesting product question for us is what happens before generation.

Which reference controls the subject?

Which one is only there for style?

What happens when two references suggest different directions?

And how much control should we expose before the interface itself becomes another creative obstacle?

We're still learning where that balance should be.

The biggest lesson so far is that supporting more inputs doesn't automatically create a better workflow. The product also has to help people understand what each input is doing.

For other builders working on generative products: how do you add more control without making the creation process feel heavier?

We changed the default video duration in one of our recent tests.

I expected it to change the output. I didn't expect it to change the prompts people wrote.

With the shorter default, prompts were usually about one thing happening:

person turns toward the camera

car drives past the building

camera moves slowly toward the product

When we tested a longer starting duration, we started seeing prompts with more going on. Someone would describe an action, then a camera change, then something else happening near the end.

I don't have enough data to call this a pattern yet. It was noticeable enough that we started looking at some of the other defaults.

We had treated duration as an output setting. I'm less sure about that now.

We also tried putting more controls on the first screen.

The idea was simple: if the workflow supports more inputs, people should be able to see them.

It looked fine when we built it.

Using it felt a little busy.

We removed some of the controls, put them back, then tried hiding a few until after the first generation. I'm still not sure which version I prefer.

The same issue showed up again when setting up a MiniMax H3 test. There were several kinds of reference input we could make available, but putting every option in front of someone immediately didn't seem especially helpful.

Right now the first step is closer to:

prompt
optional image
generate

Then the other controls become available when they're useful.

This has a downside.

If I already know that I want to work with a specific reference, I don't want the interface making me hunt for it. We've already had a couple of moments internally where the "simpler" version felt slower because we knew exactly which control we wanted.

So we may change it again.

One thing I'm watching now is what people do immediately after the first result.

Do they change the prompt?

Change duration?

Add another reference?

Generate again without touching anything?

That seems more useful than arguing about which version of the first screen looks cleaner.

We're leaving the lighter version in place for now. I want to see which controls people actually go looking for before we move them around again.

Over the last few weeks, we've spent less time building new features and more time observing how people actually use AI video tools.

One pattern kept showing up.

Most users weren't asking for dozens of editing options. They wanted to get from an idea to a usable video as quickly as possible.

That changed how we thought about the product.

Instead of asking, "What feature should we add next?" we started asking, "What's slowing someone down before they even export their first video?"

We've been simplifying the workflow, reducing unnecessary steps, and making it easier to start with either a text prompt or a reference image. The goal isn't to replace professional editing software—it's to help people create a strong first draft much faster.

We're also paying attention to where people get stuck. Prompt writing, choosing a visual style, and deciding when a video is "good enough" are often bigger challenges than rendering the video itself.

This has led us to spend more time improving the overall creation workflow instead of continuously adding new buttons and settings.

We're building Seedance 2.5, and this is the direction we're currently exploring:

• Faster idea-to-video workflow
• Simpler prompt-based creation
• Better support for visual experimentation
• Less time spent on repetitive editing

I'd love to hear from other builders.

If you're building an AI product, what's one feature you removed—or decided not to build—that actually made your product better?

Project: https://www.xmk.com/seedance/seedance-2-5