[I'll show you all the failures too] Making drawings move is so much fun and so difficult

Have you ever dreamed, 'I wish I could make the pictures I draw move'?

In my previous note, I wrote about how 'I became an AI creator in one week', and the part that made me the most nervous with anticipation and mystery was the video for the main visual. Today, I'll show you everything behind the scenes. Including the failures.

1. Isn't it amazing that the pictures you draw can move?

Honestly, this was my first time editing video.

I've been doing web design for 15 years, but video editing is a different genre. I've kind of avoided it, thinking, 'Premiere Pro? After Effects? That looks hard...'

That changed when I met Claude.
Video editing became something I could do just by giving instructions.

The moment the first image started to breathe thanks to AI, I couldn't help but laugh and think, 'This is amazing.'

I'm glad there was no one around! I must have had a creepy grin on my face (lol).

2. I used Seedance 2.0

The tool I used was Seedance 2.0. It's an AI that creates videos from images, released by ByteDance in China (the parent company of TikTok).

Roughly speaking, it's like this.

• Upload one still image

• Write a prompt (= instructions for the AI) describing how you want it to move

• Wait for a few tens of seconds to a few minutes

• The video appears

That's it. No skills or knowledge required.

There are quite a few types of video generation AI out there now.
Sora, Runway, Kling, DomoAI... some of you might have heard the names.
I imagine each one has slightly different areas of expertise.

Seedance 2.0 was chosen for a simple reason.

Because it was the talk of the town! (〃∇〃) Yay

3. ⚠️ Beware of fake sites

"Don't let yourself be fooled!"

If you search for "Seedance 2.0" on Google, a lot of fakes will come up.

At the top of the search results, there are sites with plausible-looking domains and designs. You might think for a second, "Is this the official one?"

So, I pasted the URL into Kuro-chan (my nickname for Claude Code) and said,
"I want to make an animation using Seedance 2.0."
And the answer that came back was,

"This isn't it, 🤖" Σ(´∀｀；) Huh?

I was about to accidentally believe it and create an account... Thanks, Kuro-chan!

Don't just open it from a search, but instead:

• Follow links from official announcement articles (like ByteDance releases)

• FAL.ai Search for "Seedance" on the site and enter from there

• Have an AI (Claude Code, ChatGPT, etc.) verify the URL

It's safe to open it using one of these methods.

If you want to use it via API, there is also a way to access it via FAL.ai .
FAL.ai is a platform where you can use various AI models all in one place, and you can safely use Seedance 2.0 from there as well.

With AI tools, the newer they are, the more intense the competition with fake sites. Please be careful, as things like accidentally registering your credit card are really no laughing matter.

4. How to actually make it

From here on, I will introduce the flow of how I actually made the main visual video.
This time, I made a video from a single line drawing.

Step 1: Prepare the line drawing

First, prepare the line drawing you want to animate.

It doesn't have to be a line drawing, but
in my case, I'm not good at coloring, so I use line drawings. It's something I drew for practice that didn't have a place to be used and was shelved.

You can draw something new, or pull out a rough sketch you have on hand. I think the most important thing here is to choose a picture that you have the desire to "make move!"

Step 2: Coloring line art with ChatGPT

What to provide to ChatGPT:
・Specific instructions on the scene you want
・Line art
・Reference images for clothing and positioning

・志摩がスパイ活動中のイメージにしたい
・位置はsample03のようにクレーンの上のような不安定な場所
・服装のイメージはsample01の黒っぽい服装
・ネオンのイメージはsample02
・サイズは横長で2000px×1500pxはほしいです
・志摩の髪色のベースはピンクですが、周りのカラーに合わせて暗めのピンクがかった茶色でも構いません

元データの線画もあわせて共有します。私が描いた線画なので、こちらをベースにイラストを加工してください。

Step 3: Create a storyboard

Since a video doesn't work with just "one scene," you need to think about multiple scenes.

So, first I put together a storyboard (which scenes to show and in what order).
It's like writing out a flow such as "wind blows -> hood comes off" and creating an image for each scene.
This time, I set each scene to 5 seconds.

Step 4: Create scene illustrations

Once the storyboard is decided, I send instructions for each scene like "in this atmosphere, with this composition" to generate the images for each scene.

...This was more difficult than I thought. (;´･ω･) Hmm...

The "character looks like a different person" problem

For some reason, every time it makes them aesthetic/bishonen-style.

• Big double eyelids

• Strong emphasis on eyelashes

• Red lipstick

• Sharp chin and nose bridge

• Slanted eyes

"No, that's not it...!"

No matter how many times I fixed the prompt, provided reference images, or restarted in a new session, it still leaned toward the aesthetic style.

The last resort. I opened Photoshop and made small corrections bit by bit.

⚠️ Note: Don't forget to specify the aspect ratio!

This was completely my mistake.

When generating images with ChatGPT, I forgot to specify the aspect ratio and tried to turn the resulting images directly into a video.

As a result, the ratio I wanted for the site didn't match the video ratio, leading to cropping issues or awkward margins. I learned right here that I need to decide on the image ratio before passing it to Seedance.

Step 5: Create an account on FAL.ai (a bit of trouble here)

Once all the scenes are ready, I'll create an account on FAL.ai and start preparing to use Seedance 2.0.

There was a small accident here.

FAL.ai is a pay-as-you-go service where you charge credits upfront. I tried to charge $20 (about 3,300 yen) using my Rakuten card.

...However.

I received a notification from Rakuten Card saying 'Unauthorized access detected,' and my card was suspended.

'Wait, what??' Σ(´∀｀；)

It is an overseas service, after all. It seems Rakuten judged it as 'Is this being used fraudulently?'

I hurriedly called Rakuten.

After explaining to the operator that it wasn't fraud, my card was reactivated about 10 minutes later. The charge was completed successfully.

This is probably a common Rakuten thing. I usually use a different credit card, but since it was unavailable due to an access issue, I used Rakuten and this happened.
Just sharing that this kind of thing can happen when using overseas services.

Step 6: Create videos for each scene on FAL.ai

From here on, I'll leave the work to Claude Code.

• 'I want to animate this image using Seedance 2.0's image-to-video model'

• 'I want it to be a scene like this'

The flow is to send these instructions and have it generated via API.

So, there's no manual work like "clicking around the FAL.ai dashboard...".
Videos are created just by talking to Claude Code....That's the feeling.

The incredibly convenient part of agent-based AI.
It's become an era where if you can put what you "want to do" into words, it takes shape.

Step 7: Generate and download (⚠️ This part is super important)

The video you created must absolutely be downloaded.

I will write more than enough about why in the next section.

Cost estimation (reference)

By the way, the production cost for this project was roughly like this.

• Resolution: 720p (This is clean enough for a web hero section)

• Length: 5 seconds × 5 cuts (successful)

• Failures: 3 cuts

• Total: approx. $12 (a little over 2,000 yen)

If you go to 1080p, the unit price more than doubles (around $0.65/second), so it jumps up significantly. It's important to use 720p for scenes where it's sufficient.

The current reality is that it still costs money to create videos freely.

Step 8: Edit each scene into a single video

Claude connects the video nicely for you.
Since it also easily creates several patterns, I just convey "I want to do this" and judge whether it's good or bad.

MV version (no music, partial repeat, etc.)
SNS version (with music, short)

5. The reality of failure and success

It is true that it's easy. But of course, there are failures too.

Failure 1: There is no wind

The first thing I generated was just a video where "the hood comes off".
There was no expression of wind, and the hood just slowly fell down.

“No, that’s unnatural!” ╰( ^o^)╮彡))Д´）SLAP

What I wanted to see was a scene with a bit more drama, where the wind blows strongly and the hood comes off as a result.” The reason was that I hadn't included the word “wind” in the prompt.

Failure 2: The wind is too strong problem

So, I thought I'd add wind. I wrote “strong wind” in the prompt and tried again.

As a result, I got a video where the hood flapped around like a flag.

For some reason, the size of the hood kept getting bigger halfway through, and eventually, it had the surface area of a cape.

Where did the laws of physics go? (lol) This one was also a bust.

It's really hard to convey the strength of the wind with words. If you write “strong wind,” it blows too hard, and finding a prompt with just the right temperature took a surprisingly long time.

Failure 3: I wanted to show you the failed videos...

Here is a heartfelt warning from me to all of you.

Videos generated on FAL.ai disappear from the server in 7 days.

Yes, they disappear. Completely gone.

I was planning to introduce the failed videos on note, but a week passed while I forgot to download them... and when I realized it, everything was nothing.

Truly nothing.

So, the “failed video of the hood slowly falling” and the “failed video of the hood becoming a flag” that I wanted to paste into this article can only be delivered in text.
Ugh, what a shame! (ノд｀＠) Ouch

Don't forget to download everything generated by AI on the spot.

“I'll come back for it later” is usually too late. I think the storage period for AI tools varies by service, but it's much shorter than you imagine. If you make something good, save it the moment it's done.

This is probably a lesson I will never forget from now on.

Precisely because it doesn't go well on the first try, the joy of the moment it works is huge.

Trial and error with AI is tough, but I think it's great to be able to share the joy of creation with a partner.

6. Conclusion

It's amazing when art moves.
But it's much harder than I thought it would be.

This is the most honest impression I have after making my first video.

If anyone thinks, "Can't you just do it in an instant using AI?", I would answer that it's half true and half false.
A video does come out in an instant.
But it really takes time and money to reach the "scene you wanted to see."

On the other hand, the time spent creating is much more fun and fulfilling than I expected.

The excitement of seeing your own art move is something you probably won't understand until you try it.
If you're curious, why not try it with your own art?

In the next article, I plan to write about music production. I'll write about how "Liar's Smile" was born, including the story of Suno and the lyrics, along with my failures.

Thank you so much for reading this far.