Up next in 10
My first AI micro drama generation failed badly — the characters looked wrong, the script was too complicated, and the model spent most of the video doing something I never wanted.
So I’m fixing it and generating a better version with Kling 3.0 Omni.
👉 Try the AI Photo Generator: https://yb.digital/photo
I’ll show exactly what went wrong, how I improved the character references and simplified the prompt, and why shorter, clearer instructions can make a huge difference in AI video generation.
I’ll also test Kling 3.0 Omni with enhanced multimodal references and the available multi-shot workflow to see whether the new generation can finally deliver the scene as intended.
The goal: a short, clear AI drama scene with a romantic apology, no unwanted kissing, and a cliffhanger introducing the mysterious new character.
If you generate AI videos, this is an important lesson: a bad result doesn't always mean the model is bad — your references, prompt, duration and shot structure can completely change the result.
#AIArt #ImageGeneration #PromptEngineering #photogenerator #Kling3 #Kling30Omni #AIVideo #AIMicroDrama #AIStory #AIFilm #AIVideoGenerator
Show More Show Less View Video Transcript
0:00
first episode of my microdrama series
0:03
with C-Dance
0:05
No, with Kling 3.0 Omni, and I really
0:09
don't like it. First of all, the
0:10
character the main character doesn't
0:12
look like me. Second, I specified that
0:15
the characters
0:17
should be about to kiss when the
0:19
cliffhanger happens, but here they spend
0:21
most of the video actually kissing,
0:23
which is definitely wrong. So, here's
0:26
what to do when the video generated is
0:28
not correct. First of all, you should
0:30
identify what's wrong. So, in my case,
0:33
first of all, the character doesn't look
0:35
good because I changed the pictures, and
0:38
the main picture now is not good enough.
0:40
So, I will
0:41
I will change this picture. Then, the
0:43
prompt I use is too long. It has too
0:46
many details. That's too much for a
0:48
15-second video. 15-second video prompt
0:51
should be one clear sentence what to do,
0:54
and
0:55
the AI should elaborate on one action,
0:58
not have many actions happening at the
0:59
same time. If many actions are happening
1:01
at the same time, then it's better to
1:03
use the uh multi-shot option.
1:06
And for each shot, give details about
1:08
what's happening. Otherwise,
1:11
you reach the limit of the creativity of
1:13
the model because it's different scenes.
1:16
So, first of all, fix the character. So,
1:19
I have to go back to Michelle, my
1:22
character, and I will replace again the
1:25
main picture because I put two pictures
1:29
with a Michelle holding roses, but the
1:31
face is not clear enough, and uh
1:35
the the model cannot really make up
1:39
Michelle's face clearly. So, I will
1:41
change Michelle's face to a good picture
1:43
with a closer uh
1:46
closer shot of his clear face. So, here
1:49
my character Michelle.
1:51
Now, I have that same picture here. The
1:53
full picture, I will remove it.
1:55
So, here I have a second uh good-looking
1:58
picture. I mean, with the character
2:00
clear. I will add a third um
2:03
a third picture of uh Michelle with the
2:06
face clearly visible. So, here
2:09
for example, Michelle
2:12
uh having a breakfast selfie. And a last
2:15
one, Michelle with a rose. So, it uh the
2:18
model can choose which one is better.
2:22
All right. Now, we can confirm this.
2:25
Now that I have changed the character, I
2:27
already asked the GPT to
2:30
uh create me a shorter
2:32
prompt for my first scene because the
2:34
previous prompt was too long. It was
2:36
four paragraphs. Uh we really just need
2:38
one clear sentence, what to do, where we
2:40
want to do to go, and what shouldn't be
2:42
included. After a few
2:46
a few uh iterations, I got a kind of one
2:49
uh
2:50
one-sentence prompt to create a jealousy
2:53
microdrama. Michelle arrives late, and
2:55
seeing Sally apologizes to Anna with
2:56
flowers or a creative romantic gesture.
2:59
Anna is still upset, but slowly softens
3:01
as Michelle tries to make it up to her.
3:03
I don't want them to kiss. I didn't want
3:05
them to kiss. I actually specified that
3:06
they should be about to kiss when the
3:08
cliffhanger happens. And the the model
3:11
generated a whole video about them
3:12
kissing, which is totally wrong, and I
3:14
have to regenerate. Just as as they seem
3:17
to reconcile, a mysterious woman
3:18
arrives. Only her shadow and part of her
3:21
face visible, and with Anna and Michelle
3:24
shocked.
3:25
All right. Now, this should be better.
3:28
So, I will go back to my video. I will
3:32
here uh do the regenerate. Oops. Maybe I
3:35
went too fast. So, first first, I go
3:37
back to video menu here.
3:39
Then, I select my video here.
3:41
I open the video, and here I have the
3:44
option to regenerate. I could also edit
3:47
the video.
3:49
Um so, actually, let's maybe try the
3:52
video edition. So, I selected my video
3:55
to edit and to add visual references.
3:58
So, here I will add
4:01
Michelle. Uh, so I will add again my
4:03
main picture with the good face.
4:06
Uh, and here also a better looking
4:09
picture of Michelle. That's enough.
4:12
And describe the changes.
4:16
Um,
4:18
maybe
4:19
uh,
4:20
maybe this is too much. So, let's just
4:22
use one picture. So,
4:24
use
4:26
at image one for
4:32
male character
4:35
in video.
4:37
Uh,
4:39
and now the two characters. So, here I
4:41
will reference this one, the character
4:47
from
4:48
image one
4:51
and
4:54
image two
4:56
should not be
4:58
kissing.
5:00
Uh, wow, I'm I'm
5:05
I'm afraid this might be maybe too many
5:08
details. Uh, this might be too much, but
5:11
anyway,
5:13
I will try.
5:15
Um,
5:17
should not be kissing.
5:24
At image one should
5:26
be
5:27
apologizing
5:30
deeply with Flowel. So, actually, I will
5:33
reuse the prompt I pasted.
5:36
Uh,
5:38
oh,
5:39
I will do that.
5:40
Yeah.
5:44
Create a jealous microdrama
5:47
at image one
5:49
Rosalind and apologizes to at
5:53
image two with flowers or creative
5:55
romantic gesture. At image two is still
5:57
upset.
6:00
So, let's try
6:02
for less credit than to create a to
6:04
generate a new video to save a video
6:07
that
6:08
is just wrong. They seem to reconcile.
6:11
Okay, and I need although to give the
6:14
picture of the third
6:17
character.
6:20
All right, I also need to add
6:24
the picture of the third character.
6:27
We're not sure which one is better, but
6:29
let's choose the
6:31
this one because it's our main picture.
6:33
And in the description
6:37
image three, a mysterious woman arrives
6:39
only her shoulder and part of her face
6:40
visible.
6:42
All right.
6:43
Should not be kissing. It should be a
6:47
deep apology
6:49
from
6:50
image one.
6:54
Uh
6:55
all right.
6:56
Let's try that way and see where it
6:59
goes. We will keep the audio on and
7:01
we'll edit the video.
7:03
Uh so, here also for edition we have
7:05
different models to edit videos. We have
7:07
Kling 01 and we have Seedance 2.0. So, I
7:10
never used
7:13
any to edit video. So, let's see how far
7:15
it can go. So, we have this video.
7:17
Uh I will meanwhile
7:19
let it start to
7:21
Oh, the video is too long. Okay, so it
7:23
doesn't even work. Let
7:25
Okay, um
7:27
so, we cannot even use this
7:30
uh
7:30
this video.
7:33
All right. Uh so, we have to go back to
7:36
video generation. So,
7:38
let's go back to select the video.
7:41
Here, click on regenerate.
7:44
Why regenerate? Because it automatically
7:46
selects the settings we have set before.
7:49
So, now
7:51
um
7:52
we change the characters
7:55
because I changed the
7:57
picture of the character. So, we have
7:59
Michelle, we have Anna, we have Nita.
8:03
Uh
8:04
I will keep trying with Kling 3.0 Omni,
8:07
but definitely SeaDance.
8:10
Uh especially a SeaDance 2.5 creates
8:13
much better videos.
8:16
SeaDance 2.0 is not bad as well,
8:20
uh but let's keep trying with a Kling
8:22
because it's supposed to be very good.
8:25
Uh so far, I'm not really satisfied with
8:27
the results. So, I did select the
8:30
character, but I didn't click confirm.
8:33
Let's add the characters. Now, our
8:36
prompt create a jealousy micro drama.
8:39
Michelle arrives
8:42
late and sincerely apologizes to Anna
8:48
at the date.
8:50
Uh it didn't say that it was supposed to
8:52
be a first date also.
8:55
Uh
8:56
all right.
8:58
So, now Anna is still upset.
9:03
Oh, and uh the first date
9:05
at the rooftop restaurant.
9:09
at the
9:11
sunset rooftop uh
9:14
location, uh bar for example.
9:17
Or
9:18
uh yeah, bar. Let's keep it simple and
9:20
sincerely apologizes.
9:23
Is still upset, but slowly softens as
9:27
Michelle tries to make it up to her.
9:29
Just as they seem to reconcile, the
9:31
third character, Nita, a mysterious
9:33
woman, arrives. Only her shadow and part
9:35
of her face ball and with
9:38
Anna
9:39
and
9:41
Michelle shot. All right, let's generate
9:44
this one. We want again standard
9:46
vertical 15 seconds duration.
9:48
Let's rate again with clean 3.0 Omni and
9:53
meanwhile while it is trying to
9:55
generate, let's have a look again
9:57
at
9:58
the first video I generated with Anna so
10:02
the second character arriving in the
10:04
micro drama where she is at the date
10:06
while Michelle is playing in the pool
10:09
and as we can see it the model did as we
10:12
will see the model did create several
10:14
shots in that video. That's what we want
10:16
to uh
10:18
to actually reach but this was created
10:20
with C-DANCE 2.5 as you can see here.
10:41
Hey, easy there.
10:47
Oh.
10:48
Oh, no. I'm late.
10:54
So as we can see in the previous episode
10:56
episode three, we had four shots well
10:59
two shots in each of the two locations
11:02
but because I generated a 30 seconds
11:05
video for that it was far too long
11:07
because the first uh
11:10
let's say the first
11:11
10 maybe the first 10 seconds nearly 10
11:14
seconds. So third of the video is just
11:17
Anna the second character waiting
11:19
nothing happens. That's too long. This
11:22
should have been one or two seconds.
11:24
Then
11:24
the second scene we have the main
11:27
character Michelle playing in the pool
11:29
instead of going to the date and this
11:32
also last too long. It last
11:35
Let's see. It last nearly 10 seconds as
11:37
well. So, again a third of the video.
11:40
We understood the action. 5 seconds
11:42
would have been enough. And then the
11:44
next scene Anna is still upset waiting
11:47
at the location. This is This is fine.
11:49
It's only a few seconds, but it could be
11:51
shorter. And the last scene Michelle
11:53
realizes he's late and he's leaving.
11:54
He's also too long. So, here for
11:56
example, I spent too much credit. There
11:57
was no point to make a 30 seconds
12:00
video. It's better to make shorter
12:02
videos to give a clear direction
12:04
than to have the AI fill it with
12:08
characters barely moving as we can see
12:10
in the beginning of this third episode.
12:13
So, here for 10 seconds nearly nothing
12:16
happens.
12:17
When people see these videos, they are
12:19
wondering what they are waiting for,
12:21
what's happening. It's good to create a
12:23
bit of tension, but this is too long.
12:27
Then we created the episode four, which
12:30
was totally wrong because Michelle has a
12:32
totally different face. I switched the
12:34
pictures actually and the main picture
12:35
was
12:36
not the same and not the the same
12:38
clarity on the face. So, the face is
12:40
totally different. And here the same.
12:42
Nothing happens most of the video.
12:43
>> now.
12:50
>> You also can clearly see it's AI
12:52
character.
12:58
The cliffhanger at the end is good, but
13:01
the beginning of the video is not so
13:02
good because we actually
13:04
in the description for that video we are
13:06
now regenerating, we put the focus
13:09
on the apologies. And instead the
13:11
apology here, if we analyze the video
13:13
generating now we are clean 309.
13:17
The apology is fine.
13:19
2 seconds apology, but it
13:22
Well, it's kind of good. It could It
13:23
should be longer because first we have
13:25
the 2 or 3 seconds apology and then
13:28
We have nearly 10 seconds of kiss which
13:31
should actually not have been there.
13:34
Five seconds of kiss
13:36
and two seconds the character
13:38
introduction.
13:40
Scene three seconds
13:42
>> [music]
13:42
>> the
13:43
cliffhanger happens the third character
13:45
that we introduced is
13:47
coming inside the
13:49
video. So maybe this is because the
13:51
prompt was too long. That's a
13:54
possibility. Uh
13:56
another possibility also because I I
13:58
switched the pictures I was hoping that
14:01
switching the pictures would help the
14:03
model to uh
14:06
to localize the the main character to
14:08
the current scene because previous
14:10
scenes were at the pool and
14:11
this well this micro drama episode
14:14
should not be at the pool.
14:16
But that's not what happened. Anyway,
14:19
meanwhile have a look at the credits. So
14:20
here I generate again a video for 72
14:22
credits. So at the end instead of using
14:27
seedance 2.5 which would have generated
14:31
the probably a good good video
14:34
for double the credits of
14:36
clean free online I had a bad first
14:41
video generated by clean free online and
14:42
I have to generate all over again. So
14:44
this is the case when you generate a
14:46
video and you regenerate it you have to
14:48
regenerate from scratch and spend again
14:51
the credits and there is no way to get
14:53
them back because this was not a system
14:55
error this was a user error.
14:58
Always make sure beforehand your prompt
15:01
is optimized because you don't want to
15:03
waste those credits. Now let's see
15:07
probably the new one is better. Let's
15:09
have a look and we'll see in the next
15:11
video where we go with this AI generated
15:14
micro drama. Create your own at
15:15
yb.digital/photo
15:17
link in the description
15:19
and enjoy your creations.
15:26
>> Well, and now my internet is not
15:28
loading, so I just wait for the video to
15:31
be
15:31
fully loaded before I let it play. It
15:35
has just been generated,
15:39
the first episode of that micro drama
15:43
series.
15:45
Just uh few seconds and we should be
15:47
able to launch it.
15:49
Of course, we can then later download
15:51
it, add it to
15:52
uh favorites, or delete it. Let's play
15:55
and enjoy your creations.
16:04
>> I'm so sorry I'm late.
16:06
I have something for you.
16:07
>> It's okay. I'm just glad you're here.
16:18
>> [snorts]
16:23
>> I'm so sorry I'm late.
16:25
I have something for you.
16:26
>> It's okay.
Science