Masterclass: How to use AI Agents for Filmmaking | Penumbra - Ep3
Try Agent Two by invideo: https://invideo.io/s/ypp
Watch Episode 1 here - https://youtu.be/P8YYXMUPfZ8
Watch Episode 2 here - https://youtu.be/uu1GCP9iD90
Watch Episode 3 here - https://youtu.be/4iKI-vjl_LY
*****************
Penumbra Episode 3
This is the breakdown of Penumbra Episode 3, where we get into what worked, what didn't, and what we learned. From cracking multi-character placement sequences to letting Agent Two surprise you with camera moves you never planned, plus an honest look at the continuity error that slipped through and why it stayed in.
****************
Chapters -
0:00 Intro
0:51 Placement Shots
2:27 Building with Seedance
3:54 Continuity Errors
4:55 Sync with Agents
6:05 Outro
****************
Want to learn more? Check Out: https://youtube.com/playlist?list=PLpEWYHZUE3y3ewSKfuj9uSYPmaeVp9euv&si=bk-fSH4PhVRil_Ck
Closed Caption:
Hello and welcome to the breakdown of
Episode 3 of Penumbra.
It's an invideo original series that
I'm working all by myself with the help
Of agent one, which is now agent two.
And in this episode, I've tried
Experimenting with a lot of visual
Treatment and also tried improvising on
The realism of the overall tonality and
Vibe of the series. If you have not
Watched the episode, I highly recommend
That you do along with episode one and
Two, which is on the same channel. And
Let me know in the comments whether you
Like it, you do not like it, if there is
Something specific that you want to
Learn about any of the episodes, and I
Would highly appreciate that. Let's move
On to the highlights, insights, and some
Of the very valuable experience that I
Gained while building and working on
This episode. So, the first thing and
Probably one of the most important
Things that I would like to share and
Talk about is placement shots.
Who's sitting where, who's addressing
Whom. For example, in the episode, there
Is an entire sequence where there are
Four people sitting inside an airplane
Cabin. There is also an object placed in
Between all four of them. So, there's an
Interaction going on. There is one
Person who's narrating, describing
Something, and he's not just addressing
It to one particular person, he's
Addressing it in the entire room. There
Is also the object, which is also not
Like a prop. It is object of importance,
Which will also be addressed in the
Conversation. There are many ways to
Deal with a placement sequence where
People are talking to each other, but I
Think the method that worked far better
For me, first of all, generate an image,
Wide-angle image, of all the people
Inside that frame, positioned in their
Respective places. Since this image is
Rendered and re-rendered, the quality,
If you notice, is going to be
Significantly different or outright bad
Compared to your other generations. Most
People add this image along with the
Prompt while talking to a video model,
But that is a mistake because the video
Model is most likely going to use it as
A benchmark or maybe even incorporate
That frame in the final output, which
Will break your consistency. The correct
Way to do this is to convert the same
Image into a very rough sketch along
With the name tags of whether it's
People, objects, the entrance, exit,
Directions, stuff like that. That will
Help the video model to understand the
Positions. You can alternatively also
Create like a top-down blueprint without
The characters and just the names. Now,
The next thing that I would like to
Discuss is building with SEEDANCE.
Whenever you're building a sequence with
SEEDANCE, I think the creative community
Is divided into two groups of people.
One who build individual shots and later
Stitch them together on Premiere or any
Other software. The other one who
Incorporate multiple cuts and shots
Within the same prompt and later adjust
The timings and stuff in post. I think
For 80% of the cases, I'm the other the
One. Why? Because I rely a lot on Agent
2's prompting skills and the video
Models to surprise me with different
Kinds of POVs and camera movements and
Stuff like that. I'm not a
Cinematographer, but if you notice and
If you have watched the episodes 1 to 3,
You'll notice that there are a lot of
Shots which look very intentional, as if
They are very well thought through and
Planned. I did not plan it. It is just
The Agent 2 and the video models filling
In the gap with their creativity. The
Only thing I did was to push the model
To think freely, allow the agent to take
A risk, simply just go bonkers. In my
Experience of building 40 to 45 minutes
Of Penumbra so far, I have noticed this
Mostly happens when you rely on multiple
Cuts, you express your intention and
Flow of the shot that you want to build
To the agent, and you leave the POVs,
The camera angles, the movements to the
Agent, and it's smart and creative
Enough to surprise you with very
Professional looking shots.
Moving on to the next thing, it's not
Actually an insight. It's a lesson that
I learned and which humbled me. If you
Watched the episode 3, I am not sure if
You have noticed, but there is a serious
And embarrassing continuity error. When
I generated the full body character
Sheet of Dr. Brand, the man wearing the
Coat, the collars were flipped down, but
When I asked the agent to generate the
Face character sheet, it flipped the
Collar up, which I did not notice until
Very last minute of the episode.
>> [snorts]
>> I was not able to fix it not because I
Could not, but I had to compromise a lot
On the emotional flow of the whole film
And I didn't want to do that. I
Prioritized that over this continuity
Error. As I was talking about relying on
The agents for a lot of things, it
Should never be 100% and it can make
Mistakes. It's not foolproof. It's not
Watertight. You should always pay double
Attention to the sheets, the character
Sheets, the turnaround sheets that you
Are generating and observe the details.
It can make mistake, but it's very rare.
Now, this is for people who function
Like me. I have my Premiere or any kind
Of timeline editor running in parallel
Along with the agent on the browser for
Media generations and all kinds of
Videos, etc. You have to understand this
Is just a reminder that how the agent is
Visualizing your story and the story
Flow could be extremely different from
How it's actually getting executed on
The timeline because it can't actually
See what's happening on the timeline.
So, it's very important to keep your
Agent updated about everything. It's not
Just your progress, it's also about the
Problems and hurdles that you're facing.
The agent might be following a
Particular context, particular kind of
Direction according to the screenplay
And so on and you might be doing a lot
Of improvisations and things like that.
So, it is possible it might lose that
Context. In order to help the agent
Navigate and improvise on the kinds
Of video generation and help you solve
The problems, it's very important that
You share as many screenshots as
Possible of the timeline, where you've
Reached, what you're planning to build,
Where are you facing any specific kind
Of issues, ask for advice. It will start
Adjusting according to your working
Style.
So, yeah, that's about it. I hope these
Were helpful. episode four is in
Progress. I have worked on one of the
Most complicated sequences for Penumbra.
In that, there are some learnings and
Experience that I can't wait to share
With you. So, stay tuned. I'll see you
In the next one. Take care. Cheers.
Video Length: 06:24
Uploaded By: Invideo for Filmmaking
Published: 8/17/2026
View Count: 1,266