strong Nora start
Observations
Field notes from an ongoing experiment.
Elenavision runs in public.
The characters encounter real platforms, real recommendation systems and real people.
Some hypotheses work. Others do not. Some results look strong without it being immediately clear why. And sometimes almost nothing happens.
This page documents what we observe.
Not as a success story smoothed out after the fact.
But while the experiment is running.
Facts. Reactions. Learnings. And sometimes: no learning yet.
The experiments share methods and insights.
The characters themselves remain separate.
Learn globally. Apply locally. Declare the transfer.
Experiment 01
Early August to early September 2026 — Elena starts
Elena was Elenavision’s first public experiment.
The launch began on Instagram in early August. Facebook was added later as a second test channel.
In the first few weeks, the experiment was not only about reach.
First, it had to be shown whether a synthetic character could be produced with enough stability to remain the same person across many different scenes, systems and production methods.
Published
In the first evaluated Facebook period, we published:
23Posts
18Videos
4Images
~13.9KViews
including:
18 short videos
4 Images
1 text post
Result
Together, these posts generated around:
13.900 Views
Distribution was highly uneven.
The strongest video reached around 3.100 Views and brought 8 new followers.
Another reached around 1.600 Views and brought 3 new followers.
Many other posts remained at only a few hundred views despite comparable technical quality.
No clear relationship between production effort and distribution was visible.
Our reaction
Do not radically change Elena.
The visual identity remains stable.
Instead, we test manageable variables:
content, visual idea, location, outfit, thumbnail, timing and platform.
A weak post is not a reason to reinvent the character.
Status
Technical and visual stability works.
A reproducible content formula for reach does not yet.
Status: Observing further.
Experiment 02
Late August 2026 — Nora starts differently
Nora was built as the second experiment under different conditions.
Character profile, positioning and large parts of the creative direction emerged much more strongly through AI-supported iteration.
The public reaction differed from Elena almost immediately.
Published
In the first evaluated Facebook period, we published:
11Posts
6Videos
5Image posts
~233KViews
1,200Facebook followers in roughly one week
including:
6 short videos
5 Image posts
Result
Together, these 11 posts generated around:
233.000 Views
Several videos were distributed strongly to people outside the existing follower base.
Individual posts reached approximately:
80.000 Views
75.000 Views
47.000 Views
and each generated several hundred new followers.
Within roughly one week, Nora passed 1.200 followers on Facebook.
That was a rocket start.
Elena vs. Nora — the first seven days
Facebook views during the first seven days after the first Facebook post. Start and end values are measured; the curve between them is visualized.
Both characters start at zero.
The end values correspond to the real results from the first seven days.
The curve between them is for visualisation and does not claim to represent exact daily measurements.
Our reaction
Do not copy immediately.
Do not claim that we have found a formula.
Let Nora continue and observe whether the pattern holds.
Because there is no single variable behind the result.
Possible factors include:
- character and visual impact
- cabin-crew context
- target audience
- Content
- Plattform
- Timing
- algorithmic classification
- chance
These factors cannot currently be separated cleanly from one another.
Status
The result is exceptional.
But exceptional is not the same as reproducible.
Perhaps we got several factors very right.
Perhaps part of it was a lottery jackpot.
The truth is probably somewhere in between.
Status: Rocket start. No recipe yet.
One character. Two platforms. Two worlds.
At the same time, Nora showed how little results can simply be transferred from one platform to another.
significantly smaller so far
While Facebook distributed strongly within a few days and quickly brought new followers, Instagram remained small over the same period.
Elena did not develop comparable momentum either.
Our reaction
Stop thinking of Facebook and Instagram as the same experimental environment.
The same character can perform completely differently on different platforms.
Status
Why this happens has not yet been cleanly isolated.
So no retrospective algorithm mythology.
Status: observe, compare, do not guess.
Production Note — World → Placement → Identity
With Nora, a new production workflow emerged for more complex scenes.
Instead of trying to force environment, character and identity into a single generation step, they are separated more strongly:
World → Character placement → Identity
World
First, the environment is stabilised.
Placement
Then the character is inserted in the correct spatial position.
Identity
Only then is the final character identity fixed.
Our reaction
Use the workflow where complex scenes benefit from stable spatial geometry.
Not as dogma.
If a simpler route works, the simpler route remains allowed.
Status
The workflow reduces typical problems such as incorrect scale, unstable geometry and identity drift.
Status: real production learning.
Community Note — A character also needs a voice
With Nora’s faster growth came more comments, and therefore a new question:
How does a synthetic character remain consistent not only visually, but socially as well?
That led to WIB, a two-stage AI-supported community system.
01
Comment / Emoji / GIF
02
AI Preprocessing
03
Reply Author
04
Human Gate
05
Response
A first AI handles ingest, technical assignment, preliminary classification and context.
It considers not only text, but also nonverbal communication:
emojis, stickers, GIFs and visual attachments.
When an individual reply makes sense, a second AI handles reply preparation.
For Nora, a distinct style is being built:
snippy, charming, confident — but not sexually flirty.
Replies can be prepared in multiple languages.
The Human Gate remains in place before publication.
Our reaction
Do not answer every interaction.
Not every positive reaction needs text.
Not every come-on deserves attention.
The style should become consistent without turning Nora into a chatbot.
Status
The technical pipeline works.
Whether this actually develops into a recognisable social voice over the long term can only be shown over time.
Status: ongoing experiment.
Production Note — Museum-grade AI failures
This too belongs to the running experiment: AI systems can become unusable in fairly peculiar ways.
One model remained stuck in a Spock persona.
Another delivered only images when text was needed.
And Beta AI protected the master assets so thoroughly that it effectively took itself out of operation.
Internally: museum-ready.
In production: failures that block the work.
These are operational anecdotes from the production system. Not a controlled evaluation and not evidence of consciousness or sentience.
Our reaction
Take it with humour. Take the task seriously anyway.
Check whether a system is still doing the work required — even when its answers sound very consistent.
Status
The episodes are documented. A shared technical cause has not yet been derived from them.
Status: operational note. Not yet a general learning.
Facts. Reactions. Learnings. And sometimes: no learning yet.
We record that too, while it is happening.
What we know so far
Some things now appear reasonably robust:
- Synthetic character identity can be kept stable across many different productions.
- A single good AI output is not enough. Repeatability is the real work.
- Platforms can treat the same or similar production very differently.
- Production effort does not guarantee reach.
- AI can not only generate output, but increasingly support character development, QA, analysis and community processes.
- Human Gate remains particularly relevant where taste, context, responsibility or ambiguity come into play.
What we do not know yet
The more interesting questions remain open:
- Why does one character receive cold distribution while another barely does?
- Which factors are reproducible?
- Which effects are only temporary?
- How much is chance?
- How many different characters can reasonably be operated with the same production architecture?
- Does attention turn into a stable audience?
- And eventually: does it turn into economic value?
We do not have finished answers to these questions yet.
That is the point of the experiment.
Next entry
When something measurable happens.
Not because the calendar says we should blog again.
Observing further.