Apertus
Fully open: weights, training data and recipes. Over 1,000 languages, including Swiss German and Romansh.
Hugging Face ↗What’s
happening here
Work in progressThe AI Lab is a workbench, not a showroom. We test where artificial intelligence really makes filmmaking better, and we also show where it doesn’t.
Projects come together with changing partners from film, technology and design. Some become finished work, others stay experiments. Both belong here.
Principle: craft leads, the machine assists. Every AI-assisted work is labelled as such.
01 · Research
Understand
Testing new models and methods before they reach a production.
02 · Prototypes
Build
Short films, previz and visual tests that show what is technically and artistically possible.
03 · Tools
Share
Workflows for restoration, upscaling and pre-production that help other teams too.
Screening Room
First tests · AI-generatedFirst scenes, fully generated. Three tones we are exploring: Swiss absurdist comedy, Italian-Swiss horror and analogue sci-fi with an 80s look.
Raw material, no edit, no grade. We want to see how far visual language, light and timing already go today, and where craft has to take over again.
Open. European. Understandable.
PhilosophyWe don’t just want to use European AI, we want to help build it. And we want to make it understandable: what a model can do, where its data comes from and where its limits are.
Models like Apertus and Teuken show there is an alternative to closed black boxes: open weights, documented training data, many languages instead of English only. That attitude fits film. Anyone restoring an archive wants to know what happens to every frame.
Our contribution: we test these models on real tasks from film, archives and communication, give feedback to the developer teams and document the results so that non-technical people understand them too.
Teuken
Trained for all 24 official EU languages, open source under Apache 2.0.
Hugging Face ↗EuroLLM
Open models for all EU languages, funded by EU programmes.
eurollm.io ↗Open
We prefer models whose weights and data are public.
Multilingual
Switzerland has four national languages, this site five. AI has to handle that too.
Traceable
Every step is documented and labelled.
Sovereign
Data from archives and productions stays where it belongs.
Experiments
Log · updated continuously- EXP/001In progress
Latent Reel
The animation above: a film strip made of particles that turns into a latent space and a lens. Written directly in WebGL.
- EXP/002In progress
Seeing the archive anew
Restoring and upscaling VHS and Hi8 footage with Topaz AI, side by side with classic restoration.
- EXP/003Idea
Previz in a day
From treatment to moving storyboard: concept with Claude, text with Apertus, images with generative models.
- EXP/004Coming soon
First portfolio project
The first finished project from the AI Studio. It will appear here once it is approved.
- EXP/005In progress
Neon Noir
Visual study in retro-futurism: laser grids, neon, chrome and VHS glitches. Inspired by the synth cinema of Seth Ickerman × Carpenter Brut (Turbo Killer, Blood Machines). The neon elements on this page are the first test.
- EXP/006In progress
Dezibel
Swiss absurdist comedy: a man, a sound level meter and a neighbourhood that is too loud. Three generated scenes in the Screening Room.
- EXP/007In progress
Galleria 1984
Italian-Swiss horror in the style of 1984, borrowing from body horror: a freight train, a tunnel, a man in the last carriage. Two look tests.
- EXP/008In progress
Channel 0
Analogue sci-fi horror, 1984: a technician among test equipment and CRT monitors picks up a signal that shouldn’t exist. A story very close to our restoration work with tubes and tape machines.
Stack
- ClaudeConcept · Research · Code
- ApertusOpen language model from Switzerland
- TeukenOpen EU language model, 24 languages
- Topaz AIUpscaling · Restoration
- WebGLReal-time graphics in the browser
Crew
Flavio SpeglitzProducer · Lab lead
PartnerName to follow
PartnerName to follow
You?Join in: write to me