All open roles
Research Engineer, Generative Comics
LocationBerlin, Germany
LanguageGerman
TypeFull-time
ModeOn-site
Pay€60,000 – €80,000 & VSOPs
About the role
Flikk is an AI entertainment app: personalised videos, games, and comics, all in one place. One good generated image is a solved problem. A comic is forty of them that have to agree with each other: same face, same style, same world, panel after panel.
This is a research role with a product attached. You would run experiments, most of which will not work, and be judged on the few that do, specifically on whether they still hold up when they run a hundred thousand times instead of once.
What you'll do
- Own generation quality for comics: character consistency, style coherence, and whether a page actually reads.
- Run structured experiments (conditioning, identity and reference methods, LoRAs, adapters, fine-tuning) and write down what you learned, the failures included.
- Take on the parts the base models are worst at: legible text inside a panel, hands, composition, and reading order.
- Build the evaluation that does not exist yet: an honest measure of whether a page holds together, because per-image scores miss the entire problem.
- Move what works out of a notebook and into something the backend can run inside a latency and cost budget.
- Track the field closely (papers, weights, releases) and hold a defensible view on what is worth trying here and what is noise.
What we're looking for
- Deep working understanding of modern image generation (diffusion, latent spaces, conditioning, identity preservation, LoRA and adapter methods), not just which API to call.
- You read papers and reimplement them, and you can tell a real result from a carefully chosen grid.
- Strong Python and PyTorch, at ease with GPU work and the practicalities of training and inference at scale.
- You design experiments that are capable of being wrong, and you keep the evidence either way.
- Visual taste. You can look at a page, say why it does not work, and put it in words someone else can act on.
Nice to have
- Published work in generative vision, or open-source contributions other people use.
- Fine-tuning or training image models on your own data, not only inference.
- You know how comics actually work (panel grammar, pacing, lettering) from reading a great many or making them.
- Typography or text rendering inside generated images, which is still largely unsolved.
What we offer
- Attractive VSOP options. If you create the growth, you should share in it.
- A profitable company, with a leadership team that has built and sold a similar product before.
- EGYM Wellpass: gyms, pools, and studios across Germany.
- On- and offsites, so team culture and fun don't fall by the wayside.