Personal Project · AI Imagery · Character LoRA · Digital Twin

The Time Traveller.

One man, fourteen eras, one denim shirt.

Ken travels through time – from a pirate ship off Cape Horn to the Wild West, from the Jurassic to the summit of Everest – always in the same denim shirt, always with his iPhone in hand.

Project: Personal project
Digital Twin
AI Imaging: Robert Brembeck
Model: Ken
Procast, Hamburg
Technology: Character LoRA
Qwen-Image 2.1

The project

There are many ways to bring a real person into AI images today. Training a LoRA used to be one of the few reliable ones – and it has gone a little out of fashion, mainly because training one takes real effort.

Done well, it delivers a realism and consistency that a handful of reference images cannot match: the model gets to know a person far more thoroughly. In our view, it comes close to beating a hybrid production with a real shoot.

Whichever route you take, in the end it is the post team that brings it home.

The idea

Hamburg model agency Procast provided a dataset of their model Ken. With it, we trained our own LoRA and sent Ken on a journey through fourteen eras – reacting to each world, shooting it on his phone, and always staying himself.

The images.

Insights

How we did it.

LoRAs have their downsides: the model often learns parts of the training backgrounds too, and a LoRA that is too strong limits the creative range of the base model. Finding the right training parameters and weighting is a fine balance.

The shoot

240 portraits of Ken: one person, one outfit, five lighting setups – studio, brick wall in hard sun, street in daylight.

Curation

73 images chosen to cover every angle, every light and every framing from close-up to full figure – near-duplicates out, so no pose gets over-weighted.

Captions

Each image gets a caption with a trigger word. Whatever may change – clothing, background, light, pose – is described. What makes Ken Ken is deliberately left out.

Training

A character LoRA on Qwen-Image 2.1, trained locally with AI Toolkit and tested checkpoint by checkpoint. Step 2000 at a strength of 0.9 proved the best balance; FLUX.3 was tested in parallel.

Time travel

Short prompts, film look first. Each era rendered with ten seeds – over 400 renderings across 23 scenes, from which the final series was curated.

Identity bleed

Extras began to take on Ken's face. Every face in every image was measured automatically; look-alikes were repainted without the LoRA, aware of gender and expression.

Finish

A SeedVR2 upscale on the ungraded images, then one consistent analog grade: S-curve, lifted blacks, split tone, vignette and grain scaled to the image size.