Laura Tornga.
Case Study  /  Visual Style

Training a model on my own hand.

Why I care about local artists being able to keep their storytelling lens, and create their own visual style, in the medium of AI film and cinema. A working case study in personal-model training.

SubjectIndustrial design sketches
ModelFlux LoRA · idsketch
BuiltJune 2026

Generic image models flatten artists into the mean. A personal LoRA does the opposite. It preserves the specific things that make a body of work feel like one person made it: the marker grammar, the page composition, the disciplined use of color. This page is a worked example. The artist is me. The lens is mine. The point is that the same workflow is available to any artist with thirty good drawings.

The argument

AI cinema is at the moment where the technology defaults to a global average. Pretty, competent, recognizable. It is the visual equivalent of every restaurant playing the same chillwave playlist. The interesting future is not bigger base models. It is artists training small, specific models on their own work so the system speaks in their voice instead of the median voice.

For me, that voice was forged at Art Center: marker on layout paper, pencil underpinning, a strict orange accent placed as a single intentional pop. I built a 29-image training set from eight original sketch sheets and trained a Flux style LoRA on it. The trigger word is idsketch. What came out is not a copy of my drawings. It is a model that draws the way I do, on subjects I never drew.

What it learned from
Contact sheet of 29 industrial design sketches used as training data
Training set · 29 single-subject crops from 8 sheets of Art Center industrial design work · marker, pencil, orange accent
What came out

These are LoRA outputs, not retouched sketches. None of these subjects appear in the training set. The style transferred onto products, environments, robots, and people. The orange-accent rule held: one intentional pop, not a flood. That is what a personal model does that a base model cannot.

LoRA hero frame 01
LoRA hero frame 02
LoRA hero frame 03
LoRA hero frame 04
LoRA hero frame 05
LoRA hero frame 06
LoRA hero frame 07
LoRA hero frame 08
LoRA hero frame 09
LoRA hero frame 10
LoRA hero frame 11
LoRA hero frame 12
How it was built
Base model
Flux · fal.ai fast LoRA
Dataset
29 cropped sketches
Training
1,000 steps · style mode

The whole pipeline is a Python script. Upload the dataset, train, pull weights, generate. The hard part was not the training. The hard part was the dataset: choosing which 29 crops to feed the model so it learned the rules instead of any single drawing. That is the artist work. The compute is the easy part.

The deck, How to Train Your LoRA

Originally built as a pitch to architecture-leadership about applying this exact workflow to a hospitality brand. Reframed here as the general argument for personal-style models in cinema.

Slide 1: Title
Slide 2: Thesis
Slide 3: Sketches A
Slide 4: Sketches B
Slide 5: Sketches C
Slide 6: How it was built
Slide 7: The hand
Slide 8: Every category
Slide 9: Entire worlds
Slide 10: Characters
Slide 11: Every property
Slide 12: Hand to home
Slide 13: Concept, plan, render
Slide 14: Blueprints
Slide 15: Guest moments
Slide 16: The ask
Live demo →
Why this matters

The base-model future is a feed of work that looks like everyone else's work. The personal-model future is a feed where a viewer can tell which artist made a frame from across the room. That second future is the one I want to live in, and it is the one I want to help other artists build. The bottleneck is no longer the technology. It is the artist's willingness to gather their own work and train on it.

If you are an artist, a studio, or a brand thinking about building a signature visual model from your own catalog, this is exactly the kind of project I take on. The screenplay-to-generation pipeline is the production-grade version. The AI films page is the work that comes out the other end.