Characters that hold a brand.
Construction and maintenance of digital characters, for brands, own IPs or narrative universes, with truly coherent visual identity across pieces, not similar faces that subtly change.
Creating a digital character is easy, generating a thousand different images "of the same character" is actually not the same character. Consistency between pieces, sustained over months or years, is where the real work happens. And where most AI projects fail: they generated a nice initial image and six months later they no longer know how to reproduce it.
I have several digital characters in continuous development inside my studio: Vex Muse, Valentina Serrano, Morgan Serafi. Each with hundreds of pieces produced, stable visual identity, and documented processes that allow them to keep growing. That is the real practice I apply when I build a character for a client, whether for a brand, a narrative IP or a creative project of their own.
The process always includes LoRA training (character-specific fine-tuning models), creation of a visual reference sheet, definition of its visual language (palette, angles, camera style, lighting) and production of a first set of 30–50 pieces that validate the system works. From there the character lives and can be maintained.
How we work
-
01
Concept and direction
Who is the character: what it represents for your brand or IP, what visual tone, what age, what aesthetic. Without this, everything that follows is noise.
-
02
Visual sheet and initial dataset
Generation + selection of a base set of consistent images. This dataset is the raw material to train the LoRA. The more angles, expressions and contexts, the better.
-
03
LoRA training and validation
Model fine-tuning, stress tests (does it reproduce the character well in new situations?), iterations until the consistency is solid.
-
04
Manual and initial production
Character documentation (visual bible, effective prompts, what to avoid) and production of an initial set of pieces ready to use.
What's included
- Conceptual definition and art direction of the character
- Generation + curation of the initial dataset
- Custom LoRA training
- Visual reference sheet (poses, angles, expressions)
- Visual bible with usage guidelines
- Initial set of 30–50 pieces ready to use
- Prompt and workflow documentation for future production
When we're not the right fit
- You want "an avatar for my profile" For one-off personal use, a single Midjourney image is enough. My service makes sense when you're going to produce ongoing content.
- You're after impersonating a real person I don't do deepfakes or characters based on real people without explicit, documented authorisation. It's an ethical line I don't cross.
- You need a real-time animated 3D avatar That's a different technical profile (Unreal MetaHumans, Live2D, Vroid). I can direct conceptually but the technical implementation is for another specialist.
Stack & tools
- Stable Diffusion + Flux as main bases
- Local LoRA training (Kohya, OneTrainer)
- ComfyUI + SwarmUI for reproducible pipelines
- Hardware: RTX 4070 Super, 64GB RAM, Ubuntu
- Cloud for heavy training: RunPod when it helps
- Versioning of LoRAs and prompts (Git for text, custom registry for checkpoints)
- Living documentation for each character
If your brand or narrative project needs a digital character with real, maintainable identity, tell me the context and we'll explore together how to build it well.
Let's talk →