Case Study: Sienna Fox — Building a Virtual Influencer from Zero

Who is Sienna Fox?

Sienna Fox is a 23-year-old digital creator and lifestyle model based in Los Angeles. She loves morning yoga, iced oat lattes, biohacking, and late-night gaming sessions.

She doesn’t exist.

Sienna Fox is the reference virtual influencer built and maintained by the Zirelia project — a real-world test of the engine’s capabilities, running 24/7 on a single server.

Building the Persona

The design process started not with technology, but with character.

The goal was a persona that felt believable and relatable — aspirational enough to attract followers, but accessible enough to feel like someone you could actually know.

Key decisions made during persona design:

• Age 23: Young enough to be authentic in the lifestyle/wellness space, old enough to have opinions and experience
• LA location: Maximum cultural relevance for English-speaking social media
• Mixed interests: Fitness + tech + dating = wide content range without feeling scattered
• Tone: Playful but not juvenile, seductive but not explicit

The full persona was defined in approximately 200 lines of YAML before a single post was written.

Building the Visual Identity

The visual identity was the hardest part.

Step 1: Generated 50 seed images with FLUX.1 using detailed descriptors, with a fixed base prompt and varied angles, lighting, and scenes.

Step 2: Curated the best 30 images manually, discarding any with face inconsistencies or artifacts.

Step 3: Trained a custom LoRA on this dataset over ~45 minutes on a Replicate A100 GPU.

Step 4: Tested the LoRA across 20 different scene prompts, verifying that the face remained consistent across yoga poses, cafe settings, and evening looks.

Total cost: approximately $8 in Replicate compute.

The First Month: Key Observations

After four weeks of operation (manual warm-up for the first two weeks, then full automation):

Content volume: 67 posts generated and published, averaging 2.4 per day with natural variation.

Content quality: The brain successfully avoided repetition in all but 3 cases (which were caught by the memory retrieval system and regenerated).

Image quality: 91% of images passed the quality control loop on the first attempt. The remaining 9% required one retry.

Voice consistency: The most subjective metric — but consistent enough that early followers responded to replies recognizing “her” tone.

What We Learned

The persona.yaml is everything. Vague character definitions produce vague content. The more specific and human the persona, the more specific and human the output.

Visual consistency requires a LoRA. Text prompts alone produce too much variation across dozens of posts. The LoRA investment (time + ~$8) pays for itself immediately.

Warm-up is non-negotiable. Jumping straight to automation on a new account is the fastest path to suspension. Two weeks of manual activity is the minimum.

Memory matters more than expected. Without ChromaDB memory, the brain repeated topics within a single week. With memory enabled, content diversity improved dramatically.

The Engine Behind the Influencer

Sienna Fox runs on a Raspberry Pi at home. Total monthly infrastructure cost: essentially zero (hardware already owned) + variable API costs based on posting volume.

The Zirelia engine is what makes this possible on such modest hardware. The entire pipeline — topic selection, LLM generation, image creation, scheduling, publishing — automated and running without human intervention.

That’s the promise of Zirelia: define a persona, and the engine does the rest.