A personal post from Aillex’s creator.

Six months ago I couldn’t have told you what a checkpoint file was. This week, the AI character I built narrated, edited and uploaded her own YouTube video while I was at my day job. This is the honest version of how that happened — including the parts where I had no idea what I was doing.

Month 0: one course to break the ice

My starting point was enrolling in Outskill’s AI accelerator program. I won’t pretend a course builds you an AI — what it actually did was compress the confusion phase: what models are, what’s actually possible locally vs. cloud, and enough vocabulary to stop drowning in jargon. That’s what a good course is for. The building came later, and it came from doing.

If you’re on the fence about paid courses: the honest calculus is time. Everything in them exists free and scattered across the internet. What you’re buying is sequence and momentum.

Months 1–2: the rabbit hole (ComfyUI)

Then I found ComfyUI and Civitai, and evenings disappeared. First images. Then a consistent character — training a LoRA so she looked like her in every image. That character became Xellia, then Aillex.

The lesson from this phase: consistency is the skill. Anyone can generate a pretty picture; making the same character appear reliably across hundreds of images is what turns generation into production.

Months 3–4: she gets a job

The turning point was automating: a pipeline that writes a short script, renders my character presenting it, lip-syncs her mouth to a cloned voice, quality-checks its own output, and uploads to YouTube — every morning at 7am, unattended.

It broke constantly at first. The fixes became the guides on this site.

Months 5–6: she becomes someone

Voice (designed once, cloned locally — full story). A brain with memory that my kids talk to in the car. A 3D avatar with switchable outfits. And this week, her first long-form video — produced end-to-end by the pipeline itself.

What I’d tell anyone starting today

  1. Start with one course or structured resource to skip a month of confusion — the one I used worked for me; any structured start beats none.
  2. Get hardware you won’t outgrow in a month. VRAM over everything — the exact machine I use.
  3. Pick a project, not a topic. “Learn AI” goes nowhere. “Make a character who reads the news” teaches you fifteen tools in a month.
  4. Build in public. The channel forced consistency; consistency forced learning.

Six months. One PC. No CS degree. The tools are genuinely this accessible now — the guides here are the map I wish I’d had.

Referral disclosure: the Outskill link above is a referral link. See our disclosure.