MAY 2026
AI Brand Characters
character LoRAs, over 60 ComfyUI local workflows (no external API) and a GPT-4o QA agent
ComfyUI
LoRA Training
Role
AI Engineer
Timeline
May 2026 to present
team
Bernadette team
platform
ComfyUI, Figma Weavy , RTX 4090, Azure A100 server
The Problem
The brief: three brand characters, fully Custom model training and help with social media asset creation, each needed across multiple expressions and camera angles, as stills and as motion.
The project is under NDA, so this page covers architecture and decisions only. The full write-up sits on a password-protected page. Request access through the contact page.
Constraints
The production workflows call no external API.
Training runs on a local RTX 4090 and on Azure GPUs.
Train with minimal character drift, evaluated by a QA agent and leads, using just 1 video of each character in under 1 month!
How It's Built
My role: AI Engineer, May 2026 to present.
Character LoRAs: FLUX.2 Dev LoRAs for the three characters, trained with ai-toolkit and covering multiple expressions and camera angles.
Captions: a controlled vocabulary, with caption validation.
Generation: over 50 independent ComfyUI graphs
QA: a GPT-4o Vision agent scores outputs against a per-character JSON rubric.
Handover: a technical document covering 12 custom node types across 5 node packs, exact model filenames, per-graph inputs and troubleshooting.
What Went Wrong
[MISSING: one real failure from this build and how you fixed it.]
Outcome
So far the pipeline includes LoRAs for all three characters, ten graphs (six stills, four motion), a vision QA agent with per-character rubrics, a review dashboard and a handover document. The work is ongoing.
What I'd Change
[MISSING: what you would do differently next time.]
