CUSTOM STYLE TRAINING (LoRA) PIPELINE

ID: AXO-LORA-OPS
DATE: 2026-01-30
LOCATION: Design District, Pioneer Square, Seattle, WA
CONTEXT: Workflow Optimization / Firm-Specific Branding

> QUERY: "Staff training for Revit-to-AI workflow; implementing LoRA for strict style/site consistency; setup and maintenance contract."

1. OPERATIONAL SCAN (THE PROBLEM)

Agent analyzed standard AI output vs. Firm Requirements.

Constraint Identified: Generic models (Stable Diffusion XL, Flux) default to a "Pinterest/ArtStation" aesthetic that lacks the firm's specific branding or regional context (e.g., "Pacific Northwest Modernism").
The Conflict: Without custom training, staff wastes hours "prompt engineering" trying to remove hallucinations. Inconsistent output across different project phases confuses clients.
The Agentic Fix: Train two distinct LoRA (Low-Rank Adaptation) models:
1. [FIRM_STYLE_V1]: Trained on 30+ past high-quality renders to lock in lighting, color palette, and material vibes.
2. [SITE_CONTEXT_PNW]: Trained on local flora/weather data to ensure "Grey Sky" realism rather than generic "California Sun."

2. TECHNICAL IMPLEMENTATION & CURRICULUM

> DATASET PREP: Requires curation of 20-50 high-res images from the firm's portfolio. Images must be captioned (tagged) manually to teach the AI what represents the "Firm Style."
> WORKFLOW: Revit Massing > ControlNet (Geometry Lock) > LoRA Injection > Upscale.

      [LoRA TRAINING LOOP]
      _________________       __________________
     | PAST PROJECTS   | ==> |  CAPTIONING TOOL |
     | (DATASET)       |     | (KEYWORD TAGS)   |
     |_________________|     |__________________|
                                     |
             ________________________|
            |
      ______v______       ___________________
     |             |     |                   |
     |  TRAINING   | ==> |  .SAFETENSORS FILE| <-- THE "BRAIN"
     | (KOHYA_SS)  |     |   (200MB ASSET)   |
     |_____________|     |___________________|
                                 |
                          INJECT INTO GENERATOR
    

3. COST & SCHEDULE BREAKDOWN

PHASE	DELIVERABLE	ESTIMATE
SETUP (HARDWARE)	Config Local UI / Cloud GPU Instance	$1,200
LoRA TRAINING	Data Cleaning + Model Training (x2 Models)	$2,500
STAFF WORKSHOP	2-Day Intensive (Workflow + Ethics)	$3,500
MAINTENANCE	Monthly Retraining (New Projects) + Support	$600 / mo
TOTAL LAUNCH	Timeline: 3 Weeks	~$7,800

MANUAL RENDER TIME 12 Hrs / View

AI + LoRA WORKFLOW 45 Sec / View

> DOES THIS ALIGN WITH YOUR TRAINING BUDGET?

[ BOOK TRAINING SESSION ]