Fine-Tuning & Style Prototypes — NJ MVA Creative Lab

01

Imagen 3 Style Reference

Same visual style applied to 5 different PI/MVA scenes

Reference Style

What this proves

Imagen 3's style reference feature can replicate a specific visual aesthetic across completely different scenes. This means we can establish a consistent "brand look" for all NJ MVA ad creatives without manual editing or Photoshop templates. One reference image sets the tone for hundreds of outputs.

02

Gemini Face Consistency

Same AI spokesperson across 5 ad scenes

Reference Face

Face stays consistent — same person recognizable across all scenes

What this proves

Gemini's multi-reference image generation can maintain facial identity across different ad contexts. This unlocks a recurring AI "spokesperson" for the campaign — the same trustworthy face appearing in car accident, lawyer office, medical, and settlement scenes. No stock photo licensing, no model releases, infinite variations.

03

GPT-4o Vision Scores

AI Creative Scoring — zero-shot, $0.02 for 10 images

edge_07_uber_passenger.png

8.0

Likely to perform well due to high emotional impact and clear messaging.

edge_02_truck_18wheeler.png

7.7

Likely to perform well in capturing attention and evoking emotion, but may need a stronger CTA for better conversion.

edge_03_dui_victim.png

7.7

Likely to engage viewers due to emotional impact but may need clearer CTA for conversions.

edge_01_wrongful_death.png

7.5

May perform well emotionally but could struggle with direct conversions due to unclear CTA.

edge_04_construction_zone.png

7.5

Likely to perform well in engagement but may need a clearer call-to-action to drive conversions.

edge_05_switch_lawyers.png

7.5

Likely to capture attention and evoke emotion, but may need adjustments for authenticity.

edge_06_motorcycle.png

7.5

Moderate engagement with potential for improvement in CTA clarity.

edge_08_gig_driver.png

7.5

Likely to perform well due to emotional impact, but could benefit from improved authenticity and clearer CTA.

edge_09_pedestrian.png

7.5

Moderate engagement with potential for improvement in authenticity and CTA clarity.

edge_10_pregnant.png

7.5

Likely to capture attention and evoke emotion but may need improvement in guiding the viewer to take action.

Dimension Breakdown — Top Scorer: edge_07_uber_passenger (8.0)

Scroll Stop

8

Emotional

9

CTA Clarity

7

Text Readability

9

Authenticity

6

Compliance

9

What this proves

GPT-4o can score ad creatives on 6 dimensions for $0.002 per image. This gives us an automated quality gate — only images scoring 7.5+ go to Meta Ads. The biggest consistent weakness across all 10 images is CTA clarity (avg 5.8/10), which tells us exactly where to focus optimization: add stronger overlay CTAs to every creative.

04

GPT-4o-mini Copy Fine-Tune

Training a custom model to write NJ PI ad copy in our voice

⚙

Fine-tune job running on OpenAI

Results will appear here when the training job completes.
The model is learning from our best-performing ad copy examples.

What this proves

OpenAI's fine-tuning API can create a custom GPT-4o-mini model trained on our specific ad copy style for ~$0.10. Once trained, it generates on-brand NJ personal injury copy at 100x lower cost than prompting GPT-4o with long system messages. This is the path to infinite ad copy variations that all sound like our top performers.

05

LoRA Training Feasibility

Cost analysis for custom image model training on Google Cloud ($310 credit)

Scenario	Base Model	GPU	VRAM	Training Images	Steps	Time	Cost / Run	Max Runs ($310)
A — Budget	SDXL 1.0	NVIDIA L4	24 GB	30	500	27 min	$0.32	984
B — Standard	SDXL 1.0	NVIDIA A100 40GB	40 GB	50	1,000	18 min	$1.10	281
C — Premium	FLUX.1-dev	NVIDIA A100 80GB	80 GB	50	1,500	90 min	$6.78	45

Budget usage per single run (of $310 credit)

A — Budget

$0.32

0.1%

B — Standard

$1.10

0.4%

C — Premium

$6.78

2.2%

Recommended Strategy

5x Scenario A runs to validate pipeline — $1.57
10x Scenario B runs for production quality — $11.01
2x Scenario C runs for FLUX experiments — $13.56

Total: $26.14 — Remaining: $283.86 of $310

What this proves

LoRA fine-tuning on Vertex AI is extremely affordable: a single training run starts at $0.32, and our $310 credit supports nearly 1,000 budget runs or 281 production-quality runs. The recommended 17-run validation strategy costs only $26.14 (8% of budget), leaving $284 for production training. Custom image models that generate on-brand NJ PI visuals are well within reach.