Terminal.skills
Use Cases/Generate AI Media Assets at Scale with Tested Prompts

Generate AI Media Assets at Scale with Tested Prompts

Create consistent, brand-aligned images and videos using Vertex AI media generation with systematically tested and optimized prompts.

Data & AI#vertex-ai#imagen#veo#image-generation#video-generation
Works with:claude-codeopenai-codexgemini-clicursor
vertex-media-generation
prompt-tester
$

The Problem

A content marketing team at an e-commerce company produces 200 product lifestyle images and 30 short video clips per month for social media, email campaigns, and landing pages. Traditional photo shoots cost $8,000 per session and take 2 weeks from scheduling to delivery. AI image generation could cut that to hours, but the results are inconsistent -- one prompt produces a great hero image while a slight variation generates something off-brand with wrong colors, weird proportions, or inconsistent style. They need a repeatable system where prompts are tested, scored, and optimized before running batch generation, so every asset meets brand standards without manual trial and error.

The Solution

Using the vertex-media-generation and prompt-tester skills, the workflow first builds and evaluates prompt templates against brand criteria using systematic A/B testing, then runs the winning prompts through Vertex AI Imagen and Veo to generate hundreds of on-brand images and videos in a single batch.

Step-by-Step Walkthrough

1. Define brand criteria and build prompt templates

Create structured prompt templates with variables for product type, scene, and mood, then define the evaluation rubric that every generated image must pass.

Create 5 prompt template variants for product lifestyle photography. Each template should include variables for {product_name}, {scene_setting}, {lighting_style}, and {color_palette}. The brand guidelines are: warm natural lighting, earth-tone backgrounds, minimalist composition, no text overlays, 16:9 aspect ratio. Define a scoring rubric with these criteria: brand_consistency (0-10), composition_quality (0-10), product_visibility (0-10), and usability_score (0-10).

The prompt templates separate structure from content, so you can swap in any product while maintaining consistent style. Template A might use "editorial product photography" framing while Template B uses "lifestyle scene with product in natural use."

2. Test prompt variants systematically

Run each template variant through the prompt tester to generate sample outputs, score them against the rubric, and identify which template consistently produces the best results.

Test all 5 prompt templates with 3 different products each (wireless earbuds, ceramic mug, leather wallet). Generate 2 images per combination using Vertex AI Imagen 4.0. Score each output against the brand rubric. Show me a comparison matrix of average scores per template, and identify which template performs best overall and which performs best per product category.

The systematic testing eliminates guesswork. Instead of running 50 ad-hoc generations and picking favorites, you generate 30 controlled samples and let the scoring rubric identify the template that consistently hits brand standards. Template C might average 8.4/10 overall while Template A averages 6.9 -- a difference invisible from looking at two cherry-picked examples.

3. Generate image assets in batch

Run the winning prompt template through Vertex AI Imagen to produce the full month's image assets in a single batch.

Using the winning prompt template, generate product lifestyle images for all 40 products in the Q1 catalog. Generate 3 variations per product at 1024x1792 resolution. Use Imagen 4.0 with style_preset set to "photographic" and negative prompt "text, watermark, blurry, distorted, cartoon." Save to ./generated_assets/{product_slug}/ with descriptive filenames.

The batch job tracks progress and flags quality issues automatically:

Batch Generation Report — Q1 Catalog
=======================================
Total products:     40
Variations/product: 3
Total generated:    120 images

Quality Scores (automated rubric):
  Passed (>= 7/10):  106 images (88.3%)
  Flagged (< 7/10):   14 images (11.7%)

Flagged breakdown:
  Low product_visibility:  6 images (product partially occluded)
  Low brand_consistency:   5 images (color palette drift)
  Low composition_quality: 3 images (awkward cropping)

Generation time:  52 minutes
API cost:         $43.20 (120 images x $0.36/image)
Storage:          2.1 GB (1024x1792 PNG)

Flagged images queued for regeneration with adjusted parameters.

The three variations per product give the design team options without requiring multiple prompt iterations. Flagged images get regenerated with adjusted parameters -- tighter cropping guidance for visibility issues, explicit color hex codes for palette drift.

4. Generate short video clips with Veo

Produce 15-second product showcase videos using Vertex AI Veo, using the same tested prompt patterns adapted for video generation.

Generate 15-second product showcase videos for the top 10 products using Veo 3.1. Adapt the winning image prompt template for video: start with a close-up of the product, slowly pull back to reveal the lifestyle scene, end on a hero composition. Use 1080p resolution at 24fps. Generate 2 variations per product.

Video generation uses the same prompt principles validated in the image testing phase, adapted for temporal storytelling. The consistent style between still images and video clips creates a cohesive brand presence across channels.

Real-World Example

An e-commerce brand producing 200 lifestyle images monthly switched from agency photo shoots to this AI pipeline. The prompt testing phase took one afternoon and identified that "editorial flat-lay" templates scored 34% higher than "lifestyle scene" templates for their minimalist brand. Batch generation of 200 images through Vertex AI Imagen cost $45 in API credits and completed in 90 minutes. The design team rejected 12% of outputs (vs 8% rejection from traditional shoots), but the cost per usable image dropped from $38 to $0.26. Monthly video production went from zero (too expensive at $2,000 per clip) to 20 clips per month at $3.50 each. Total monthly content production cost dropped from $8,000 to $340, and turnaround went from 2 weeks to same-day delivery.

Tips

  • Test at least 5 prompt template variants with 3 products each before committing to batch generation. A small upfront testing investment prevents regenerating hundreds of off-brand images.
  • Include explicit color hex codes from your brand guide in the prompt rather than color names. "Warm beige" is ambiguous; "#D4C5A9 background" is reproducible.
  • Generate 3 variations per asset and let the design team pick. The marginal cost of extra variations is negligible compared to the time spent iterating on prompts for a single perfect output.
  • Save the winning prompt templates in version control. When the brand guidelines update, you can diff the old and new templates to see exactly what changed.