Training multimodal AI systems to generate product advertisement video concepts from image and text inputs requires more than visual description; it demands creative reasoning grounded in real product inputs, structured to reflect how effective advertising actually works.

The client needed a dataset that could teach models to translate product images and descriptions into coherent, shot-by-shot video concepts, complete with camera direction and voice-over narration. This required solving challenges that go beyond standard annotation:

• Creative grounding without invention: Generating original, imaginative shot concepts derived strictly from provided product inputs, without introducing unsupported product features
• Structural coherence at scale: Ensuring every storyboard followed a logical advertising arc, from product introduction through feature demonstration, lifestyle relevance, and final brand reinforcement
• Voice-over discipline: Producing narration lines that highlighted real product benefits, matched the visual content of each shot, and remained naturally paced for short-form video delivery
• Consistency across product types: Maintaining creative and technical standards across diverse product categories, ad archetypes, and visual styles without sacrificing originality

Turing deployed a team of trained evaluators operating under a structured storyboard authoring and quality assurance workflow designed specifically for image-grounded creative generation tasks.

1. Input validation and product extraction

Each task began with verifying that product images and descriptions were complete and usable. Tasks were rejected if inputs were missing, broken, insufficiently detailed to determine product use, or contained mature or inappropriate content.

Evaluators then reviewed all provided images and descriptions to extract core product features, selling points, physical design details, and relevant use contexts. This extraction step grounded all subsequent creative decisions in the actual product inputs rather than general assumptions.

2. Structured storyboard planning

Evaluators mapped a logical advertising arc across all 15 shots before writing began:

• Each segment was assigned a visual purpose, such as product introduction, feature demonstration, user interaction, lifestyle relevance, or brand reinforcement
• Ad structure conventions were followed: wide establishing shot, close-up detail, feature demo, lifestyle moment, and closing frame
• Evaluators selected an ad archetype suited to the product type, choosing from lifestyle-driven, product-first showcase, feature-led explainer, or playful and animated styles

3. Original, cinematic shot authoring

Each shot was written to communicate a single, uncut sequence of action achievable within the target video format. Shot descriptions covered:

• Scene setting: environment, time of day, mood, and relevant props
• Character actions: specific behaviors, gestures, and product interactions
• Product placement and activation: how and where the product appears in frame
• Camera motion: selected from a defined taxonomy including static, pan, tilt, zoom, tracking, arc, and handheld, with movement direction and smoothness specified

Evaluators were required to imagine and describe original scenes inspired by product inputs.

4. Voice-over authoring

At least 12 of 15 shots required a voice-over line, with each line required to:

• Highlight one specific product feature or benefit visible in the shot
• Match the tone of modern ad narration and remain naturally paced for short-form delivery
• Avoid slogans, invented statistics, or claims not supported by the provided product inputs

5. SOTA model testing and iterative rubric calibration

To validate that delivered storyboards produced high-quality video outputs, we tested scripts against SOTA video generation models, including Nano Banana, VEO, and Seedance. This step confirmed that storyboard descriptions translated into coherent video sequences without hallucinations or generation artifacts across models, providing an additional quality signal.

In parallel, rubric dimensions were progressively tightened based on observed failure patterns in model outputs, ensuring that quality standards kept pace with the task complexity and the capabilities of the models being tested against.

6. Human-in-the-loop quality assurance

All tasks passed through a two-layer quality system combining evaluator-level self-review with independent human QA:

• Evaluators completed a structured pre-submission checklist before tasks entered the review queue, catching common errors such as  mechanical adaptation of product features without natural creative integration, unsupported feature invention, and voice-over disconnects at the source
• Dedicated human reviewers independently assessed every task against a four-dimension rubric covering:
  a. Conceptual accuracy: shots reflected actual product features from the provided inputs, with no speculative or unsupported claims.
  
  b. Visual clarity and creativity: scene descriptions were specific, cinematic, and varied, with appropriate camera motion and logical transitions between shots.
  
  c. Ad structure and purpose: the storyboard followed a coherent advertising arc and maintained a consistent style and tone throughout.
  
  d. Technical and instructional adherence: shots met word-count requirements, voice-over coverage targets, timing feasibility, and prohibitions on on-screen text or third-party logos.

• Delivered more than 7,500 original shot descriptions across 500+ storyboard tasks, each grounded in real product images and descriptions with no invented features or copied inputs
• Produced over 6,000 voice-over lines matched to specific product features or benefits, maintaining natural pacing and narration tone throughout
• Maintained 90%+ quality score across delivered tasks, reflecting consistent creative and technical discipline at scale
• Applied a four-dimension rubric covering conceptual accuracy, visual creativity, ad structure, and technical adherence, with rework enforced before delivery

The client received a structured, high-quality storyboard dataset grounded in real product inputs and built for multimodal AI training and evaluation. With original shot-by-shot descriptions, camera motion specifications, timed voice-over lines, and consistent advertising arc structure, the dataset provides clean creative signal for training systems that generate or reason over product video content.

This foundation supports:

• Training multimodal models to generate product advertisement video concepts from image and text inputs
• Evaluating model outputs for creative coherence, product grounding, and structural ad logic
• Benchmarking image-to-video generation quality across feature alignment, visual creativity, and voice-over relevance
• Scaling storyboard data production across product categories, ad archetypes, and brand styles using a validated evaluator workflow