← Practical guides

AI RADAR / PRACTICAL GUIDE

How to test an AI video demo before using it on a paid project

Design a three-shot trial to move from an impressive showcase to deliverable work. This original workflow evaluates product accuracy, control, and actual workload separately, with support from official documentation. It does not report a completed performance test.

  1. 01Write down delivery criteria.
  2. 02Generate three shots under controlled conditions.
  3. 03Evaluate raw output and correction work together.
  4. 04Make a project-specific decision to use, limit, or defer.

1. Separate what the demo actually shows

A promotional video shows the appearance of the published example and the workflow being described. It may not disclose how many attempts preceded the selection, whether the footage received additional work, or whether you can obtain the same result with your product. Record these as questions for your own trial, rather than treating them as signs of failure.

For example, seeing a smooth camera orbit does not establish that packaging text will remain correct in every frame. An attractive single shot also does not show whether the same product will survive a client's request for another angle. Frame your evaluation around which part of this delivery the tool can handle reliably, rather than whether it is impressive overall.

2. Write acceptance criteria before generating

Our example task is to prepare three short shots of a beverage bottle on a table. This is a test design you can adapt to your product, not a case study with measured results. Write down the actual delivery's framing, required usable duration, resolution, and product details that must remain unchanged. Specify details such as cap shape, label placement, and brand color.

Also set limits on the time and generation spending allocated to the trial. The numbers can vary by project; what matters is not relaxing the limit after seeing an attractive output. If the client will approve a finished edit, evaluating only the tool's preview is insufficient.

3. Separate the reference image from motion instructions

Runway's image-to-video guide explains that the input image establishes composition, subject, lighting, and style, while text guides motion and progression over time. Treat preparing the product photograph and writing motion instructions as separate tasks. Check the guide's model scope, too: advice written for one version should not be generalized to every model.

For this trial, choose an image with a readable label and clear product edges. Avoid requesting several dramatic movements in the first attempt. The example below separates a stationary product from a slow camera approach. It is a starting suggestion; only the resulting video can establish whether the requested movement occurred.

EXAMPLE PROMPT / ADAPT TO YOUR BRIEF

The bottle remains upright and still on the table. The camera slowly moves closer. Soft studio lighting remains consistent throughout the shot.

Source: Runway — Image to Video Prompting Guide

4. Use three shots to expose different risks

Prepare three separate shots of the same product. Give each shot one main question. This prevents a successful camera move from overshadowing product distortion in another shot. When viewing the three together, also check whether their color, scale, and material appearance suggest the same product.

  • Shot A — Stationary product, slow approach: Do the label, cap, and body proportions hold throughout the shot? Compare the opening and closing frames with the reference.
  • Shot B — Small lateral camera move: Are the bottle's edges, reflection, and contact with the table consistent? Pause and inspect particularly when surfaces hidden in the original photograph become visible.
  • Shot C — Client revision of the same shot: Request shot A again, changing only the movement to a slower pace. Can the requested change be made while preserving the product's appearance and framing? Record the additional attempts the revision requires.

5. Fix the controls and keep a trial log

Choose the reference type for its actual purpose. Adobe describes composition reference in terms of edges, depth, and layout. Its separate motion reference documentation covers transferring camera movement. Do not interpret either control as a guarantee of perfect product identity. Before testing, verify whether the feature is available in your selected model and interface.

For every output, save the model name, date, input file, exact prompt, duration, framing, resolution, and seed if available. Within a comparison, change only the variable you want to examine. When switching models, record settings that cannot be matched directly. Keep all attempts: retaining only the attractive selected clips conceals the workload required.

Source: Adobe — Use video as composition reference

6. Inspect raw output inside the actual edit

Adobe's known limitations page identifies text and symbol distortions in Text to Image. That statement is not a conclusion about all video models, but it is a concrete reason to verify labels separately when generating a starting image. Correct a faulty reference before adding movement to it.

Watch outputs at delivery size and normal playback speed, then pause problematic frames. Do not treat a few flawless frames as representative of the entire clip. Place the clips in the edit, measure the usable duration you need, and record any color matching, masking, stabilization, or label correction required. Count selection and correction time separately from generation time.

Source: Adobe — Known limitations in Firefly

7. Make a project decision instead of assigning one score

Write separate judgments for product accuracy, motion control, response to revisions, and delivery workload. Avoid a combined score that lets good lighting offset a critical label error. The thresholds below are an editorial decision framework, not measured success rates or a performance ranking of tools.

  • Use: Required details remain correct for the necessary duration, revisions are feasible, and the total workload stays within your limit.
  • Limit: Results are acceptable only in certain framings; restrict the tool to those shots and include the necessary corrections in the work plan.
  • Defer: Critical product details distort, revisions are unreliable, or obtaining acceptable output exceeds the workload limit. Choose another production method for this delivery.

Before you deliver

  • Acceptance criteria and a trial budget were written before generation.
  • Raw outputs, prompts, and settings for all three shots were saved.
  • Labels, product shape, and consistency between shots were inspected.
  • A client revision was tested separately.
  • Total workload, including selection and correction, was recorded.
  • The decision applies only to this product, model version, and delivery conditions.

Sources and method

This guide combines the official documentation below with original workflow suggestions from AI Radar. Example prompts are not production test results. Prepared with AI assistance; this is not a claim of independent product testing or human review.

Sources checked: 17/11/2026

Follow the latest news on this topic →

Share your discovery.

Send this story to a friend.

Read next

@airadarhaber

Your radar is on Instagram, too.

Fresh AI developments, dedicated visuals and product examples. Explore our posts and catch up in Stories.

Follow on Instagram

GRAPHIC. DIGITAL. MOTION.

Let’s bring your next idea
to life.

Explore my graphic design, branding
and motion work.

Explore the portfolio