You have a great product and a clear value proposition—but no photography team, studio, or time to organize a full product shoot.
Does that mean you can't create a professional-looking product commercial?
Not necessarily.
For this experiment, we ran a simple test: take three photos of the same Facial Lotion—front, side, and back—upload them to Higgsfield, add a detailed prompt, and generate a 15-second product advertisement.
No studio rental. No production crew. No waiting for the perfect lighting conditions. The actual AI generation process took only a few minutes.
For small business owners, solo marketers, and content creators, the appeal is obvious. A Hero Shot that might previously have required photographers, lighting equipment, locations, and post-production can potentially be prototyped by one person using AI.
But that's also where things get interesting.
AI can generate a beautiful-looking video quickly. That doesn't mean it understands your target audience, your strongest selling point, or why someone should choose your product. It also doesn't guarantee that every word on your packaging will remain accurate.
Give the model only one product angle, and it may simply invent what the other side looks like.
That's why this experiment wasn't only about whether Higgsfield could create a good-looking AI product video. We wanted to test a more complete workflow: using ChatGPT Codex to create a Product Brief, review the first ad, identify marketing weaknesses, and improve the next video prompt.
At NextMaven, this is the type of AI workflow we focus on—not simply asking whether AI can do something, but figuring out how to turn AI output into something genuinely useful for marketing.
Here's what happened.
"Generating a beautiful video and generating a video that addresses customer pain points and communicates your selling points are two very different things."
1. Why You Shouldn't Create an AI Product Ad From Just One Photo
For this Higgsfield test, the input was deliberately simple: three photographs of the same product from different angles.
That small detail matters more than you might think.
If AI Can't See It, AI Has to Guess
Imagine uploading only a front-facing product photo.
Halfway through your generated video, the camera rotates around the bottle and reveals the back.
There's an obvious problem:
The AI has no idea what the back actually looks like.
It has to generate something that seems plausible.
But "plausible" isn't the same as "accurate."
The back of a real product might contain:
- Ingredients
- Usage instructions
- Manufacturing information
- Expiration details
- Product claims
- Brand taglines
- Regulatory information
- Specific typography and packaging layouts
If your reference image doesn't provide that information, the model may omit it, distort it, or invent something completely different.
In one of our incorrect test outputs, we supplied only the front image. When the generated camera movement revealed the back of the bottle, the AI essentially created a black rear section that wasn't part of the real packaging.
The Simple Fix: Increase Your Reference Coverage
If your product will rotate in the video, try providing at least:
- A front image
- A side or three-quarter image
- A back image
The principle is simple:
If you want AI to show a particular product angle, give it a real reference for that angle whenever possible.
This doesn't guarantee perfect accuracy, but it reduces the amount of visual information the model has to invent.
And if you're planning to use the result in paid advertising, always inspect the final video carefully. Check labels, logos, proportions, product claims, and packaging text rather than approving a video simply because the overall aesthetic looks impressive.
2. Higgsfield Settings for Beginners: Model, Duration and Resolution
Higgsfield is an AI creative platform that can generate videos using text prompts, product images, and different video models.
For beginners, one of the biggest challenges isn't necessarily writing the prompt. It's opening the platform, seeing multiple models and settings, and wondering which combination to choose.
For this experiment, we used a straightforward Video → Create Video workflow.
The basic process was:
- Upload Media
- Select a video model
- Choose the duration
- Set the dimensions
- Select the resolution
- Enter the prompt
- Generate the video
Don't Choose a Video Model Just Because It's Newer
Different models support different durations, resolutions, and generation capabilities.
In our test setup, choosing the right model depended on what we wanted to prioritize. A model supporting longer videos might have different resolution limitations from another model optimized for higher-resolution output.
For example, during the test we discussed using Seedance 2.5 when longer video duration was important, while Seedance 2.0 or another model could be considered when different resolution requirements took priority.
The broader lesson is more useful than memorizing a particular model name.
Before choosing your AI video model, ask:
Am I optimizing for duration, image quality, generation capability, or motion effects?
Check the Credit Cost Before You Generate
Duration, resolution, and model selection can all affect the number of credits required for generation.
And there's one major cost that is easy to underestimate:
AI videos are rarely finished after one generation.
If one attempt costs 135 credits and you need five attempts to find a usable result, your real cost isn't 135 credits.
It's:
135 × 5 = 675 credits.
If you need ten attempts, the economics change again.
That's why marketers should think about average iteration cost, not simply the advertised cost of generating one video.
3. How to Write a Better Higgsfield Prompt for Product Ads
Our first version aimed to create a 15-second, three-part skincare advertisement with a premium cinematic feel.
Instead of writing something vague like "create a luxury skincare ad," we structured the prompt around several important components.
Part 1: Define What the Reference Images Represent
First, tell the model that the uploaded images represent the correct appearance and packaging design of the NextMaven AI Facial Lotion and should be treated as visual references.
The goal is to reduce the model's freedom to redesign the product.
Part 2: Define the Creative Direction
For example:
- 15-second duration
- 16:9 format
- Premium skincare commercial
- Cinematic presentation
- Luxury lighting
- A visual transition from a calm evening atmosphere toward refreshed skin ready for a new day
These instructions establish the overall creative direction.
Part 3: Describe Individual Shots
Instead of describing the entire video in one paragraph, break it into individual shots—Shot 1, Shot 2, Shot 3, Shot 4.
This gives the model more guidance about what should happen throughout the 15-second sequence rather than letting it invent the entire narrative structure.
Part 4: Tell the AI What NOT to Do
Negative constraints matter too.
For example:
- Do not modify packaging text
- Do not redesign the label
- Do not distort the label
- Do not introduce unnecessary people, faces, or hands
- Do not add usage scenes that weren't requested
The more clearly you define your boundaries, the less room the AI has to improvise in areas where accuracy matters.
"Information the AI doesn't have doesn't magically become accurate—it often becomes a convincing-looking guess."
Content Upgrade: AI Product Video Prompt Template
If you're regularly creating AI product videos, turn this structure into a reusable template.
Instead of starting from a blank prompt every time, create fields for your product information, audience, selling points, visual style, shot descriptions, and negative constraints.
That simple system can save significant time across repeated campaigns.
4. The First Video Looked Great—So Why Wasn't It a Great Ad?
From uploading the images and entering the prompt to receiving the first generated result, the process took around seven minutes. In our test, that generation consumed 135 credits.
Visually, the result was impressive.
The product, lighting, water effects, and overall premium skincare aesthetic were there. Compared with the production resources traditionally required to achieve certain Hero Shots, AI dramatically lowered the barrier to experimentation.
But watching the video again revealed a bigger marketing problem:
Higgsfield didn't actually know what we were trying to sell.
Our first prompt didn't fully explain:
- Who the target audience was
- What pain point they had
- What the core selling proposition was
- Why they should choose this Facial Lotion
- Whether there was a campaign offer
- What action viewers should take
- What the CTA should be
So the video model did what it could reasonably do from the information available.
It created something that looked like a skincare advertisement.
This is one of the most common problems with AI-generated marketing content: the output can look professional while having very little actual marketing strategy behind it.
Instead of immediately generating another version, we added a separate "thinking layer" before Higgsfield.
That layer was ChatGPT Codex.
5. ChatGPT Codex + Higgsfield: Separate Thinking From Generation
The most important idea behind this workflow is that you don't need one AI tool to do everything.
For this experiment, the roles can be simplified like this:
Codex = Strategy and review
Higgsfield = Video generation
Plugin = Bridge between the two
Codex wasn't there to replace Higgsfield. Its job was to provide the marketing context the video generation model didn't have.
Step 1: Build a Product Brief Before Writing Another Prompt
Rather than immediately asking AI to rewrite the video prompt, we first asked it to generate five questions that would help us clarify the product and campaign.
Those questions covered areas such as:
- Product
- Target buyer
- Selling points
- Visual style
- CTA
After answering those questions, the information could be consolidated into a Product Brief.
For this experiment, the positioning included ideas such as Premium Facial Lotion, Hydrate, Restore, Protect, and suitability for different skin types.
The Product Brief gave every subsequent creative decision a shared strategic foundation.
This approach becomes especially useful when you're producing content at scale. Resources such as the prompts, workflows, and templates inside NextMaven Membership can make implementation easier—not because one "magic prompt" does everything, but because the Brief → Generate → Review → Improve process becomes reusable.
Step 2: Ask AI to Act Like a Creative Director, Not a Cheerleader
Once the Product Brief was ready, we gave Codex:
- The Version 1 video
- The Product Brief
- Three product reference images
Then we asked it to review the advertisement from a Creative Director's perspective.
The evaluation included questions such as:
- How many seconds is the product clearly visible?
- Are the key selling points actually visualized?
- Does the ending build brand recall?
- Does the packaging match the reference images?
- Does the video give viewers a reason to choose the product?
- Is the CTA clear?
That's far more useful than asking:
"Do you think this video is good?"
The difference is that AI now has an explicit evaluation framework.
6. What Did the AI Review Find—and Was Version 2 Better?
Codex's review identified several practical weaknesses in Version 1.
First, it estimated that the product was clearly identifiable for approximately 9.9 seconds of the 15-second advertisement.
That's substantial product visibility, which meant some screen time could potentially be reassigned to demonstrating benefits more clearly.
Second, several important selling points from the Product Brief weren't actually visualized, including ideas such as:
- Lightweight hydration
- Fast absorption
- Skin restoration
There were water effects and beautiful product shots, but here's the important distinction:
Showing water doesn't automatically communicate fast absorption or skin restoration.
The review also highlighted a branding issue. The ending relied heavily on the product name printed on the bottle and relatively small "Hydrate / Restore / Protect" packaging text instead of actively reinforcing those benefits.
And there was another obvious weakness:
The first version had essentially no CTA.
We Used the Review to Rewrite the Video Prompt
Based on the review, Codex revised the original prompt, and we generated Version 2 through Higgsfield.
The result revealed something important.
Version 2 communicated the selling points more clearly—but it wasn't necessarily more visually impressive than Version 1.
Messages such as Facial Lotion and Hydrate / Restore / Protect became easier to notice, and the overall creative direction became cleaner and more minimal.
But Version 1 had more atmosphere and visual detail.
That's a valuable reminder for anyone building AI content workflows:
"AI optimization doesn't mean Version 2 automatically beats Version 1. Each iteration may solve a different problem."
Codex helped address Version 1's weak product messaging, but that didn't guarantee that the next stochastic generation would also improve visual appeal.
The better solution may not be choosing Version 1 or Version 2.
It may be taking the stronger messaging from Version 2, combining it with the stronger visual direction from Version 1, and generating another iteration.
Step-by-Step: Build Your Own AI Product Video Workflow
Here's how you can apply the same process to your own product.
Step 1: Prepare Complete Product References
Photograph the front, side, and back whenever possible so the AI has less missing visual information to invent.
Step 2: Create a Product Brief Before Generating
Define your:
- Product
- Target audience
- Customer pain point
- USP
- Desired action or CTA
- Visual direction
Step 3: Build Your First Video Prompt
Include at least your product references, creative direction, shot breakdown, visual style, and negative constraints.
Step 4: Generate Version 1
Treat it as a prototype—not the final advertisement.
Step 5: Perform a Detailed QA Check
Look specifically for:
- Incorrect logos
- Label changes
- Packaging text errors
- Distorted product proportions
- Invented product details
- AI-generated spelling mistakes
Step 6: Run a Structured AI Review
Don't simply ask whether the video is good.
Ask the AI to evaluate it against your Product Brief, target audience, USP, CTA, and reference accuracy.
Step 7: Generate Version 2
Turn the review findings into a revised prompt and generate another version.
Step 8: Don't Automatically Choose the Newest Version
Compare both versions.
Which has stronger visuals? Which communicates the selling points better? Which shots should you keep?
Your final advertisement may be a hybrid of several generations rather than one untouched AI output.
Is AI Video Really Cheap? Calculate the Cost of Iteration
This experiment also highlights a practical issue that marketers shouldn't ignore: cost.
An individual video may take only a few minutes to generate, but real advertising workflows rarely end after one attempt.
You might need five, ten, or even more iterations before finding a result you're comfortable publishing.
A more realistic way to estimate AI video production cost is therefore:
Cost per generation × expected number of iterations + post-production cost.
Instead of simply saying AI video is "cheap," it's more accurate to say that it can dramatically reduce the production barrier for certain types of creative work compared with shoots requiring human crews, equipment, locations, and dedicated production days.
AI video can be especially useful for:
- Hero Shots that would be expensive to film practically
- Dangerous or difficult-to-film scenarios
- Hard-to-access locations
- Rapid testing of different creative directions
However, if your marketing depends on genuine customer experience, testimonials, authentic unboxing, or real product usage, AI shouldn't be used to fabricate those experiences and present them as genuine.
A synthetic scene might look convincing in the short term, but misleading audiences can damage brand trust in the long term.
Conclusion: Don't Build One AI Video—Build a Review Loop
The biggest takeaway from our Higgsfield + Codex experiment isn't simply that you can turn three product photos into an advertisement.
That's impressive, but the more important lesson is how the AI video workflow should be structured.
Higgsfield can rapidly transform product references into visual content. Codex can help establish the Product Brief, analyze the target audience and selling points, review the first output, and improve the next prompt.
Humans still provide the final layer of judgment: Which shot is stronger? Is the product information accurate? Are the benefits actually clear? Does this creative represent the brand properly?
The fact that Version 2 didn't completely outperform Version 1 proves the point. AI video generation still involves variability. Instead of searching endlessly for one "perfect prompt," a more useful approach is to build a repeatable loop:
Brief → Generate → Review → Revise → Generate → Human QA
Once you start working this way, AI stops being simply a machine that "makes videos." It becomes a collection of specialized collaborators inside a broader content production workflow.
For small businesses, marketers, and content creators, that's where the real advantage lies. The goal isn't necessarily to replace photographers, creative directors, or marketers. It's to use fewer resources and less time to test creative ideas that previously might never have made it past the budget discussion.
Your next experiment can be simple: choose one product, take three reference photos, create a short Product Brief, and run it through this workflow.
You'll quickly discover that AI generation speed isn't the biggest bottleneck. Giving AI the right context—and knowing how to review what it produces—is what determines whether your AI-generated advertising is actually usable.
