Amlan Das recently shared something that caught my attention: "Most AI image generators are useless for DTC brands because they cannot spell." He even gave the painfully familiar example: you ask for a bottle of "Vitality Vitamins," and the model returns a label that reads "VITALIII."

If you have ever tried to generate ecommerce-ready creative with AI, you know that this is not a minor issue. Misspelled packaging text is the fastest way to make an image feel fake, unsafe, or scammy. And in direct-to-consumer (DTC) marketing, trust is the conversion rate.

Amlan also pointed out something important: the situation is changing. With the Feb 26 release of NanoBanana 2 (Gemini 3.1 Flash Image), his team stress-tested the model across common DTC briefs, and found the text rendering capability is finally production-ready. That is a big claim, and it is worth unpacking, because it signals a real shift in how performance teams can create and iterate on ads.

Why spelling is the hidden blocker for DTC AI creative

For years, AI image generation has been strong at vibes and weak at specifics. It could give you a beautiful "lifestyle" scene, but it could not reliably render:

Brand names
Ingredient lists
Supplement Facts panels
Claims and disclaimers
Variants and flavors
Small type on packaging (the stuff regulators care about)

That limitation made AI feel like a toy for DTC teams. You could concept, mock, and brainstorm, but the minute you needed a usable image for Amazon, Meta, TikTok, or an email hero, you were back in Photoshop fixing gibberish.

Amlan’s point is that accurate text changes the ROI math. When the model can handle complex packaging copy, you are no longer generating "inspiration". You are generating production inputs.

"The text rendering capability is finally production-ready; it handles complex packaging copy without requiring extensive Photoshop cleanup."

What NanoBanana 2 changes (and what it still does not)

Amlan’s stress test is a good way to think about AI models: not as a single prompt, but as a system that must hold up across repeatable creative briefs.

From his examples, the model performs across:

Lifestyle shots
Supplement hero shots
Pet product macro views
Fabric detail callouts

In other words, it is not only good at one aesthetic. It can support multiple DTC verticals.

But Amlan also gives a reality check: even with improved text, the model can still drift into generic, stock-like compositions if you do not control lighting, framing, and product presentation.

"But the model still requires strict guidance on lighting and composition to avoid looking like generic stock footage."

This is the part many teams miss. Text accuracy is necessary, but not sufficient. Ads fail for more reasons than spelling.

The real skill: moving from "prompting" to directing

Amlan said it plainly: if you type something vague like "bottle of shampoo on a rocky surface," you will get garbage.

That is not a knock on the model. It is a reminder that professional product photography is a controlled environment: lens choice, depth of field, shadow softness, specular highlights, angle, props, background texture, and color temperature are decisions. If you want AI to output something that competes with a photographer, you need to bring those decisions into the prompt.

Think of it like this:

Amateur prompting describes an object.
Professional prompting specifies a shot.

The technical parameters that matter most for DTC ads

When I translate Amlan’s advice into a practical checklist, I end up with a handful of controllable variables that consistently separate "usable" from "obvious AI":

Lighting: softbox vs hard sunlight, directionality, rim light, shadow density, reflection control for glossy packaging.
Composition: camera height, angle (front-on vs 3/4), negative space for copy overlays, rule of thirds.
Lens and depth: macro vs 50mm product feel, shallow depth for lifestyle vs deep focus for Amazon-style clarity.
Surface and background: material realism (stone, wood, fabric), microtexture, color harmony with the label.
Brand consistency: color palette, typography vibe, prop selection, and the level of "clinical" vs "cozy".

Amlan’s post implies something else: teams should standardize these decisions into repeatable templates, not reinvent them per campaign.

A prompt playbook is really a creative operating system

The most actionable part of Amlan’s post is not the model name. It is the workflow: a playbook of 25 copy-paste prompts that force the model into near-Pro output across DTC verticals.

Here is why that matters. Performance marketing is not a single perfect ad. It is a volume game:

Many concepts
Many angles
Many iterations
Many placements

If your AI workflow cannot reliably output "good enough" creative at scale, it does not help your testing velocity.

Amlan listed what the playbook covers, and you can see a full-funnel perspective in it:

Product hero shots optimized for unbranded discovery on Amazon
Lifestyle context shots that place products in real settings without artifacting
Flat lays, unboxing shots, and subscription box reveals for social
TikTok-native vertical content that looks like UGC, not a brand ad
Before/after splits, testimonial cards, and retargeting FOMO creatives
A Brand Research Prompt that teaches the AI your exact visual identity

That last item is the sleeper feature. Most brands do not fail with AI because the model is weak. They fail because the output is inconsistent. If your creatives look like five different companies from one week to the next, your brand equity leaks.

The batching workflow: where the 80% time savings comes from

Amlan claims the workflow reduces production time for performance creative testing by 80%, with a batching process that can generate 50 variants in under 30 minutes.

That makes sense if you think about where time goes in a typical creative cycle:

Briefing and concepting
Shooting or sourcing assets
Retouching and resizing
Exporting multiple ratios
Creating variations for hooks, claims, and offers

AI does not remove all of that, but it can compress the middle, especially the "get me five more options" step.

A practical way to implement batching without wrecking quality is:

1) Lock the variables that should not change

These are your brand constants:

Packaging layout and copy
Color palette and background range
Lighting style
Camera angle

2) Rotate one variable per batch

This is your testing variable:

Different surfaces
Different props
Different lifestyle context
Different crops (close, medium, wide)
Different placement safe zones for ad text

3) Build a quick QA gate

Even if spelling is better, you still need a review pass:

Verify all label text and disclaimers
Check for warped edges, melted typography, or impossible shadows
Confirm the product silhouette matches the real SKU
Ensure the image does not look like a stock photo clone

AI will not replace your photographer, and that is the point

I like that Amlan explicitly says: "No single AI model replaces your brand photographer (nor should it)." That is the healthiest frame.

The best way to use AI in DTC creative is not as a replacement for brand-building assets. Use it as a high-speed testing engine:

Generate controlled variations to find winners
Prove which angles convert
Then invest photography and video budget into scaling the validated concepts

This approach protects quality while increasing velocity.

The competitive advantage is time, not novelty

Amlan ends with a strategic warning: brands figuring this out now will have a six-month head start.

That is believable because the advantage compounds. Faster testing produces:

Better learnings
Better creative briefs
Better offers and positioning
Better retargeting sequences

And once your team has templates, prompt libraries, and QA habits, your competitors cannot copy that overnight.

A final takeaway

If you are evaluating AI image generation for ecommerce, do not ask, "Can it generate a pretty image?" Ask:

Can it render my packaging text accurately?
Can it match my lighting and brand style consistently?
Can my team reproduce results with templates?
Can we batch variations without quality collapsing?

That is the difference between AI as a demo and AI as a creative production system.

This blog post expands on a viral LinkedIn post by Amlan Das, ceo @ DAS | audience monetization for retail & ecommerce. View the original LinkedIn post →