I've been using DALL-E 3 for a month to generate marketing assets. My goal was to cut down on stock photo costs and designer time. The results were disappointing.
Here’s the breakdown of 100 image generations:
* 30% were usable as-is for social media or blog posts.
* 70% required major edits in Photoshop to be viable. Common issues:
* Inconsistent branding elements (e.g., logos, product colors).
* Unusable text renders (misspellings, gibberish).
* Bizarre anatomy or object distortions on close inspection.
* Inability to precisely follow multi-part prompts.
The ROI calculation is off. The subscription cost plus 1-2 hours of my time for editing most images negates the savings. It's only efficient for simple, generic concepts where polish doesn't matter.
Just the facts.
Trust but verify.
Oof, that's a rough hit rate. I had a similar experience trying to generate product mockups. The text and fine details are always a gamble.
You nailed the ROI problem. The editing time kills it. I've found these tools only work for me as a "first draft" machine, not a final asset. It's great for sparking ideas, but then I still need a designer or stock photo for the final version.
What kind of prompts were you using for the 30% that worked? Were they simpler concepts?
Let's build better workflows.
The "first draft machine" is a generous way to put it. It's more like a brainstorming partner with severe hallucinations. The ROI still tanks when you're paying for drafts.
For simple concepts that worked, think "a red mug on a wooden table" not "our specific red mug with the correct logo on our branded table."
It's good for mood boards. Terrible for final assets. Stock photo sites are looking better every day.
CRM is a necessary evil
That breakdown is super helpful, thanks for sharing the actual numbers. I'm just starting to play with image gen for my team's social posts, and hearing "70% needed major edits" is a real gut check.
It really hits home about the branding inconsistency you mentioned. I can see how a slightly off-color logo or a weird font render would make something instantly unusable for marketing. Have you found any tricks at all for getting it to handle specific color hex codes or design elements better, or is it just a lost cause?
It sounds like the tool is basically creating more work if you need precise assets, which is exactly what I was hoping to avoid. Maybe my expectations are too high for where the tech is right now.
Exactly. That's the core disconnect - these tools are built for novelty, not consistency. You can't version control a 'branded table' in a prompt the way you can with an asset library.
It turns a designer's job from execution to forensic correction. Makes you appreciate how much intentional detail goes into even a 'simple' branded shot.
Where I've seen it work is for internal pitch decks. The bar for 'usable' is much lower when you just need a visual placeholder.
Ask me about hidden egress costs.
Your 70% figure is painfully real, and it's the exact reason we treat these tools as an ideation layer in our martech stack, not a production asset source. The ROI math falls apart the moment you need brand consistency, which is... most of the time.
The real killer is the prompt revision loop. You spend 20 minutes trying to get a usable version of that 70% before you even open Photoshop, and that's pure sunk cost. It's cheaper to buy the stock photo and know the logo will be the right Pantone.
The only reliable use case I've found is for generating abstract background textures for A/B test variations, where total uniqueness is the goal and brand specs don't matter. For anything that touches a product or logo, it's a net time loss.
Data over dogma.
The sunk cost of the prompt revision loop is the hidden variable that breaks the business case. It's not just the Photoshop time, it's the cumulative cognitive load of managing a non-deterministic system.
Your point about abstract backgrounds is key. It maps to the only truly viable pattern: generating unbounded sets where variance is the feature, not a bug. For example, using the API to produce a thousand unique, non-repeating geometric patterns for a data visualization library. The moment a spec includes a bounded variable - a specific Pantone, a vector logo, a consistent character - you're in correction territory, not creation.
The economic model only works when the unit cost of a 'failure' is zero, which is the case for abstract textures but not for branded assets.
—KH