← All articlesCase studies

What Changed in AI Generation Over a Year: Notes from Product Work

Ryan MitchellNovember 13, 20264 min read
What Changed in AI Generation Over a Year: Notes from Product Work

A year-over-year look at AI generation for product content: what got genuinely better, what's still stuck, and which old habits deserve a second look.

It's easy to describe a year of AI generation progress in vague terms. It's more useful to look at specific product tasks: where things got noticeably easier, and where nothing really moved.

Where things got noticeably better

Hands. A year ago, a product-in-hand shot was a gamble. Now it usually lands on the second or third try. We covered where things stand in detail here.

Text in images. Text in images used to be unreliable across the board, but several models have been fine-tuned specifically for it, and short captions now render cleanly far more often than a year ago.

Speed. Fast models cut wait times from minutes to seconds, and that changed the whole workflow. It's now cheaper to generate several variations than to polish one prompt.

Object consistency. When you generate a scene around a real product photo, the product itself drifts away from the original far less than it used to.

Video with sound. A year ago this was a novelty. Now it's a standard feature.

What hasn't changed

Video length. Still measured in seconds. A longer clip still means stitching several generations together.

Lip sync on dialogue in languages other than English is still uneven, based on what we've tested so far. It holds up for short lines but breaks down on anything longer, more on that here.

Diagrams and structured relationships. A model will draw something that looks similar, not something that's actually meaningful.

Consistency across runs. The same prompt can still produce different results from one generation to the next.

Factual accuracy. A model still doesn't know your actual product and fills in the blanks with something it's never seen.

Old habits worth reconsidering

A few rules that used to be solid advice aren't anymore.

"Always write prompts in English" matters less than it did. On simple scenes, the gap between English and other languages has mostly closed.

"Never bother generating hands" is no longer an absolute rule.

"Pick one model and stick with it" has become a costly habit. The gap between fast and heavyweight models has widened enough that running two tools side by side now makes sense, there's a guide on picking the right one for the job.

Spending a long time perfecting a prompt before hitting generate now loses out to just generating more on fast models.

How to track changes without the noise

Check in once a quarter, not every week. Less actually shifts between releases than headlines suggest.

Keep one test task from your own work and run it against every new model release. That tells you more than any third-party comparison.

Pay attention to improvements in the specific thing that's been a pain point for you, not to general leaderboard rankings.

And don't jump to something new right away. Getting used to a model's quirks takes time, and that investment doesn't pay off instantly.

What none of this changes

A real product in the shot still can't be replaced by generation alone.

Checking output before you publish is still non-negotiable.

The time you save still comes from how well you prepared your source material, not from whichever model you picked.

I went back through my notes from a year ago, and I wouldn't give half of that advice today. The other half hasn't changed at all, and that's the part about preparation and review.

I wouldn't be shocked if that half is outdated too, a year from now. I've given up trying to predict this space.

FAQ

What improved the most? Hands in frame, generation speed, keeping the real object consistent in a scene, and sound in video.

What stayed the same? Video length, lip sync on non-English audio, structured diagrams, and run-to-run consistency.

Which habits should I drop? Always translating prompts to English, avoiding hand shots by default, and relying on a single all-purpose model.

What still matters no matter what? A real product in the shot, checking output before publishing, and preparing good source material.

Should I chase every new model release? No. Getting used to a new model takes time, and that time doesn't pay off right away. Reviewing your toolkit every six months is a reasonable pace.

What won't change going forward? The need for a real product in frame and checking your output before it goes live. It comes down to process, not the tool itself.

Sources

  1. Flami: Putting Products in Hands with AI
  2. Flami: AI Lip Sync
  3. Flami: How to Pick the Right Model for the Job

About the author

Ryan Mitchell

Reviewer at Flami

Read next

Get 15 credits for free

Use them to generate images and videos:

≈ 11 × Nano Banana 2, ≈ 12 × GPT Image 2, ≈ 1 × Seedance 2

Get free credits