Yesterday I was trying to explain something we do at work to my kid, and it turned into a fun little story (Inspired by a TV show ‘Cook with Komali’, we were watching yesterday). Sharing it here too.
By definition, LoRA is a lightweight fine-tuning technique that modifies a small subset of a model’s weights to teach it new characters, specific motions, styles, or physics without retraining the entire neural network. Now, here is the analogy:
Meet the AI chef
Think of the video-making AI like a master chef who already knows how to cook thousands of dishes — perfectly. It knows how to chop, season, plate, everything. It learned all this by watching millions of cooking videos. Now imagine you want this chef to also make your grandma’s secret recipe, exactly the way she makes it.
Option 1: Send the chef back to cooking school
You could send the chef back to school and make them relearn cooking — but only teach them your grandma’s recipe this time, over and over. Bad idea. The chef might get your recipe perfect, but forget how to make anything else. Also — cooking school takes years. Way too slow, way too much effort, just to learn one new recipe.
Option 2: A little extra practice, not a whole new education
Here’s the smarter way — and it’s basically what LoRA does.
Instead of retraining the chef from the beginning, you have them do a little bit of extra, focused practice — just on grandma’s dish, just a handful of times. But here’s the important part: while they practice, you only let them adjust a few specific hand movements — say, exactly how they stir and how they season. Everything else about how they cook — their chopping, their timing, their instincts for heat and texture — stays exactly as it was. Locked. Untouched.
After this short practice, the chef has picked up a small, extra “muscle memory adjustment” — just for those few movements. You can think of it like a small add-on skill sitting on top of everything they already knew, rather than a card of instructions they read and follow. It blends in automatically once they’ve practiced it — it’s not a rulebook, it’s a trained habit, just a very small and specific one.
That’s the actual trick: change only a tiny slice of what the chef knows, through real (but short) practice — not instructions, and not a full re-education.
The tricky part with videos
Now here’s where it gets interesting for video. A video isn’t just one dish (may be a frame/image can be compared to a one dish). Video is more like a chef plating the dishes across many moments in a row, where each moment has to flow smoothly into the next. That smooth flow is a separate skill from how the dish looks or tastes on its own.
The real skill in teaching a video AI something new is being very deliberate about which small set of skills get that extra practice — the ones about appearance and identity — while carefully leaving the skills that control that smooth flow between moments completely alone.
Why this is exciting to me?
Currently this is one of the exiting thing we work on at Fabeo — helping AI video tools pick up a specific character or style – through this kind of small, focused practice, without breaking everything else the AI already knows how to do beautifully. In other words, we are not rebuilding the chef. Just a little extra, careful practice in exactly the right place. A small skill added. Same great chef underneath. Perfect new dish, smooth flow and all.