Predicting reel views before posting: my model with 89 features
Big models give me a paragraph of opinion. Small ones give numbers. From numbers I built a model that rates my reels before I post. What works and the catch.

On Wednesday I rendered v1 of a reel about AI models. 48 seconds. Before I showed it to anyone, I got a report: first cut too late, one shot in the middle held too long, the window in the background a bit blown out. And a forecast: about 1.36 times the median views of my profile.
Sounds like magic? It isn’t magic. It’s a few small models, a lot of number crunching and one simple model at the end. I’ll tell you how it works, why I got into this at all and where the catch is. Because there is a catch.
Why don’t I ask the chat whether a reel will catch on?
Because a big model gives a paragraph of opinion, not a number I can compare with v2 or check a week later.
The first thing I did, of course, was ask a big model. You paste in the transcript, a description of the shots and ask: will it catch on or not? You get a nice paragraph. Strong hook, good pace, maybe worth trimming the middle.
Ok, but what do I do with that paragraph? I can’t compare it with yesterday’s paragraph. I can’t work out whether v2 is better than v1. I can’t check after a week whether it was right, because it said nothing you can measure.
I needed something that returns numbers. And here’s the thing hardly anyone talks about: small models have a superpower. They hand back raw data. Not text, but vectors and probabilities.
- DINOv2 gives a vector for every frame, a description in numbers of what is in the shot.
- Whisper gives word timings, down to a fraction of a second.
- AST recognizes 527 classes of sound: speech, music, silence, effects.
- V-JEPA2 describes motion over time, not single frames.
Claude and GPT won’t give you those numbers. They give you an answer. And I wanted material to train on.
89 features from one reel
One reel gives me 89 features. Some are simple: when the first cut lands, how long a shot lasts on average, how much silence there is, how fast I talk. Some come from the vectors: whether the frame changes, whether the face is close, whether the picture is aesthetically consistent.
I won’t give away the whole list, but the rule is simple. Every feature has to answer a question I ask myself while editing. Is something happening at the start? Am I sitting too long in one shot in the middle? Is the sound going flat? If a feature doesn’t translate into an editing decision, it’s out.
At the end sits CatBoost. Depth three, three hundred iterations, five models in an ensemble. Nothing fancy. A small model on a table of numbers that learns only from my own material. Everything runs locally, on a Mac, with no API. There’s a separate hub on the portal about local AI models.
Why so simple? Because with a few hundred examples a complicated model will learn things by heart. I’d rather have shallow trees that catch a few strong patterns than a deep network that memorizes that reels with a green sweater got a lot of views.
The material: my reels and three baskets
The model is only as good as what it learns from. I train it only on my own reels: on what I recorded, edited and published myself, along with the statistics from my account. I split them into three baskets: bronze, silver and gold. Gold are the ones that clearly stand out above the median of my profile. I don’t count bare views, because those also depend on the day, the hour and how many people happened to follow the account at the time.
This matters, so I’ll say it again: I compare a reel with the median from the same period. The question isn’t “is this viral at all”, it’s “did this reel do more than usual for me”. That way the model learns form, not the fact that the account grows month by month.
Preparing the material took me longer than the training itself. First the export of recordings and statistics, then the transcripts, then the vectors. With 32 GB of RAM you can’t hold all the models at once, so it goes one after another. The whole thing grinds away for a few hours. I start it in the evening and look in the morning, same as an overnight Blender render of my 3D films.
What it looks like in my week
So this doesn’t sound like a lab, I’ll show the normal run of one reel.
I record on my phone in 4K, in HDR. I sit by the window, in yellow glasses, with the microphone at my chin. Usually three takes of the hook, because the hook is half the success and the first take is almost always too slow. How I set up the hook, the cover and the end card in my 3D series reels, I described separately.
Then grading. Phone footage in HDR looks gray and washed out on Instagram, so I bring it down to SDR with a LUT, lift the contrast a little and sharpen. For me that’s a script too, not turning sliders by hand, because with ten reels a week turning sliders by hand is a wasted evening.
Next comes transcription with Whisper, cuts, captions, b-roll. Render. And only on the rendered file does the forecast get calculated. That matters: the model rates the finished reel, not the idea. The idea I rate myself.
I get the report in a few minutes. I read it like notes from an editor who has watched all my reels, remembers which ones worked and has a soft spot for fast cuts. With part of it I agree right away. With part I don’t, because sometimes a longer shot is on purpose, for example when I’m showing something on screen and the viewer needs time to read it. Then I leave it and note down that I knowingly went against the grain. After five days I see whether I was right.
How much time does it take? The report itself, almost none. The fixes after the report are usually a quarter of an hour of editing. For a reel with a chance of getting one and a half times more views, that’s a good trade.
The forecast is a curiosity, the list of fixes is the value
The forecast number is nice to put in a reel. But the real value is elsewhere. The model tells me what stands apart from my gold reels. In my v1 it came to five things:
- First cut in the third second. My gold reels usually cut before a second and a half.
- No change of picture at the transition between the hook and the content.
- A middle that’s too static: almost ten seconds in one shot.
- Shots in the middle that are too long, over eight seconds on average.
- A weaker-looking frame: a blown-out window in the background.
Each of these has a concrete fix. A cut before 1.3 seconds. B-roll in the seventh second. Shorter shots in the block about models. A tighter frame, less window. This isn’t “maybe worth considering”. It’s a to-do list for the edit.
The model doesn’t tell me what to record. It tells me where the viewer will probably drop off. And that I know how to fix.
And here’s the loop I like most. I make the fixes, render v2, the model runs the numbers again. I see whether the forecast moved. I iterate v1 into v2 on numbers, not on a hunch at two in the morning.
What are the limits of a model like this?
Four: RAM, a small sample at the start, the fact that the model only knows me, and the fact that it knows nothing about the topic.
Now, honestly. First, RAM. Big vision and text models don’t fit into 32 GB at once. It can be done, but you have to line them up in a queue and accept that it takes time.
Second, and this is more important: on a small sample the model looked better than it is. With the first few dozen reels the results were beautiful. Too beautiful. The model had simply memorized a few of my strongest reels. Only with more material, and with testing on reels it hadn’t seen, did it show how much it really knows.
The more material, the surer the result. That’s not a slogan, it’s arithmetic. That’s why every reel of mine goes back into the set after publishing: the forecast next to it, the real result after five days next to it. The model is supposed to be wrong less and less. If it isn’t improving, it means one of the features is lying.
Third: the model knows only me. That’s a strength and a limit at once. It learns my pace, my framing and my viewers, so its advice fits me, not someone with a million followers. But it won’t suggest a form I’ve never tried. New things I have to come up with myself, record a few times, and only then does the model have something to compare them with.
Fourth: the model knows nothing about the topic. A superbly edited reel about something nobody cares about still won’t fly. Form is half. The other half is the topic and whether you have something to say.
What I take from this
The most important change isn’t technical. I stopped treating publishing like a coin toss. I have a stage before posting where I can still fix something, and I have a number I can compare myself against a week later.
If you want something similar for yourself, don’t start with a model. Start with a table. Take 30 of your reels and write down three things for each: when the first cut lands, how long the longest shot is and what the result was against your median. You’ll see a pattern in that alone. The model comes later and just counts it faster.
One warning from me. Don’t edit for the model. The moment you see a number, the temptation shows up to cut every reel the same way, because “the model likes fast cuts”. A month later you have a feed where everything looks identical, and viewers feel that faster than any algorithm. The model is there to catch my mistakes, not to pick my style for me. That’s why the decision about what stays is always mine.
Right now I’m taking v2 of this reel back to the workbench. The first cut moved forward, the middle got b-roll. Let’s see what the model says. And let’s see what people say, because they have the last word.
I work with people 1:1 under the delta240mvt brand. On your project, from the first step.
See the offer ↗

