One Base Model And A Fleet Of Looks

A 50 mm prime on the desk, 1 base checkpoint open in the interface, and 4 looks on the drive by the end of the afternoon. That is the shape of the whole thing, and it fits on 1 line because the training is not the hard part. The base model arrives already trained on a broad world. Every face, every material, every hour of light is in it already, and it came out of somebody else's compute bill. You are not building a model. You are adding 1 look to a model that works. That is what a small adapter is. It trains on top of the base and it holds 1 look and nothing else. Photorealism in 1 adapter. A drawn register in a second. Whatever look you have a use case for in a third. The base stays 1 file and the fleet grows around it. The economics are the whole argument. Training from nothing costs what the first base cost. Training on top of it costs the small run. The gap between those 2 numbers is the reason a solo operator can hold several different looks instead of 1. 1 base model plus a handful of cheap adapters is a fleet of sellable style-specific models. Small effort, wide spread.

Each adapter is a separate file, and a separate file is a separate asset with its own name, its own reference sheet, its own listing and its own buyer. This is the part most people skip. They train 1 good look, use it on 20 renders, and never put it in the shop window. 1 look is a trick. A catalogue of 20 named looks is a business, because the buyer is not shopping for a model. The buyer is shopping for the exact look they already described to somebody and could not produce. So the file gets 3 things before it goes anywhere. A name that says what it does. A short reference sheet with the images that define it. A credit line that holds every render made with it, so the asset has proof attached rather than a promise. Name the look by what the frame shows and never by what the run contained. Nobody buys a run. They buy the frame it produces. 1 look is a trick. 20 named looks is a catalogue.

The order is 4 steps and only the first 1 is expensive. Step 1, the base. Step 2, the small run that carries the look. Step 3, capture flat. Step 4, the grade in post. Reverse any 1 of those and the pipeline charges you for it later. The training set itself is small because the base carries the world and the adapter only carries the look. A tight set that holds 1 register beats a large set that holds 3, and it beats it every time, because a mixed set teaches the adapter to average. An averaged look is mud, and it reads as mud in the render. The cheapest way to lose the run is to feed it 3200 Kelvin interiors and 5600 Kelvin exteriors in the same folder and ask it for a look. It will give you the average of the 2 rooms and call it a style. Hold 1 register in the set. The base already knows the rest of the world.

Keep the footage flat in capture and grade in post. This is the 1 rule that keeps the fleet honest, and it survives because the 2 things are not the same thing. A look is a decision you make before the render. A grade is a decision you make after it, on a file that still holds the latitude to be pushed. The grade is what sells the realism. Realism comes out of the grade, never out of the adapter. A look burned into the capture cannot be revisited. The grade can. A 10-bit flat frame still carries the 3200 Kelvin interior and the 5600 Kelvin window as 2 separate decisions when the grade opens, and 1 of them can be pulled while the other holds. Bake the look in and both numbers are gone. You get 1 frame, and it is the frame the look decided. So the order matters more than the taste does. Base, then the small run, then flat capture, then the grade. The grade is the last decision, never the first.

One base, then the fleet. The base is the same base for everybody.

4 failure modes cover most of the wreckage, and each 1 names its own repair. The look sold twice. Two adapters trained from the same folder and listed under 2 names produce near-identical renders, and the second listing is a refund waiting to happen. Keep 1 folder per look and 1 look per listing. The run trained on a mixed register. Warm interiors and cool exteriors in the same set teach an average, and the average has no audience. Split the set by register and train 2 adapters. The look baked into the capture. It looks wonderful on the day and it cannot be re-lifted 3 weeks later when the client wants the same shot at dawn. The latitude was spent at the moment of capture. The library with no names. 20 files called final-look, final-look-2 and final-look-good are not an asset base. They are a folder, and a folder cannot be listed, priced or handed to anybody. Diagnose before you retrain. Retraining without a diagnosis is a re-roll with more steps in it.

3 checks before a look is filed, and none of them is a glance. Read the render back against the base and confirm the register held, because a look that drifts into the neighbouring register is 2 looks fighting over 1 file. Then confirm the capture is flat, which means the shadows still hold detail and no colour decision has been spent early. Then confirm the grade survives a second open, because a grade that only works once is not a grade, it is a guess. Here is the arithmetic behind all of it. Built companies across 12 countries. Restructured a industrial group. Deployed 210 energy systems across Africa and Asia. Took a company through a Frankfurt IPO. 16 years of field work from rural Nigeria to Fukushima. Entrade built 200 wood-gasification machines across the UK and Europe. Signed a personal guarantee on a company that folded on Christmas Day 2008, and lost savings in the 2013 Cyprus bail-in. 393 pages in paperback and Kindle, published 8 April 2026. None of it arrived as 1 look. All of it arrived as 20 small runs nobody watched. The probability that a single look carries an entire body of work is close to 0. The probability that a library of small looks does is close to 1. Most people price the first at 100 and the second at 0, which is why they arrive at the listing with 1 file and no buyers. One base, then the fleet. The base is the same base for everybody.