Design Around The Model

Every render that comes back wrong was predictable before it was rendered. Not in principle. In writing. The model has a short list of things it cannot do, and the list does not change between jobs, between clients or between renderers. Waxy skin, a plastic sheen, the CGI look, a deformed hand, a morphing face, drift between 2 renders of the same subject. 6 failures, and they repeat. The negative block is where that list is kept. Tier 1 holds 44 terms and stays attached to every still and every video. Tier 2 holds 17 body and anatomy terms for nude and sexual work. Between the 2 tiers sits almost the whole visible failure surface of a generated frame. The first mistake is treating the block as a settings file. A settings file is something you tune. This is closer to a shot specification, because it names the tells before they arrive rather than describing the mood you hope to get. That reorders the work. The block is written before the shot is framed, and the shot is then framed so it never leans on anything the block lists. 1 decision, made 20 minutes early, removes an entire class of failure from the shoot. The alternative is the expensive route. Render it, watch it, diagnose it, rewrite it, render it again. 4 passes where 1 would have held, and the diagnosis lands on the same 6 failures every time. The probability that a shot survives to the final cut is close to 1 when it was written against the block, and close to 0 when it was written first and corrected afterwards. The block is cheaper than the correction by an order of magnitude.

The 44 global terms are not a jumble. They break into 7 families, and each family is a different kind of failure. Skin, the first family, 6 terms. Glossy, airbrushed, plastic skin, waxy skin, over-smooth, beauty filter. This is the family that makes a real face read as a mannequin. Renderer artefacts, 8 terms. CGI, 3D render, digital painting, uncanny, doll-like, wax figure, cartoon, anime. Every 1 of these describes the render announcing itself. Text and marks, 4 terms. Text, watermark, logo, caption. When lettering renders almost correctly it reads worse than lettering that fails outright, because almost right reads as a mistake rather than a choice. Motion, 4 terms. Jittery, fast motion, motion blur, warped. Motion blur is the quiet 1. It looks like craft and it hides a model that could not hold the move. Anatomy, 11 terms. Extra limbs, extra fingers, fused fingers, six fingers, deformed hands, extra digits, mutated hands, morphing face, identity drift, cloned face, cloned subject. 11 terms, and this is the family your audience notices in a single frame. Capture quality, 4 terms. Out of focus, blurry, low quality, oversaturated. Each 1 is a request to the renderer to stop pretending a phone snapshot is a camera. Composition and gaze, 7 terms. Halo, vignette, looking at the camera, looking at the lens, fourth wall break, direct address, eye contact with the lens. That last family is where 90 per cent of amateur generated work gives itself away. Not 1 of the 44 is a repair. All 44 are a design brief. Read them as a list of shots you will not write, and the block stops being a filter and starts being a plan.

Waxy skin is not a style the renderer prefers. It is over-correction, the model suppressing surface noise it cannot resolve. Which means the words that cause it are the flattering ones. 6 triggers account for most of it. Beautiful. Glowing skin. Flawless. Perfect. Radiant. Photorealistic. Add 2 more from the over-pumped end of the prompt stack, hyper-detailed skin and a 4K face close-up, and you have written the exact instruction that sends a realism biased model into gloss. None of those 6 words says anything false. That is why they are dangerous. Each 1 asks for a version of a face that has never existed on film. The repair is 9 words on the positive side. Documentary film. 35mm. Natural skin texture. Visible pores. Unretouched. Natural light. Negative fill. Slow deliberate movement. Single subject. 9 words, and they pull the render back toward the texture of a photograph. Negative fill deserves its own line, because it is the 1 that most people skip. It means removing light rather than adding it. A cheek with a shadow on 1 side has structure. A cheek lit from 3 directions has none, and no amount of texture language will put it back. Then the rule that closes the loop. Post process cannot rescue a bad render. 3 hours of retouching will not put a pore into a cheek that was smoothed at generation, because the information was never in the pixels. So when a face comes back waxy, the answer is not a retouch pass. It is a re-render with the 9 defaults and the 7 cues written in. That is a 1 line fix against an afternoon of masking.

On a face, the language that works is biological rather than poetic. 7 cues carry almost all of it. Visible pores. Fine peach fuzz catching the rim light. Capillary redness on the cheeks and the nose. A slight oil sheen on the T zone only. Fine laugh lines at the corners of the 2 eyes. Uneven melanin. No makeup. They work because they are what a camera records and what a model otherwise smooths away. Asking for 7 of them by name gives the renderer detail it would rather average out. The word only is doing real work in that list. Oil sheen on the T zone only is skin. Oil sheen everywhere is a wet face, and the renderer will deliver the wet face if you leave the scoping out. No makeup is the 1 that reads as a risk and is not. It removes the layer the model uses to hide smoothing. An unretouched photograph has texture that a beauty default erases, and the erasure is the tell. Uneven melanin is deliberate. Skin that is 1 flat tone across the whole face is the fastest machine read there is, on any face, in any light. 7 cues, 6 words of prompt each at most. Written into the first pass, they cost about 20 seconds. Added after the render, they cost the render.

The negative block is not a settings file. It is the specification.

Everything above is prompt craft. This section is not, and it is the part that actually holds. 1 rule runs underneath the whole block. Design the shot around what the model cannot do. Do not prompt your way out of a shot the renderer cannot hold, because the prompt is not the constraint. The constraint is the shot. Applied, it reads like this. If a shot needs a face, it stays medium framed and short. If it needs a hand, it is the wrist or the back of the hand, with no fingers visible in the frame. If a passing hand would be a nice detail, the better answer is a shot where no hand appears at all. A wrist and a back of the hand each read as a hand. Neither carries fingers. So neither can carry the error that fingers arrive with. The cost of that rule is 0. You lose a finger close-up that would have failed anyway. You keep a shot that renders on the first pass, and the audience never notices the substitution, because the audience reads intention and not anatomy. This is also why the shot list is the real instrument and the negative block is the supporting 1. The block tells the renderer what to avoid. The shot list means there is nothing to avoid in the first place. Nothing in post adds pores that were never in the pixels. Nothing in post restores a hand that was never fully rendered. Both decisions were already made by the time the frame existed.

The arithmetic behind this chapter is 30 years of building, in 12 countries, under 2 rules: work with what the constraint allows, and refuse the version of the work that looks better on paper than it can look in the room. Built companies across 12 countries over 30 years. Restructured a EUR 75 million 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. Signed a EUR 20 million personal guarantee on a company that folded on Christmas Day 2008. In the building for the 2003 blackout. 250 operators trained and deployed. 393 pages in paperback and Kindle, published 8 April 2026. Not 1 of those was built by doing a thing the machine could do well. Every 1 was built by designing around what was possible, in the order the constraints arrived, and turning down the plan that required a capability nobody had. That is the same rule applied to a render. The block names the failures. The shot list removes them. The frame rule means they never enter the shot. What stays is the 1 habit that transfers. Before the render button, read the shot and ask which family it leans on. Skin, face, hands, text, motion. If the answer is hands or a tight face, the shot gets redesigned and not rendered. Probability first, budget second. A shot written against the block survives to the final cut close to 100 per cent of the time. A shot rendered first and rescued afterwards survives close to 0 per cent of the time, and it costs 4 passes to learn that. 44 terms and 1 rule are cheaper than 4 passes.