Never Ask The Model To Fail

AI video fails predictably. Not often, not randomly. Predictably, in 4 specific places, on nearly every render that comes back wrong. That single fact changes where the work happens. A generation that melts a hand is not bad luck and it is not a bad seed. It is a shot that was written down badly, or a shot that should never have been written down at all. The repair belongs in the shot list. Not in the prompt, and not in post. 4 danger zones break in a fixed order of severity: hands first, faces second, skin third, and text with reflections and fine pattern fourth. The ordering is not cosmetic. It is a ranking of how visibly each zone fails. A waxy cheek reads as polish. A 6 fingered hand reads as a machine, and the audience stops believing the film in a single frame. The window is tight. 63 per cent of the videos that hold their audience hook inside 3 seconds, and 70 per cent intro retention is the bar that reads as strong. A hand that fails at second 2 lands inside the exact window the whole edit is built to win. So the discipline is simple to state. Look at every shot you have planned, and ask which of the 4 zones it leans on. If the answer is hands or a tight face, the shot needs redesigning before it needs rendering. The best AI film is the one that never asked the model to fail.

Hands break first because they are the most information dense moving object in the frame. Extra fingers, fused fingers, wrong counts. There are 27 bones in 1 human hand, 8 in the wrist, 5 in the palm and 14 in the fingers, and the model has to hold every 1 of them consistent across 2 frames at 24 per second. 3 fixes, in order of how much they cost you. First, keep hands out of tight focus. A hand that is soft, partial or at the edge of the frame is a hand the model does not have to finish. Second, describe the start and the end position, not the gesture. Write the hand at the door and then at the shoulder. What happens between those 2 points is where the model invents a motion path, and the motion path is where fingers fuse. Third, and this is the cheapest fix available, crop it. The back of the hand reads as a hand. A wrist or forearm crop reads as a hand. Neither one carries fingers, so neither one can carry the error. No fingers visible is the most useful 3 word phrase in the whole toolkit. It costs nothing, it removes an entire class of failure, and it changes the shot in no way an audience will notice. The shots to strike off the list entirely: hand close-ups, sign language, hands holding text or signs, precise gestures, fast hand motion. 5 categories, and every 1 of them is a shot the model cannot do. Rewrite it or lose it.

A face filling the frame forces the model to invent detail it does not have. That is the whole problem in 1 sentence, and it explains almost every uncanny face on screen. An extreme close-up asks for skin at a resolution the model cannot deliver, so it produces 1 of 2 outcomes. Waxy, or cratered. Medium framing is the safe zone, because it asks for less and gets all of it right, and the anchor that holds it is the lens. 85mm for the face, 35mm over the shoulder, 24mm for the room. A long static hold on a face is the second fault. Hold still for 5 seconds and the face drifts on its own. The jaw shifts. The nose widens. The features stop being 1 person. Face shots stay at 3 seconds or less, or they keep moving, and 3 seconds is generous. A full profile turn is the third. Rotation past 90 degrees smears the features, because the model has to reconstruct a side of the head it never rendered. Turn the body, not the head, or cut before the turn finishes. Then the crowd problem. Multiple faces in 1 frame means the background faces melt, because the pixel budget was spent on the hero. 1 hero face per frame is the limit, and everyone else is soft, turned or out of frame. Last, stacked emotions. Crying and smiling and grimacing inside 1 shot reads as morphing rather than feeling. Pick 1 emotion per shot. 6 rules, and each 1 removes a failure the audience can see.

Waxy skin is not a style the model prefers. It is the model suppressing noise, over smoothing the frame to hide what it cannot resolve. Which means the fix is the exact opposite of beautiful language. Every flattering word in the prompt pushes the render further into gloss, so the words that produce real skin are the words of an unretouched photograph. Natural skin texture. Visible pores. Unretouched. No makeup. The cues that carry it are biological rather than poetic, which is why they work. 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. Those are the signals a camera records. They are also the 1s a model otherwise smooths away, and naming them is what stops it. The trap is vocabulary. Beautiful, flawless, radiant and even photorealistic are the triggers that send a realism biased model into over correction, because each 1 of them is asking for a version of a face that has never existed on film. And when a render still comes back waxy, the response is not a retouch pass. It is a re-render with the biological cues written in. That is a 1 line fix against an hour of masking.

A bad render is not bad luck. It is a bad shot that was written down.

The 4th zone is the quiet 1, because it fails in ways that look like production rather than error. Text first. The model garbles lettering, and it garbles it confidently. Storefront signs, phone screens, book covers, tattoos, licence plates, screens in the background. Every 1 of those is a shot that will come back with lettering that is almost right, which is the worst outcome available, because almost right reads as a mistake rather than a design choice. Either remove the word or remove the frame. Reflections second. Mirrors, glass and water do not match the scene in front of them, so a mirror that shows a different room, or a glass that shows a different light, breaks the film quietly. Unless the mismatch is the point, keep the reflective surface out of the shot. Fine patterns third. Teeth, picket fences, brick walls, mesh and fabric weave all produce moire, because a repeating pattern at the size of 1 pixel inside a frame that is 1920 pixels wide is the hardest thing there is to resolve. Pull back, defocus, or change the angle until the pattern stops repeating across the frame. None of the 3 can be rescued later. Each 1 is decided at the moment the shot is written.

The shot-design rules are 6 and they fit on 1 line each. Face shots are medium framed, 3 seconds or less, movement present. Hand shots are wrist or back of hand, soft focus, no fingers visible. Walking and body shots are medium wide with the hands at the sides stated in the prompt. Object and wide shots are full creative freedom, because those are the safe shots. From behind and silhouette are safest of all, because a face that is not visible cannot drift. And any face, hand or body clip stays at 3.5 seconds or less, cutting to a new angle or an object rather than holding, which at 24 per second is a cap of 84 frames and a number you can hold in your head. The cut rhythm suits it. 19 shots inside the first 30 seconds is the benchmark the short form work already runs on, and a 3.5 second cap forces that rhythm anyway, which means the fix the model needs and the pace the platform rewards point in the same direction. Then the rule that overrides the tools. Post process cannot rescue a bad render. High-pass overlays, frequency separation, grain, none of it adds pores that were never in the pixels. The biology has to be in the prompt. 3 hours of retouching will not put a pore into a cheek that was smoothed at generation. Here is the arithmetic this article is built on. Built companies across 12 countries over 30 years. 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. Signed a 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 through a thing the machine could do well. Every 1 was built by designing around what was possible, in the order the constraints arrived, and refusing the version of the work that looked better on paper than it could ever look on screen. That is the same rule here. The probability that a shot survives to the final cut is close to 1 when it was designed around the 4 zones before rendering, and close to 0 when it was rendered first and rescued afterwards. The difference is not budget and it is not talent. It is asking the right question before the render button.