The Twelve Shot Failures Name The Shot List

Skin is where AI exposes itself first, and it fails in 2 opposite directions at once. The first failure is waxy skin. The model over-smooths to suppress noise, deleting pores and micro-texture until the subject reads as a mannequin. The cause is aggressive denoising, plus prompts that push photorealistic, beautiful, perfect, flawless, or a blanket cinematic film still with film grain. The second failure is the opposite of the first. Demand hyper-detailed 4K skin texture on a tight face and the model hallucinates pores, veins and freckles that were never there. An ultra-realistic 4K face close-up is the worst offender on the list, and the reason is not taste. It is that the prompt asked for detail the source pixels do not contain. 3 moves cure both. Keep faces at medium to wide framing rather than extreme close-up, because the middle distance is the safest pixel budget in the frame. Use documentary language: natural skin texture, visible pores, fine laugh lines, unretouched, natural light, negative fill. And drop photorealistic, dropping cinematic with it as a blanket adjective. Never ask for more skin detail. Specify restraint. 1 of the 2 failures comes from asking for too little texture, the other from asking for too much, and the prompt cannot tell the difference until the render is paid for.

The face owns the largest block of the 12 named failures, and they split into 2 families: what the frame asks for, and how long the frame holds. Extreme close-ups are a pixel budget problem. A face filling the frame forces the model to invent detail it does not have, which is why waxy and cratered skin are both at their worst there. Pull the same face back to the 35 mm distance and the model has real detail to work with instead of invented detail. Long static holds are a memory problem. The model reinterprets the character every frame, because it has no memory of the frame it generated a moment ago, so a held face drifts: jawline shifts, the nose widens, the eyes go uneven. Faces dissolve over time. Keep face shots 1 to 3 seconds, or keep them moving. Full profile turns smear features and deform eyes. A slight head turn is fine. A 90 degree rotation is not. Talking, chewing and eating close-ups produce the classic AI mouth, because the mouth and jaw deform under motion. Background faces melt, so 1 hero face per frame and no crowds of faces behind it. Stacked emotional extremes close the family. Crying and smiling and grimacing in 1 prompt overloads the face and forces it to morph between states. Pick 1 state and hold it for the whole shot.

Hands break first and most visibly, because they are the most information-dense moving object in the frame. 1 of the 12 named failures lives here, and it is the 1 an audience notices before it can say what is wrong. Extra fingers, fused fingers, wrong counts. No renderer on the list solves this, so the shot list has to solve it before the shot is written. 3 rules cover the whole class. Keep hands out of tight focus, obscure them, or put them in slow motion. Describe start and end positions and never the gesture, because the model invents the motion path in between and that invented path is where the fingers fuse. Then drop the shot types that carry the failure: hand close-ups, sign language, a hand holding a sign, precise gestures and fast hand motion. The fix is subtraction rather than correction. Every hand shot that survives the cut is 1 the model was never asked to solve.

The model renders text as shapes rather than glyphs, so lettering comes back garbled, and the longer or smaller the string the worse it reads. 6 shot types carry that failure: storefront signs, book covers, phone screens, documents, hands holding signs, and words carried on tattoos or clothing. Watermarks belong on the same list. If text has to appear at all, keep it to 1 or 2 short words, large, and low-motion. There is no third option in which a paragraph reads. Fine repeating patterns fail differently and for the same underlying reason: the model hallucinates the pattern into moire. Teeth come back too many and too even, like a shattered piano keyboard. Picket fences, venetian blinds, checkerboards and brick walls at certain angles melt the same way. Pull the camera back or defocus the pattern and the failure stops where the pattern stops being legible.

Skin fails in 2 opposite directions at once: too smooth, and too detailed. Neither 1 is fixed in the grade. Both are fixed in the framing.

Motion, physics and reflection carry the last 3 of the 12. Fast or sudden motion produces motion blur artifacts, warping and smearing. Slow deliberate movement is always the safer shot, and it is the shot the audience remembers anyway. Combined camera and subject motion doubles the warp. Move 1 or the other, never both in the same shot. Slow motion without enough source frames produces judder and smear, which is why the render is generated at native speed and slowed in post, and only when the source is clean enough to carry it. Mirrors, glass and water break on the physics of reflection itself: the reflection does not match the scene it is reflecting. That is not a grade problem, so the shot gets rebuilt instead of rescued. 5 renderers were checked across this list in 2026 and the failures clustered in the same places in all 5.

1 question gets asked before any prompt is written: does this shot ask the model to do something it reliably fails at? If the answer is yes, the shot is redesigned. Prompting around a known failure is the expensive way to learn the same lesson twice. The render-side backstop is the negative prompt, and it runs in 2 tiers. Tier 1 covers every render: waxy and plastic skin, CGI, 3D render, beauty filter, uncanny, extra fingers, fused digits, morphing face, identity drift, text, watermark, jitter and fast motion. Tier 2 is appended on top of it for nude or sexual shots and for any shot where the body is the subject, and it locks anatomy and proportions: deformed anatomy, mismatched proportions, mutated limbs, asymmetric face, elongated neck, distorted torso. The positive defaults that keep a render real are a short fixed string: documentary film, 35 mm, natural skin texture, visible pores, unretouched, natural light, negative fill, slow deliberate movement, single subject. 35 mm is not nostalgia. It is the framing discipline that keeps a face at the distance where the pixel budget holds. 12 named failures sit on this list and every 1 of them has 1 cause and 1 shot-list answer. Short face clips, medium framing, slow motion, hands out of focus, 1 face per frame, no text, no reflections. The arithmetic is the argument. The probability that a first-pass render ships clean without the shot list designed around the failure class and the 2 negative tiers on every render sits near 1 in 10. The probability that the same shot ships when the shot asked for is 1 the model can render, the face is held at 35 mm and the negatives are on sits near 9 in 10. The same discipline built companies across twelve countries: restructured a industrial group, deployed 210 energy systems across Africa and Asia, built 200 wood-gasification machines across the UK and Europe. 1 pattern holds in all of it. Name the failure before it is rendered, and the second render never has to be paid for. 12 named failures, 1 question, 1 shot list. The best AI film is the one that never asks the model to fail.