The Model Breaks In Known Places

The failures are not random. They cluster in a fixed set of 6 places, and every cluster has a known cause: skin, faces, hands, text, motion and repeated fine patterns. That single fact changes what a shot list is for. You do not fix waxy skin in the grade, because the grade never made it. You do not shoot the shot that produces it. The probability that a row asks the model to do something it reliably fails at can be read off the page before a single frame exists, which means the list can be audited at 0 cost and fixed at 0 cost. Run that 1 question against every row and the document gets shorter, not longer. The 180 line holds screen direction across every cut and the negative prompt holds the look, and both are written before anything is rendered. 6 families, 6 causes, 1 question per row. The same 6 families show up across the 5 renderers in the 2026 study behind this list: Wan, Veo, Sora, Kling and Runway. Everything below is the same argument from 6 directions: the failure is mechanical, it is predictable, and it is cheaper to design around than to repair.

Skin is where the model shows itself first, and it fails in 2 opposite directions. Push it toward smooth and it over-smooths, deleting pores and micro-texture until the subject reads as a mannequin under a beauty filter. Push it toward detail and it invents: demand a hyper-detailed 4K skin texture on a tight face and the model hallucinates pores, veins and freckles that were never in the room. Both directions come from the same prompt habit, and the habit is the word photorealistic, or flawless, or a blanket cinematic film still with 4K gloss on top. The fix is restraint. Faces sit at 35mm or 50mm middle distance, not at the extreme close-up, and the language goes documentary: natural skin texture, visible pores, fine laugh lines, unretouched, natural light, negative fill across 1 cheek. The 85mm frame is where a face is held, and 85mm is also the restraint: the glass isolates without demanding 4K detail the model cannot see. 24mm and 35mm carry the room, 50mm carries the body, 85mm carries the face, and 3 lens groups are never mixed on the same subject inside 1 scene. Ask for less and the face arrives. Ask for 4K detail and you get a 4K invention.

The 2026 generation of models has no memory of the frame it drew a moment ago. It reinterprets the character every time, so a held face drifts: the jawline shifts, the nose widens, the 2 eyes stop matching each other. Faces dissolve over time. The fix is arithmetic. Face shots run 1 to 3 seconds, or they keep moving, and the move is slow enough that the 3 seconds read as a hold rather than a pan. Everything renders at 24 fps, so a hold that is short in seconds is still 48 frames long, and the drift compounds across every one of them. 1 hero face per frame is the ceiling, because background faces melt into each other and 4 faces in a crowd become 1 face and 3 smears. A head turn past 90 degrees smears the features and deforms the eyes, so the turn stops at 60 or 70 degrees and never reaches the full profile. Talking, chewing and eating close-ups deform the mouth, which is the oldest tell in the field. Stack 3 emotional states into 1 prompt and the face morphs between them. Pick 1 state and hold it.

Hands are the most information-dense moving object in the frame, so they break first and they break visibly: extra fingers, fused fingers, and counts that are simply wrong. The fix is a writing rule rather than a rendering rule. The list names the start position and the end position, and leaves everything between them unspecified, because the path is where the model invents the extra digit. Hands stay out of tight focus, or they are obscured, or they move slowly enough that a 24mm or 35mm frame at middle distance never resolves them. The material that breaks most reliably is the material that asks for precision: hand close-ups, signing, a hand holding a sign, a precise gesture, fast hand motion. Describe 2 positions and let the middle be approximate. The audience reads intent from the shoulder anyway.

The failures cluster. That is the whole opportunity.

Text garbles in proportion to its length, which makes it a design decision rather than a hope. Signage, book covers, phone screens, documents, a hand holding a sign, a tattoo, a shirt with 4 words on it, a watermark: all of it leaves the frame. Where text is unavoidable it is 1 or 2 short words inside a 16:9 frame, large in the frame, and low motion, and even then it is a risk. Reflections carry their own separate failure. A mirror, a window, a wet street and a water surface each have to agree with a room the model cannot see, so the reflection and the reflection's subject drift apart inside the same shot. Repeated fine patterns are the last cluster: teeth too even and too many, venetian blinds, a picket fence, a checkerboard, brick at certain angles. Pull the camera back to 24mm or 35mm, or defocus the pattern at 85mm, and the moire stops before the shutter does.

Slow and deliberate is always the safer register. Combined camera and subject motion doubles the warp, so 1 of the 2 moves and the other stays still. Slow motion without enough source frames judders and smears, so the render happens at native 24 fps and the slow is taken in post, and only when the source is clean enough to survive it. Fast and sudden motion smears the frame in ways no retime repairs. This is why the list carries 2 global strings on every row: 43 negative terms that name the failure in the model's own vocabulary, and 9 anchors that name what the frame should be instead, pasted verbatim rather than retyped. The 43 terms are pasted and never re-typed. A negative prompt rewritten per shot drifts, and the drift is what puts a wax face in shot 7 that was not in shot 4. EX Venture runs the same shape on the operational side, from 16 years of field work: 35 staff, 200 plus interns a year, work built across 12 countries, 393 pages in the book published 8 April 2026, and 450 plus generated articles and films behind the pipeline, with every order passing a named check before it ships. Set the framing, hold the lens, name the negatives, keep 1 thing moving. The best AI film never asks the model to fail.