Four Of Six Is Not A Pass

The ask was 1 image. A candid night-street frame of a woman outside a club, no interviewer in shot, real skin, unposed, and the brief named the register precisely: a genuine frame from a real viral street interview, not an AI-generated fashion portrait. 19 candidates came back across 5 prompt iterations and every 1 was scored against the same 6 checks. The best frame of the run passes 4 of them. The 2 it fails are the 2 the brief cares about most. It does not read as an unposed amateur phone frame. It reads posed and cinematic. And the background does not stay legible: it dissolves into circular bokeh. 1 finding landed before any frame was rendered and it is not about renderers at all. The submitted reference image is candid footage of a real, identifiable person, and photoreal generation of that person, or of a fictional character styled to match her hair, wardrobe, jewellery and street, is likeness replication, which does not route here. The genre is shareable and the genre is what was built. Nothing was carried across from that photograph. Not its hair. Not its top. Not the styling. Not its street. The 19 frames are an authored in-house woman, and the honest result of the whole run is 4 of 6 with 2 checks still open.

The 5 iterations read as a ladder rather than a spread of attempts. The straight brief, anchored to the project's own woman, scored 1 of 6. Mechanism language about gaze and mouth, with the negatives folded into the prompt, took it to 3. Adding an explicit instruction not to smooth her skin, plus an inverted bokeh instruction, held it at 3. The change that moved the number was structural: drop the subject anchor and demand deep focus. That produced the first 4 of the run. The reason is checkable. The anchor plate is itself a polished close-up with creamy bokeh, so every prompt carrying it handed the renderer a reference for the exact look the brief rejects. The 5th iteration added an anti-pose action and a point-source light description, and it held the 4 rather than raising it. Then a local candid grade was applied to that best frame, dropped it to 3 of 6, and was reverted. 5 iterations, 19 frames, 2 changes that mattered, and the same 2 checks open at the end.

Continuity and the phone register are in tension on this lane, and here the register won. The frame that finally reads as real skin is of a new original synthetic woman rather than the woman in the delivered films, because the plate that carried continuity was also carrying the portrait grammar the brief is trying to leave behind. 1 upload round-trip normally buys identity across a cut. On this lane it bought the wrong register, which is why dropping it was the only move that lifted a score. Check C on the scorecard is real skin texture, pores, uneven tone, and it is the first genuine pass of the run. It arrived the moment the anchor came out, and it did not arrive in the 3 iterations that kept it. The lesson is not that anchors are bad. It is that an anchor is a decision with a look attached, and the look travels into every prompt that carries the file. 1 file, 1 look, every prompt.

2 mechanical findings came with the run and both are about the tool rather than the brief. The image lane has no negative-prompt field. The negative block written into iteration 1 was silently discarded, so every candidate in that batch rendered with no negative guidance at all. No error. No warning. No line in the log. 1 batch of frames was never actually testing the prompt it was given. The prompt is also capped under 1500 characters, which means the negatives have to be squeezed inside the prompt itself as an avoid tail. Both facts are now written down as checks instead of hopes from the scorecard of 2026-09-24: confirm the field exists before writing into it, confirm the whole prompt fits under the cap, and read the number back out of the tool rather than assuming it.

19 candidates across 5 prompt iterations. The best frame of the run scores 4 of 6.

The scorecard on the best frame is honest at the ceiling rather than rounded up. Pupils clearly away from the lens: pass. Mouth closed, lips together, no teeth: pass. Real skin texture, pores, uneven tone: pass. The subject is the only person in frame, no interviewer, no microphone, no hand: pass. Then the 2 that matter. It reads as an unposed amateur phone frame: fail, because it reads posed and cinematic. Deep focus with a legible background: fail, because the background dissolves into circular bokeh. Across 5 differently-worded iterations and 19 frames, the renderer returned a posed subject against circular bokeh every time. Asking for no bokeh is read as a style word. Asking for deep focus changes how the subject is rendered and not what the background does. So the 2 open checks are the model's portrait grammar rather than the prompt's phrasing, which is exactly why 5 rewrites moved 4 checks and none of the 2 the brief cares about. A local grade cannot recover it either, because the bokeh is real blur in the pixels, and the studio's own grade pass lowered the frame from 4 of 6 to 3 of 6 and was reverted. This matches a finding already recorded on the studio's night-interview brief: image models carry a film-look prior and resist phone cues. 1 model is recorded as obeying the phone register, and it carries no credits. The other lane reads zero on all 5 accounts. So the ceiling of this run is real and it is reported as 4 of 6.

3 routes close this and only 1 of them is a prompt. Fund a lane that obeys the phone register, which is 1 named model on 1 funded account and 2 lanes that are unfunded today. Or shoot it, because the brief is physically producible in 1 evening: 1 consenting adult model, 1 black outfit, 1 club doorway and 1 phone. Or take the frame from video, where a mid-motion frame is inherently less posed than an image model's output, untested on this brief and worth 1 render. Everything else from the run is kept rather than deleted: 19 candidates, every iteration folder, and the graded variant documented as kept and not delivered because it scored lower. That is the shape the studio runs on, work built across 12 countries, 35 staff and 200 plus interns a year, 393 pages in the book and 450 plus generated articles and films behind the pipeline, and every order passing a named check before it ships. The probability that 19 frames across 5 wordings all fail the same 2 checks by accident is low. They fail because the renderer has a prior and the prompt was never the variable. Write the brief. Fix the checks before the first render. Then pick the lane that can actually meet them.