Codex catches anatomy-placement error in Sora 2 reference-frame QA

• Kai Gani said he had been using Sora 2 castings to generate stronger reference material for character cameos.
• Gani's comparison post said Codex beat Claude and Gemini at spotting an anatomy and placement mistake in the generated clip.
• The winning check was still-image QA, not full video understanding: Gani clarified that GPT 5.5 analyzed contact sheets and selected frames.
• The attached Codex output flagged four error classes: generative artifacts, vehicle geometry drift, malformed plate text, and road reflection continuity.

Gani's Sora 2 casting note is the creator workflow clue: use the video model to stage references, then run a separate model over frames as a quality-control pass. His follow-up is the constraint: the model did not listen to audio or track motion directly. It inferred problems from a contact sheet, then zoomed in on individual frames.

Gani described Sora 2 castings as a way to get strong references for character cameos. That is the useful part for filmmakers and AI artists: the generation step becomes a casting board, not only a finished clip.

The post does not describe the full prompt stack, seed process, or editing pass. It only establishes the workflow direction: generate believable character reference material first, then use that material downstream.

Gani's test compared Codex against Claude and Gemini on one concrete job: find the visible mistake in a generated video. He said only Codex caught the anatomy and placement error that a human viewer could see.

The attached frame shows a woman leaning across a parked station wagon in a desert-road shot. The reported failure was physical coherence: her body and arms appear to intersect with the car structure instead of sitting in a plausible pose.

Codex's written output grouped the clip problems into four buckets:

1. Generative visual artifacts: implausible body position, incoherent support, arms intersecting with the hood or window area.
2. Vehicle continuity and geometry: grille, headlights, hood ornament, antenna, roof rack, side panels, and windshield relationships shifting across the shot.
3. Malformed readable text: the license plate rendering as warped, nonsensical text.
4. Road and reflection continuity: lane markings and road geometry sliding across the hood in ways that do not match reflection or vehicle motion.

For creators, that list is more useful than a pass or fail score. It turns a vibe problem into a shot-review checklist.

Gani's clarification narrowed the claim: the system was using GPT 5.5 as image processing, not native video analysis. It looked at a contact sheet of multiple frames, then inspected specific frames.

That distinction changes the workflow. The QA step can catch anatomy, geometry, text, and frame-to-frame visual consistency issues visible in stills, but Gani said it could not understand audio or motion except by extrapolating from individual frames.

The source clip post carried the original linked media that Gani used as the object of the QA comparison. The clip circulated widely enough to give the frame-level test a shared visual target rather than a private prompt result.

The evidence pool does not include the prompt, model settings, or the full contact sheet. What it does show is a reproducible pattern: generate a reference clip, export representative frames, ask a vision-capable model to identify physical and continuity breaks, then use the error categories as revision notes.