A00:00:04:04MAJORSkin Under Mixed Light
Which AI video model holds believable skin under mixed light?
FINAL EYE · AI video, judged by a cinematographer.
Independent model tests and pre-delivery QA for studios that generate a lot of AI video. Every defect timecoded, judged by a cinematographer and colorist.
Updated 19 Nov 2026
A00:00:04:04 · f100MAJOR

Which one are you?
Prompt study on your brief, $349: your scene written as a DOP-level prompt and tested on 3–4 models, so the model is chosen before hundreds of episodes depend on it. For a full production: model selection from $1,200.
Brief a studyAI video QA from $35 per clip: every defect timecoded with location and severity, caught before the client sees it.
Book a reviewFootage study, $349: your frames run through 3–4 models, to see which one keeps your look, your light and your faces.
Send your footageQuick model check, $39: your prompt on 3 models, with scores, ranking and the top 3 defects per clip.
Check my promptStudies run on your brief and your material. Public proof so far: Test01, skin under mixed light. New tests join the archive as they’re judged.
Why not just a machine
Leaderboards count votes; software scores how pleasing a frame looks. Neither says why a shot fails: a lamp that lights nothing, a shadow too big for the raindrops casting it, a reflection sharper than its object. Every clip is also checked against broadcast standards: legal levels, frame rate, resolution, colour space.
Latest tests
Test01 P1 from FINAL EYE: five AI video models, one hard close-up under mixed light, judged on a fixed 9-axis rubric with blind letters.
A00:00:04:04MAJORWhich AI video model holds believable skin under mixed light?

Test01 P2–P4 · in progressHow it's judged
Every test compares AI video models, or versions of the same model, on the same prompt. The formats change from test to test; the rubric and the blind judging stay fixed.
Prompt adherence /100 plus craft axes, set before any clip is seen.
Clips are judged as A, B, C… The model names are revealed only after scoring.
Every defect is timecoded and located in the FINAL EYE comparison app.
In Test01 P1 the app's automatic scan missed two slight camera pans; the human eye caught them.
Tested via Higgsfield, first generation, no retries.
Where it helps
FINAL EYE tests AI video the way footage is judged on a film set. That pays off when:
You're about to run many generations: picking the right model first means credits go to one that can actually do the shot.
A clip is about to go to a client: problems get caught before the client sees them, so a missed defect is far less likely to cost your company the contract.
You have several takes or models to choose from: each one checked frame by frame, with every issue timecoded, so it's clear which one holds up.
A new model comes out: you see whether it's worth switching before spending on it.
You want to judge on your own: every report shows what to look for and where, so you can compare your next clips yourself.
Services · after the proof
Per clip: timecoded defects with location and severity, and what to regenerate.
The verdict, the scorecard and the major defects per model, with the frames: see what a FINAL EYE review catches before paying for one.
Get the free report1 clip, up to 60 s
Up to 4 clips, each up to 60 s
Pilot of 8 clips, each up to 30 s
Before you generate
A wrong model costs days of regenerations. The study shows which model handles your scene before the schedule depends on it.
Book a reviewFinishing on a locked cut, from $95, on request.
Free report
The Test01 P1 report, sent by email when you subscribe to the FINAL EYE newsletter. 10 pages: verdict, scorecard and the major defects per model.
FAQ
Most teams rely on whoever generated the clip, which is the person least likely to see its flaws. FINAL EYE is an independent pre-delivery review, read the way a production reads footage: by a cinematographer and colorist checking light, skin, lens look, motion and continuity frame by frame. Every defect is logged with its timecode, its location in the frame and its severity, together with what to regenerate. The question behind every review is simple: would this shot survive a client, a big screen and a colorist?
Per clip: timecoded defects with location and severity, and what to regenerate. Each defect is tied to one of nine axes: prompt adherence, reference fidelity, visual integrity, light and colour, motion and physics, temporal stability, camera and optics, performance, and audio. The client's brief is scored once, on adherence; every other axis judges only what is on screen, so the same miss is never counted twice. Every finding points to a frame that can be checked, so nothing has to be taken on trust.
Every clip is checked against broadcast delivery standards: legal video levels, frame rate, resolution and colour space. An in-house tool measures exposure and colour on an RGB parade and a BT.709 vectorscope, so "underexposed" or "off-hue" is a measured value, not an impression. An automatic scan flags flicker, frozen or repeated frames and camera drift. In Test01 it caught one delivery error directly: Wan 3.0 delivered 30 fps where 24 was asked, a mismatch that shows up as stutter once the clip lands in a 24 fps timeline.
$35 for one clip up to 60 s, $120 for up to 4 clips (each up to 60 s), and $250 for a pilot of 8 clips up to 30 s. Extra clips after any package cost $25 (up to 30 s) or $30 (up to 60 s). Before generating, a quick model check runs a prompt on 3 models for $39; a prompt study or footage study on a brief costs $349; model selection for a full production starts at $1,200. Turnaround is agreed per request.
In FINAL EYE's blind test of 19 Nov 2026 (Test01 P1: one close-up of a woman under mixed light, a warm 2700K lamp against cyan neon through wet glass), Gemini Omni Flash 1.2 rendered the most believable skin, with execution at 7.42/10. Wan 3.0 came second (6.67), holding up except for plastic highlights and a slight orange cast. Kling 4.0 and MiniMax H3 went plastic, and Seedance 2.5 came out over-softened. No model reached the 85mm lens look the prompt asked for, and none matched the 4:1 contrast ratio: measured values ranged from about 1.8:1 to 80:1. Full results: /tests/test01-p1/
Yes. No model developer sponsors, pays for or reviews FINAL EYE tests, and rankings are never sold. Clips are judged blind as letters A, B, C…, and model names are revealed only after scoring, on a rubric fixed before each test. Clips are generated through Higgsfield; any affiliate link is disclosed at the top of the page where it appears and never placed next to a ranking.
A single-clip review ($35) or a quick model check ($39) is paid on Ko-fi; the clip or prompt then goes to hello@finaleyestudio.com, a reply follows within one working day, and turnaround is agreed per request once the material matches the item ordered. For larger projects, the brief and clips go to hello@finaleyestudio.com first. Within one working day, a reply confirms the scope, the fixed price and the delivery date, with a secure payment link, and work starts once payment clears. Every report arrives by email or as a private download link: a branded PDF with every defect timecoded and what to regenerate first.
Fixed price, agreed before any work starts. Small orders are paid on Ko-fi; projects from $120 are paid by PayPal invoice, by card or PayPal, with no PayPal account needed. Full payment upfront; projects from $1,200 are split 50/50.
No work starts until the material matches what was ordered: for example, a clip longer than 60 s on a single-clip review. The choice is then the buyer's: pay the difference, send material that fits, or take a full refund.