Seven crafts, one standard · FinalGrade
FINALGRADE
Benchmarks

Seven crafts, one standard

Camera and Lighting launch first. Five more crafts are on the way, from production design to stunts, each graded by the people who do the work.

A film is made by departments, and each one has its own language. A production designer talks about period, palette and set dressing. A costume designer talks about silhouette, fabric and how a garment moves. A sound designer talks about room tone and perspective. A stunt coordinator talks about timing, contact and safety. Each of those vocabularies is precise, and each one is a test AI video will eventually have to pass.

FinalGrade is building a benchmark for seven departments on a set, in four groups:

Figure 1Seven crafts, three stages
Live
CameraPreliminary results
LightingPreliminary results
Grading now
Production DesignResults December 2026
Costume, Hair and MakeupResults December 2026
In development
Sound DesignRubric and panel forming
ScoreRubric and panel forming
StuntsRubric and panel forming
Every benchmark uses the same method: prompts in the craft's own terms, one dimension per question, working professionals grading blind.
Group Crafts Status
Cinematography Camera, Lighting Live: preliminary results
Design Production Design; Costume, Hair and Makeup In progress: results December 2026
Sound and Music Sound Design, Score In development
Action Stunts In development

One standard

Every benchmark follows the same rules:

  • Prompts are written in the craft's own terms, by people who practice it.
  • Each prompt tests a single dimension and holds everything else fixed.
  • Every dimension is tested twice, as a fresh shot and as a note on an existing clip.
  • Working professionals in that craft judge the results blind, side by side, ranking every model's clip and drawing the line between delivered and not.
  • The same prompts go to every model, with no tuning per model.

Because the method is the same, results are comparable across departments. A lab can see that its model is strong on camera movement and weak on costume continuity, measured the same way.

Why camera and lighting first

We started with camera and lighting because they are the controls filmmakers reach for first, and the ones where today's tools fall shortest. They are also the crafts where the vocabulary is most standardized: a "Rembrandt" or a "rack focus" means the same thing on every set in the world. That makes them the right place to prove the method.

What the design crafts test

Production design and costume, hair and makeup are grading now. Their rubrics cover period accuracy, palette, set dressing that fits the character, silhouette and fabric behavior, skin and hair realism, and continuity of wardrobe and makeup across shots. Continuity is already a recurring theme: our camera and lighting evaluators flag wardrobe and set changes as defects when models revise a shot.

Who evaluates each craft

Each benchmark is evaluated by working professionals in that craft, invited on their credits. Camera is evaluated by cinematographers and camera operators. Lighting is evaluated by lighting professionals and cinematographers. Production design is evaluated by production designers and art directors. Costume, hair and makeup is evaluated by costume designers and makeup and hair artists. Every panel follows the same rules: anonymous, paid, opt in to every round and blind to the model.

What a design direction looks like

Design directions follow the same pattern as camera and lighting: a plain scene and one specific instruction in the craft's terms. A production design direction might specify a period and a place, such as a kitchen in a 1970s Midwestern farmhouse, and the panel checks whether the props, finishes and set dressing are right for the period and consistent across shots. A costume direction might specify a character's wardrobe and how it should age, and the panel checks fabric behavior, silhouette and continuity from the wide to the close-up.

Sound, score and stunts

Sound design, score and stunts are in development. Sound and score depend on models that generate audio with video, which is arriving now. Stunts test physical plausibility, timing and contact, some of the hardest problems in video generation. Each will open when its rubric and panel are ready, not before.

If you work at the top of one of these crafts and want to help shape its benchmark, join a panel.

Keep reading