Introducing FinalGrade-Camera
13 dimensions, from shot size to lens character. One example prompt for each, written the way a DP would say it.
FinalGrade-Camera asks one question: can a video model deliver the shot a cinematographer asks for?
It is built from the vocabulary cinematographers use on set. Each prompt isolates one of 13 dimensions, so a failure points to a specific control rather than a vague sense that a shot looks off.
The 13 dimensions
The full rubric covers everything a DP specifies about the camera: what the frame contains, where the lens sits, how it behaves and how it moves.
| Dimension | What the panel checks | Example direction |
|---|---|---|
| Shot size | The frame matches the size asked for, from extreme wide to insert | Medium close-up, chest up |
| Framing and composition | The subject is placed with intent: headroom, lead room, balance | Frame her on the left third, looking into the space |
| Camera height | The lens sits at the requested height | Camera at seated eye level |
| Camera angle | Tilt and angle are right: high, low, Dutch, overhead | Low angle, looking up at him |
| Focal length | Perspective and compression match the lens | Shoot it on a 24mm, close to her |
| Depth of field | Focus range is as shallow or deep as asked | Shallow, only the eyes sharp |
| Focus and focus pulls | Focus is on the right subject and pulls land on time | Rack from the glass to her face on the line |
| Camera movement | The move is the one asked for: push, pull, pan, track, crane, orbit | Slow dolly in |
| Movement speed and easing | The move starts, travels and settles at the right pace | Ease in, then hold |
| Stabilization | Handheld feels handheld; locked-off stays locked | Locked off, no drift |
| Aspect ratio | The image is composed for the ratio | 2.39:1, use the width |
| Motion blur and shutter | Blur matches a cinema shutter | 180-degree shutter, natural blur on the pan |
| Lens character | The image carries the lens's look | Anamorphic, oval bokeh, horizontal flares |
| Dimension | Dailies | Notes |
|---|---|---|
| Shot size | ||
| Framing and composition | ||
| Camera height | ||
| Camera angle | ||
| Focal length | ||
| Depth of field | ||
| Focus and focus pulls | ||
| Camera movement | ||
| Movement speed and easing | ||
| Stabilization | ||
| Aspect ratio | ||
| Motion blur and shutter | ||
| Lens character |
How a camera question is built
Each question has a simple scene, one direction and three yes/no checks that define what "delivered" means. For a tracking shot of a swimmer, the checks are that the camera travels alongside her at water level, that the foreground water passes faster than the background, and that her distance to camera stays roughly constant. A clip can look beautiful and fail all three.
Every question is tested two ways: as a fresh shot from a written setup (Dailies), and as a note on an existing clip (Notes), where everything except the note must hold. Seven models take the fresh-shot test. Five support revisions and take both.
What round 1 found
The first round covered 20 camera questions across movement, stabilization, depth of field and focus pulls, evaluated blind by working cinematographers and camera operators.
Three patterns stand out.
Simple moves are close to solved. Precise ones aren't. A slow pull back to reveal a room was delivered in 95% of graded clips. A push in given as a revision, "slow push in on him and rest on a frame where he fills it," was delivered in none. The difference is precision: the pull back only has to happen, while the push in has to start, travel and land on a specific frame.
Focus is the weakest area on revisions. Asked to throw the background out of focus, rack to an object or bring the whole office into deep focus on an existing clip, models mostly re-rendered the shot instead of moving the focus plane. Deep focus as a revision was delivered in none of the graded clips.
Rankings change by control. One model led on focus and depth of field and finished last on camera movement. Another led on movement. A single camera score hides this, which is why the results page breaks the leaderboard out by control and by test type.
Hardest and easiest directions
The extremes of round 1 show where the line between understood and not understood currently sits.
| Direction | Test | Delivered |
|---|---|---|
| Slow pull back from him to reveal the study | Fresh shot | 95% |
| Rack focus from the near player to the far player as he moves a piece | Fresh shot | 67% |
| Very shallow focus, the lights behind him fall to large soft discs | Fresh shot | 57% |
| Pan with him | Revision | 20% |
| Go wide open: keep him sharp, let the glass in front and the bar behind fall off | Revision | 20% |
| Shift focus to the shelves behind him, let him go soft | Revision | 13% |
| Deep focus: bring the office behind her into focus too | Revision | 0% |
| Slow push in on him and rest on a frame where he fills it | Revision | 0% |
Every one of the hardest directions is a revision. Models that can produce a convincing rack focus from scratch often can't add one to a shot that already exists.
Reading the camera leaderboard
On fresh shots, Kling led camera with a score of 69.6, ahead of LTX and Gemini Omni. On revisions, the order changed completely: Wan led at 60.2, followed by Seedance, with Kling fourth. Within camera, the leaders also differ by control: one model was first on focus and depth of field and last on camera movement.
That is why the results page shows the leaderboard four ways: overall, fresh shots, revisions and by control. For a team choosing a model, the right view depends on the work. If you generate shots from scratch, read the fresh-shot board. If you iterate on shots you already have, read the revisions board.
What comes next
Later rounds add the remaining nine dimensions, starting with shot size, focal length and camera height, and more evaluators per question. Every round is graded the same way, so a lab can compare one release to the next.
Preliminary results are live on the FinalGrade-Camera page.