Kinema-2 Motion generates the world. Kinema Faces keeps the people in it the same people. Below: what each model is, how we measure it, and the notes we wrote while building them.
A latent video diffusion transformer trained for sequences, not clips. Temporal attention spans the full 60-second window, and a physics prior keeps cloth, water and hair moving like they weigh something.
An identity encoder that conditions Motion on a character embedding, not a pasted face. It carries age, bone structure, wardrobe and gait across lenses, lighting and time of day — and refuses to lock a real face without a release.
Comparisons use the public KinoBench-24 suite: 240 prompts, 5 raters per shot, blind. Competitor names are withheld in the public card; they are available under NDA.
Mean rater score, 0–100, on 12-second shots with a moving camera.
Percentage of shots scoring ≥ 95 on the identity metric, as the sequence grows.
Median seconds on production hardware, October 2026. The 4K pass is optional.
Mean deviation in centimetres between the solved camera path and the path recovered from the rendered shot.
Most video models generate at 8, 12 or 24 frames per second and interpolate up. It looks fine on a phone. On a 40-inch grading monitor the seams show: hands that smear for two frames, rain that stutters, hair that briefly becomes a different hair. Editors call it “the swim”. Directors call it “no”.
Kinema-1 interpolated. We measured the swim with a simple metric — optical-flow discontinuity between consecutive frames — and found it accounted for 61% of the rater complaints on KinoBench. So Kinema-2 Motion was trained to emit 48 real frames per second from the start.
Every frame is a real frame. There is nothing between them to get wrong.
Doubling the frame rate and quadrupling the shot length means 2,880 frames in a single attention window. The naive cost is quadratic. We use a two-level scheme: dense attention within 2-second chunks, and a sparse set of 96 “anchor” latents that carry global state — lighting, camera pose, character embedding — across the whole shot.
Anchors are what make a note like “same shot, one hour later” cheap. The light anchor changes; the chunk latents are re-denoised from step 18 of 40 instead of from noise. That is the 20-second re-render you see in Studio.
Cloth, water and hair are where video models look most like video models. We added a lightweight auxiliary loss during training: a learned physics critic predicts whether a 0.5-second window of motion is plausible for the material it sees. It is small — 180M parameters — and it moved the “material realism” rater score from 71 to 88.
The critic is also why coats in Kinema have weight. It penalises fabric that moves faster than the body wearing it.
Honest numbers: 48 fps native is 2.1× the compute of 24 fps interpolated at equal resolution. The anchor scheme claws back about 40% of that. Net, an 8-second shot costs 1.3× what it did on Kinema-1 and takes 71 seconds instead of 55. We decided the swim was worth more than 16 seconds.
60 fps native is in training. Longer than 60 seconds is not — past a minute, our users cut, and Studio is built for cutting. The next note will be about Kinema Faces v4 and what “the same person” means across a ten-year age change.
Short, specific write-ups from the people training the models. We publish what we measure, including the misses.
Research roles in Lisbon and Los Angeles. Bring a benchmark you do not trust and tell us why.