Calibration of Generative Models for Cinema-Grade Output

Mechanism

Color, motion, and frame-rate fidelity targets for video-generative models intended to share a delivery pipeline with traditionally captured cinema. Diffusion-class and flow-matching-class video generators trained on web-scraped datasets exhibit color-space drift (sRGB or Rec. 709 priors leaking into PQ HDR output), motion-vector inconsistency (sub-pixel trajectory noise that defeats encoder MV prediction), and frame-rate quantization artifacts (judder against 24p and 48p cadence) that pass through the standard consumer encoding pipeline but fail under cinema-grade evaluation. This research line characterizes those drift modes and proposes calibration objectives that align generator output with cinema-grade encode targets, the same targets the DLC dual-layer compression encoder optimizes against on the delivery side.

Why this matters

• Generative models that share a delivery pipeline with cinema content must satisfy the encode pipeline's PQ HDR color volume and motion-vector tolerances. Consumer encode chains will silently drop quality signals (clip out-of-gamut, smooth over MV noise, resample to 60p) that cinema-grade encode chains will not. Without explicit calibration, the resulting bitstream is either uneconomic or visibly degraded at the consumer endpoint.
• The work is co-located with TsugiCinema's filed work on the cinema delivery pipeline (Trinity decode and DLC dual-layer compression). Calibration targets come from the encode and decode specifications directly rather than being reverse-engineered from sample content, which is the standard external-research posture and which has known failure modes around metadata and reference-display assumptions.
• Status: Research. Filing decision deferred pending characterization of drift modes and cross-model generalization. The expected output is a public preprint and an evaluation dataset, not a near-term provisional.

Status and what is next

Characterization of drift modes is in progress on a small set of open-weights diffusion-class video generators. Cross-model evaluation across diffusion-class and flow-matching-class generators is planned. Honest disclosure: this work has not been validated against a cinema-grade test bed at the time of writing. The cinema-grade test bed is the same Trinity-plus-DLC pipeline that is filed and operating on the delivery side, but co-locating it with the generator characterization rig is a separate engineering pass that has not yet started. Expected output is a preprint and dataset release within the next twelve to eighteen months. A filing decision will follow that release if the calibration mechanism is non-obvious over the public preprint.