Lightricks' LumiVid Turns Ordinary Video Into Cinema HDR With 300 Clips
Lightricks and Tel Aviv University adapt a pretrained video diffusion model to output cinema-grade HDR by aligning log-encoded footage with existing latents.
ProTakeaways
- LumiVid adapts a pretrained SDR video diffusion model to generate scene-linear HDR video with only a LoRA.
- LogC3 cinematic encoding aligns HDR distributions with the frozen VAE, avoiding encoder retraining.
- Camera-mimicking degradations during training force the model to hallucinate clipped highlights and shadows from priors.
- Achieves 30.05 dB PU21-PSNR and JOD 8.22 on UPIQ, beating HDRTVNet, Lediff, and X2HDR.
- Outputs float16 EXR scene-linear files ready for professional grading; supports SDR-to-HDR and text-to-HDR modes.
- Single-image variant LumiPic is open on Hugging Face; project page at hdr-lumivid.github.io.
Turning a standard 8-bit clip into cinema-grade HDR footage usually means training a bespoke model from scratch on scarce paired data. A team from Lightricks AI and Tel Aviv University just showed you can skip that entirely. Their new method, LumiVid, takes an off-the-shelf video diffusion model and teaches it to emit true scene-linear HDR video with a LoRA that touches less than 1% of the parameters and roughly 300 training clips.
No new architecture is involved. The move is a representation choice borrowed from film cameras that happens to line up almost perfectly with what generative video models already understand.
Why HDR breaks generative models
High dynamic range imagery offers a rich and faithful representation of scene radiance, but remains challenging for generative models due to its mismatch with the bounded, perceptually compressed data on which they are trained. Every large video diffusion model you can name was trained on standard dynamic range footage: 8-bit values squeezed into a [0, 1] range and gamma-corrected for display. HDR works the other way, storing raw scene radiance as unbounded float values, the kind of numbers a colorist needs when re-exposing a shot or compositing VFX plates.
Feed that unbounded signal into a pretrained VAE and things fall apart. The encoder was never asked to compress a sun 10,000 times brighter than a shadow. Paired SDR-HDR data is also rare, and most available datasets provide display-referred HDR rather than the raw scene-linear values needed for generative modeling. Distribution mismatch plus tiny datasets has kept HDR generation stuck in bespoke, from-scratch territory.
The LogC3 alignment trick
LumiVid's core observation is that the film industry already solved half of this problem. LogC3 is ARRI's logarithmic encoding curve used inside Alexa cinema cameras to squeeze wide-latitude sensor data into a manageable range while preserving highlight and shadow detail. The researchers noticed that LogC3, already an industry standard for ARRI Alexa cameras, maps HDR images into a distribution that naturally aligns with the latent space of pretrained video diffusion models.