HDR-Q: Quality-Aware HDR Video Assessment via Multimodal LLMs

  1. HDR-Q: Quality-Aware HDR Video Assessment via Multimodal LLMs
  2. Can AI see and reason about HDR video quality?
  3. SDR encoders cannot see what HDR adds
  4. Where current methods break
  5. Three obstacles, and the third surprised us
  6. Beyond8Bits, the data HDR-Q trains on Saini et al., CVPR 2026
  7. HDR-Q has one part per obstacle
  8. The encoder separates HDR from SDR
  9. Adding HDR input made standard GRPO worse
  10. HAPO pushes HDR and SDR outputs apart
  11. HAPO adds three terms to GRPO
  12. 0.921 SROCC on Beyond8Bits, and it holds zero-shot
  13. The baseline invents artifacts, HDR-Q names real ones
  14. Contrastive KL is the component that matters most
  15. What we learned, what is still open
  16. Takeaways

← → or swipe to move, click the slide to advance, F for full screen. Add #12 to the address to link to slide 12.