PhD Defense: Perceptual Quality Assessment and Enhancement of Visual Media using Generative Priors

  1. PhD Defense: Perceptual Quality Assessment and Enhancement of Visual Media using Generative Priors
  2. Candidate and committee members
  3. Dissertation roadmap
  4. Beyond8Bits and LGDM
  5. Beyond8Bits and BrightRate (review)
  6. LGDM turns a diffusion model into a zero-shot quality judge (review)
  7. Five contributions
  8. HDR-Q, a multimodal LLM for HDR video quality
  9. HDR is a different signal space
  10. Where current methods break
  11. Three obstacles
  12. HDR-Q architecture
  13. Adding the HDR input made GRPO worse
  14. The contrastive KL at the core of HAPO
  15. The HAPO objective
  16. HDR-Q reaches 0.921 SROCC on Beyond8Bits
  17. HDR-Q names real artifacts where the baseline invents them
  18. The contrastive KL matters most
  19. Takeaways
  20. LumaFlux: Structure-Aware SDR-to-HDR Conversion with Monotone Luminance Flows
  21. The question
  22. What gets lost, on three frames
  23. From the SDR input space to the HDR output space
  24. Prior work and where it fails
  25. A frozen FLUX transformer, steered by what the frame measures
  26. PGA is a LoRA told what it is looking at
  27. Scene context from SigLIP, and a monotone tone decoder
  28. A 314,396-pair corpus of user and professional video
  29. LumaFlux is best in every column of the primary track
  30. Three scenes, every method
  31. In the ablation, the tone decoder matters most
  32. Video without temporal training
  33. What holds up, and what falls short
  34. Takeaways
  35. Rectified-CFG++, geometry-aware flow guidance
  36. The question
  37. Why we need guidance at all
  38. Why CFG works on diffusion but breaks on flows
  39. Why CFG fails and how we fix it, geometrically
  40. Extrapolation versus interpolation
  41. The two algorithms side by side
  42. Theoretical guarantees
  43. CFG versus ours, property by property
  44. Watching the artifacts grow, or not
  45. Baseline, CFG and Rectified-CFG++ on three Flux prompts
  46. CFG versus ours on Flux-dev, eight pairs
  47. Text rendering, the stop sign test
  48. Results on MS-COCO 10K
  49. User study
  50. Sweeping the number of sampling steps
  51. Takeaways
  52. Summary of contributions
  53. Connecting the contributions
  54. Publications
  55. Questions?
  56. Thank you

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