Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Sep 2·5 min read
DiffHDR: Re-Exposing LDR Videos with Video Diffusion Models
TL;DR
DiffHDR is a novel framework that enhances low dynamic range (LDR) videos by converting them to high dynamic range (HDR) using video diffusion models. It significantly improves radiance fidelity and temporal stability, enabling realistic HDR video generation.
✦ Why It Matters
Implement DiffHDR in your video processing pipeline to enhance LDR videos for HDR displays.
Key Takeaways
- Utilize DiffHDR for improved HDR video quality.
- Explore Log-Gamma color space for better radiance recovery.
- Leverage text prompts for controllable video conversion.
- Generate HDR training data from static HDRI maps.
How It Works
DiffHDR operates by formulating the LDR-to-HDR conversion as a generative radiance inpainting task. It uses a pretrained video diffusion model to synthesize plausible HDR radiance in regions that are over- or underexposed.
By working in Log-Gamma color space, it effectively recovers continuous scene radiance from the quantized pixels, allowing for realistic HDR video generation.