AI-Driven SuperResolution: Using AMD ROCm, FSR & Real-ESRGAN

Reviving retro games and old videos with AI! Using Real-ESRGAN, AMD ROCm, and FSR, we restore and upscale footage to HD with GPU-accelerated speed on Radeon RX 9070 XT and Ryzen 9 9900X.

Comment

Inspiration

I’ve always loved classic games and old video footage, but seeing them blurry and grainy on modern screens felt wrong. I wanted to restore them-not just upscale them-and prove that AMD hardware could power real-time AI video enhancement.

What it does

AI-Driven SuperResolution revives low-quality footage by sharpening details, restoring textures, and upscaling every frame to pristine HD. It blends Real-ESRGAN, AMD ROCm, and FidelityFX Super Resolution (FSR) to achieve GPU-accelerated, real-time restoration.

How we built it

The pipeline runs on a Radeon RX 9070 XT + Ryzen 9 9900X setup.

  • Real-ESRGAN reconstructs and enhances details frame-by-frame.
  • ROCm enables GPU acceleration through PyTorch.
  • FSR (via FFmpeg + libplacebo) applies the final polish for sharpness and latency-free scaling.

I optimized tile sizes, memory usage, and filter integration to make the model run efficiently on consumer AMD GPUs.

Challenges we ran into

ROCm support for newer RDNA3 cards was tricky-driver mismatches and packaging conflicts slowed setup. Real-ESRGAN needed tuning to avoid artifacts, and integrating FSR through FFmpeg required custom filter chaining. Each step demanded debugging, profiling, and careful optimization.

Accomplishments that we're proud of

I achieved 4× perceptual upscaling with consistent detail recovery and low latency-proving that consumer AMD GPUs can deliver research-grade results in real time. The final demo restored classic clips and retro games with impressive clarity and smoothness.

What we learned

Hardware-aware optimization is key. Balancing model complexity with memory tiling, precision, and throughput turns a research model into a deployable product. I also deepened my understanding of GPU compute, ROCm internals, and real-time media pipelines.

What's next for AI-Driven SuperResolution

I plan to extend it with temporal consistency (frame-to-frame coherence) and motion-aware restoration, enabling smoother video and even better artifact suppression. The long-term goal is an open-source, plug-and-play AMD-accelerated video restoration suite for creators and archivists.

Built With

  • computervision
  • deeplearning
  • fsr
  • python
  • pytoch
  • rocm
Share this project:

Updates