ML Denoising With Tensor Flow
summary
An AI denoiser built with TensorFlow Estimator and trained on noisy/render pairs (including albedo, normals, depth) uses a U‑Net style autoencoder with skip connections and ≈1.5 million parameters. Trained on NVIDIA Titan hardware during an AMD collaboration and shown at SIGGRAPH 2018, it cuts multi‑hour/frame path‑tracing convergence to near‑real‑time perceptual results, is quantizable for mobile, and uses a parallel tf.data pipeline and standard gradient optimizers to reach usable quality.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 3:41Noise in rendering
- 6:54AI denoising solution
- 7:14Neural network basics
- 11:11Autoencoders for denoising
- 13:40Skip connections
- 14:26TensorFlow components
- 16:18Training in TensorFlow
- 18:13Training process
- 19:39Data processing in TensorFlow
- 23:55Data loading optimization
- 25:01Hardware support
- 26:04Installing TensorFlow
- 27:38Q&A
related talks
26mHit-reco: ProtoDUNE Raw Data Denoising with DL models
01h 30mDeep Learning Techniques for Monocular Depth Estimation
29mHow AI boosted audio processing?
57mTensorflow, deep learning and modern convolutional neural nets, without a PhD
01h 07mDeep Learning with TensorFlow
42mWhat's New with TensorFlow?
20mTensorFlow.js Bringing Machine Learning to the Web and Beyond
54mDeep Learning in JS