AI Denoise

AI Denoise

AI denoising uses machine learning models trained on millions of image pairs to remove noise from renders. Instead of using mathematical filters like traditional denoisers, AI denoisers have learned what clean images look like and can reconstruct the clean image from a noisy input with remarkable accuracy. The results are often much better than traditional denoising, especially for very noisy images.

The training process works by taking clean, high sample count renders and adding artificial noise to create noisy versions. The AI model learns to predict the clean image from the noisy one. Over millions of examples, the model develops an understanding of what kinds of patterns are real detail and what kinds are noise. This allows it to remove noise while preserving fine details that traditional filters would blur away.

Nvidia's OptiX denoiser is the most widely used AI denoiser in VFX. It runs on Nvidia GPUs and can denoise a frame in milliseconds. OptiX uses the beauty pass along with auxiliary passes like normals and albedo to guide the denoising. The results are impressive, often producing clean images from renders with very few samples. Many renderers have built in OptiX support.

AI denoising is not perfect. It can sometimes hallucinate detail that was not in the original render, creating patterns that look plausible but are not actually correct. This is called AI artifact. For final frames, it is safer to render enough samples that the AI denoiser only needs to do light cleanup. For look development and previews, AI denoising is a game changer because it lets you see a clean result almost instantly.

The future of denoising is AI based. As models get better and hardware gets faster, AI denoisers will become the standard for both preview and final frame rendering. Some renderers are already experimenting with neural rendering, where the AI does not just denoise but actually generates parts of the image. For now, AI denoising is an essential tool that every VFX artist should understand and use.