Watch the full tutorial

What is Neural Restore in Darktable 5.6?

Darktable 5.6 introduces Neural Restore, an AI-powered denoising system that works fundamentally differently from the traditional denoise modules. Rather than operating within the processing pipeline, it functions as a plugin that generates a brand new DNG file — clean from the sensor data up. If you’re coming from Lightroom, think of it as the equivalent of Lightroom’s AI Denoise button, which also generates a new DNG.

This makes it particularly powerful for heavily noisy images — think ISO 10,000 and above — where standard denoise modules struggle.

Three Modes: Which One to Use?

Neural Restore offers three modes, found on the left panel in both Lighttable and Darkroom:

Raw Denoise — works at the very beginning of the pipeline, directly on your raw sensor data. This is the recommended starting point for new images. It generates a clean DNG that you then process normally with your exposure, AgX, cropping and other modules.

Denoise — works at the end of the pipeline, on an already processed image. Useful if you want to apply AI denoising to an image you edited months ago without redoing all your work. It exports a TIFF file rather than a DNG.

Upscale — covered in a dedicated tutorial, watch it here:

For most workflows, start with Raw Denoise on new images. Use Denoise only to salvage already-processed files.

Setup: Activating the AI Models

Before using Neural Restore, you need to activate it in Darktable’s settings:

  1. Go to Settings → scroll to the AI section near the bottom
  2. Download the Raw Denoise model (and optionally the Denoise model)
  3. Enable the models you downloaded
  4. Activate AI features globally

Mac users: Darktable can struggle when left to choose its own acceleration. Manually select Apple Core ML — it works reliably for both AI masking and Neural Restore, and processing is now very fast (a few seconds at most).

Using Raw Denoise: Step by Step

  1. Open your image in Darkroom
  2. Find the Neural Restore plugin on the left panel
  3. Click Generate Preview — Darktable renders a small preview area rather than the full image
  4. Move the peeker to inspect different areas of the image (if it stops responding, restart Darktable — a known bug that will be fixed)
  5. Adjust the Strength slider (0–100%)
  6. Set your output folder for the new DNG (default is same folder as original)
  7. Click Process — a new DNG appears in your Lighttable in about 5 seconds

Finding the Right Strength

There is no universal number — it depends on your image and your taste. Here’s a practical guide:

  • 100% — maximum denoising, but risks a plastic, detail-losing look especially on organic textures
  • 80% — a good starting point, strong noise reduction with acceptable detail retention
  • 60–70% — better detail preservation if 80% feels too aggressive

The classic denoising trade-off applies here too: the more noise you remove, the more detail you risk losing. Generate a few test DNG files at different strengths and compare them directly in Lighttable — it’s easier than judging from the small preview.

One key advantage of Raw Denoise: because it works directly on sensor data before any processing, you can later lift shadows aggressively without noise creeping back in — something that’s much harder to achieve when denoising later in the pipeline.

Darktable AI Denoise vs Lightroom AI Denoise: My Take

Both tools produce excellent results at high ISO. Lightroom’s AI Denoise is polished and well integrated. But Neural Restore in Darktable 5.6 is genuinely competitive — and it’s completely free, open source, runs locally, and requires no subscription.

The Darktable team themselves recommend a sensible workflow: use the classic denoise modules first for moderately noisy images, and reach for Neural Restore only when standard denoising isn’t enough. That’s good advice — Neural Restore is a precision tool for difficult files, not a replacement for the full denoise workflow.

The model is still improving, and the speed improvements since early 2026 previews are significant. It’s worth experimenting with on your own images to find what works for your shooting style.