OMATOOLS/OmaRAW guide

Guide / Retouch and AI cleanup

AI denoise explained

Understand the RAW preview, strength and separate editable DNG result.

Read the full reference ↓ · All tools

On this page · full reference

Film recorded with an earlier version; the written reference is current. AI models and runtimes are now included; any first-use download shown or mentioned in the recording no longer applies. This lesson explains the controls and workflow. See the recording notes below for exactly what is demonstrated.

Where it is

Develop → Prepare → AI denoise

Quick overview

  • Supported Bayer and X-Trans RAWs select the model automatically. Preview Strength at 100% on texture and smooth areas.
  • Apply creates a separate full-size linear DNG, keeps the RAW and copies colour, crop and masks as editable settings.
  • Denoise strength is baked into the copy. The film explains the real controls using JPEG samples; it does not show a RAW denoise completing.

In this film

Read the transcript · Download captions

  1. AI denoise needs a supported RAWThis JPEG sample shows the controls, not a finished denoise result.
  2. Open AI denoise in PrepareAn optional download supplies local models and tools.
  3. Strength balances noise against textureAround 60% is a starting trial; lower values leave more grain.
  4. The sensor selects the modelBayer and X-Trans are detected from the RAW, not just its brand.
  5. Use GPU when faster checks the hardwareIt can fall back to CPU; untick it to force CPU.
  6. ISO override is for missing Bayer metadataUse the actual capture ISO; otherwise leave it blank.
  7. Preview a detailed and a smooth areaOn a RAW, Preview denoise opens the detail comparison.
  8. Compare at the same position and zoom100% reveals lost texture as well as remaining noise.
  9. Global colour and exposure are includedCrop, masks, retouch and other detail effects are excluded from this preview.
  10. Apply creates a separate full-size DNGThe original RAW stays unchanged and is stacked with the copy.
  11. The new DNG keeps editable adjustmentsExposure, colour, crop and masks can still be changed.
  12. Denoise strength is baked into that copyReturn to the RAW for a different strength or demosaic result.
  13. Save copy as chooses a name and locationAllow space for a large floating-point DNG and temporary files.
  14. Use denoise before object removalContinue editing the denoised copy and judge ordinary sharpening again.

Watch the current tool demonstration for the complete workflow in the current interface.

What AI denoise is for

For a concise control-by-control lookup, see Prepare → AI denoise.

AI denoise uses a model suited to the sensor pattern to reduce noise while converting a supported RAW into a full-size linear DNG. Use it when noise in a high-ISO or strongly lifted photograph needs more help than ordinary Noise reduction. The models and runtimes are included in current OmaRAW packages and work offline after installation.

The source RAW remains unchanged. The denoise strength and demosaicing are baked into the new DNG, while exposure, colour, crop, masks and retouch settings are copied as editable adjustments. Return to the RAW to choose a different denoise strength. A new DNG is not a backup of the original sensor mosaic.

A complete working sequence

  1. Open a supported RAW in Develop. Use Prepare → AI denoise. If the panel reports missing or damaged bundled files, reinstall the OmaRAW package.
  2. Begin around 60% Strength. This is a trial starting point, not a requirement. Lower values retain more noise and can preserve a more natural texture.
  3. Choose Preview denoise. Inspect a textured area and a smooth shadow. Drag on the preview to move through the photo; it refreshes after you stop.
  4. Compare Before / After at the same position and zoom. Use 100% for texture and Fit for the whole preview area. Scroll zooms under the pointer, double-click switches Fit/100%, and Ctrl+scroll steps zoom presets.
  5. Check skin, hair, fabric and fine lettering, not only a smooth wall. If they look smeared, lower Strength and preview again.
  6. Choose Apply AI denoise for an automatically named copy beside the source, or Save copy as… for another name or destination. Existing names are not overwritten.
  7. The app adds the DNG to the catalog, stacks it with the source and opens it. Continue editing that copy. Reassess ordinary noise reduction and sharpening instead of applying strong extra smoothing automatically.

What the comparison includes

The preview includes current global colour and exposure, including Auto, so you can judge noise under the intended brightness. It excludes crop, local masks, retouch and other detail effects. Those exclusions avoid confusing the denoising comparison with separate effects; they do not mean the settings are lost from the saved copy.

After changing global colour or exposure, update the denoise preview. Compare the same region at the same scale; a fitted view hides both noise and lost detail.

Sensor detection, ISO and GPU

Bayer and X-Trans sensor patterns select the appropriate model automatically. Selection is based on the sensor, not merely the camera brand. The detected sensor appears in the panel. If a Bayer file lacks ISO metadata, enter the actual capture ISO in the override field; otherwise leave it blank. X-Trans does not need that override.

Use GPU when faster checks Vulkan against the CPU and can fall back. Untick it to force CPU. Full-resolution processing can require substantial memory and time; large X-Trans files can take particularly long on CPU. Preview first and cancel if necessary.

Already developed linear DNGs, monochrome sources and some DNG calibration maps are unsupported. A JPEG tutorial photo cannot demonstrate RAW denoising. The tool reports unsupported input rather than converting it into an equivalent sensor RAW.

Disk space and versions

Float32 DNGs are large, approximately 12 MB per megapixel before considering any storage compression, and processing needs temporary space too. Keep room for the source, temporary work and final DNG. Previewing is useful before committing to a batch of large copies.

If you also want AI object removal, denoise first. The original and derived versions can carry independent edits. Select the version you intend to export and keep both when you want to revise the denoise decision later.

Recording scope

This film explains the real panel and the workflow using the pack's JPEG sample collection. It does not show a RAW denoise finishing or present JPEG smoothing as a model result. The supported-input, preview and output details below come from the application's current denoise reference.

Local AI denoise

Open Prepare → AI denoise in Develop. The runtime and both sensor models are included in the package, ready for offline use. Denoising runs on your computer; photographs are not uploaded and no online account is needed.

Preview and apply

Choose the strength and preview a detail area before applying it. Drag the preview to another part of the photo, use the wheel to zoom, and compare Before and After at the same position. The preview includes global colour and exposure settings, but crop, local masks, retouch and other detail effects are excluded. Update the comparison after changing colour edits.

Apply AI denoise creates an unused -denoised.dng file beside the original, adds it to the catalog, stacks it with the original and opens it for editing. Save copy as… lets you choose a different filename. Existing files are not overwritten.

The original stays intact. Current exposure, colour, profiles, crop, masks and retouch transfer to the new copy as editable settings. Demosaicing and denoise strength are baked into its pixels; make another copy from the original to change those. Sensor-only settings remain with the original.

Support and performance

Photo versions

Photo versions above the image keeps the original, intermediate results and latest AI result together. Choose Open original or Open latest AI version to move between them; each retains its own edits. Apply denoise before object removal for the normal RAW workflow.

Sources and licences

The included Bayer TreeNet and X-Trans Restormer models come from RawForge's ONNX release, distributed by the MIT-licensed upstream project. The package build checks pinned hashes and converts the models to float32 before installation. Original and converted hashes, conversion details and the upstream project licence are recorded in the third-party notices and model manifest.

RawForge and RawHandler code is MIT-licensed. The adapted Malvar filters come from colour-demosaicing under BSD-3-Clause. Full attribution and licence text are in the denoise notices, installed with OmaRAW as AI-DENOISE.txt. The processing implementation and model manifest are included in the application's source.

See AI tools for masks and object removal, or the command-line guide to use denoising through an agent.