OMATOOLS/OmaRAW guide

Guide / Masks and local adjustments

AI object masks

Select an object with points, a box or AI brush, then create editable paths.

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.

Where it is

Develop → Masks → AI object mask

Quick overview

  • Select an object with points, a box or AI brush. Use Paint for exact additions and erasing.
  • Refine and review the draft, then Create editable mask converts it into paths with ordinary local controls.
  • Inference runs locally. Inspect fine or ambiguous edges; the result is not a hair or transparency matte.

In this film

Read the transcript · Download captions

  1. AI estimates an object outlineThese optional models run locally; this demo uses a copied runtime.
  2. Start AI object maskThe first-use Download button installs tools, not your photographs.
  3. Click inside the cupA point is a hint to the model, not a painted dab.
  4. Inspect the blue selectionRight-click unwanted areas to exclude them; add points for missed parts.
  5. Try a box around the cupA close box can disambiguate an object.
  6. Switch to AI brushA short stroke inside the object asks AI to find its outline.
  7. Brush inside the cup, then releaseGreen is your guide; blue is the model’s resulting coverage.
  8. Switch to Paint for exact correctionsPaint adds exactly the brushed area; right-drag erases.
  9. Paint a small correction at the edgeA fresh AI prompt would recompute the selection and replace manual paint.
  10. Undo selection restores the previous draftUse it to compare prompts and manual changes.
  11. Refine edges before acceptingIt follows nearby colour edges; weak boundaries still need inspection.
  12. Create editable maskKeep the draft visible while conversion completes.
  13. Now it is an ordinary editable localShapes & overlay shows paths whose points you can revise.
  14. Use its local Colour controlsAI selected the object; you choose what adjustment it receives.
  15. Push gently towards a warmer colourThe overlay hides so you can judge the actual correction.
  16. Preview the accepted selection againInspect edges at 100%; this is not a fine-hair transparency matte.

What AI object masking does

AI proposes an object selection from hints you place on the photo. It does not change exposure or colour by itself. You review the selection, correct it if necessary, then choose Create editable mask. That converts the outline into ordinary paths in Masks, where all local Colour, Tone and Detail controls become available.

Use it to speed up selecting a cup, person, animal or other recognisable subject. It produces a binary object outline that can be feathered afterwards. Fine hair, motion blur, glass and transparency do not become a detailed opacity matte.

Setup and local processing

Open Develop → Masks → AI object mask. Current OmaRAW packages include the selection/removal models and their runtime. No separate first-use download is needed. The selection and removal calculations then run on this computer without an assistant account.

CPU works without a suitable GPU. Use GPU for selection enables optional Vulkan acceleration; integrated graphics are not always faster. This switch applies to selection, while removal runs on CPU. If the panel reports missing or damaged bundled files, reinstall the OmaRAW package and read its status message before retrying.

Select object: points and a box

  1. Start AI object mask, with Select object active.
  2. Click inside the intended object. Inspect the blue selection, especially where the subject touches similar colours.
  3. Add another positive click on a missed part. Right-click a mistakenly included region to exclude it.
  4. Alternatively, drag a box around the object when a point is ambiguous. Keep the box reasonably close to the subject.
  5. Refine with a few useful points. Repeated clicks on the same difficult edge may be less effective than a manual correction.

The points and box are prompts for a new model estimate, not painted coverage. A point outside the outline can change a much larger part of the result. Undo selection returns to the previous draft, making it useful for comparing two prompts.

AI brush

Choose AI brush, paint a short stroke inside the object and release. The model uses the stroke as an object hint and finds the outline. The green stroke is your guide; the blue overlay is the actual result. Right-drag over an unwanted region to provide an exclusion hint.

A short stroke well inside the intended subject is often clearer than tracing its entire boundary. Avoid crossing into the background unless it belongs to the selection. Brush size is a radius expressed as a percentage of the photo's long edge, so reduce it for narrow subjects and close detail.

Paint for exact draft corrections

Choose Paint when you want to add exactly what you brush. Left-drag adds coverage, right-drag erases. Paint does not ask AI to extend the mark to the object boundary. This is useful for a missed strap, a small gap or a stray background island.

Switching tools keeps the current draft. However, a new AI click, box or AI brush stroke recomputes the AI selection and replaces manual paint with that new result. A practical order is AI estimate → refine → manual Paint → accept. Undo selection can restore the previous draft if a fresh estimate loses useful corrections.

Refine, invert and clear

Action What it does When to use it
Refine edges Moves the draft outline towards nearby colour edges, using more detailed rendering when zoomed in Improve a coarse boundary before accepting it
Undo selection Restores the preceding selection state Compare a new prompt, paint stroke, inversion or refinement
Invert Selects the opposite area of the draft Start a background correction from a subject selection
Clear Removes the current selection so you can start again The hints have led to the wrong object
Cancel / Escape Ends the AI draft Leave without converting a selection

Refine edges can help where the foreground and background separate clearly. It cannot reliably decide ambiguous or transparent boundaries. Compare before and after with Undo selection and correct weak edges manually.

Create the editable mask

  1. Choose Create editable mask after reviewing coverage. The selection stays visible while conversion is running.
  2. On success, Shapes & overlay opens with the new local's editable points and coverage visible. This is the confirmation that the selection has become an edit.
  3. Move points or curve handles, add or subtract shapes, and set Feather as needed. Then open Colour, Tone or Detail and apply the correction.
  4. Hide the mask display and compare the local off/on. Check both the detailed boundary and the full photograph.

If conversion fails, the message beside the button explains the problem and keeps the draft. Refine or simplify the selection and try again. A busy button is not an instruction to keep clicking. If the accepted mask is visible but the photograph looks the same, its adjustment values may still be neutral.

A converted mask is saved with the photo's edits and supports Undo, history, snapshots, copying and export. It no longer requires the model to redraw that saved path. An unaccepted draft is temporary: cancel, changing the photo or changing its edit clears it. Finish or cancel the draft before adjusting other tools.

Example: a local colour push

Select the person, accept the mask, open Colour and push gently towards warmth. If you want only skin rather than clothing and hair, sample a Colour Range within the accepted local and inspect the intersection. AI identifies an object; colour range narrows it by pixel colour. Neither guarantees that every selected pixel is skin, so check hands, lips, clothing and similar-coloured details.

What the film records

The lesson runs the installed models locally on a credited sample photograph in a separate demonstration catalog. It shows real selection and conversion into paths. Processing waits are shortened; playback duration is not a speed benchmark. No real user catalog or photo is used.