RAW and AI model preferences

These pages of File › Preferences… choose how RAW files are decoded and which models run for masks and search. The defaults suit most photos. For how the dialog works, see Change preferences.

RAW & Color

Setting	Default	What it does
Demosaic	AAHD	How the sensor’s mosaic becomes RGB. The photo reloads when you change it.

Demosaic choices:

Choice	Strength
AAHD	Adaptive AHD. The best detail on fine texture, with few zipper artifacts.
DCB	Often cleaner through smooth gradients and foliage; can soften the finest detail.
AHD	Classic adaptive homogeneity-directed interpolation. The fastest of the three.

Denoise works on the RAW data before demosaic, where sensor noise is still easiest to separate from detail. Its model is chosen under AI Models, below. See Editing for RAW Process.

AI Models

The page lists the models you added, then which model each AI feature uses.

Your models

Some features can run a model you supply as well as the one that comes with DarkRay. Click Add Model…, or drop a file or folder on the page, and point at either:

• an ONNX file (.onnx; a large one has a data file beside it, which must stay in the same folder), or
• a downloaded model folder: its weights as .safetensors beside the config.json that names its type.

DarkRay checks the model as soon as you choose it: that it opens, which feature it fits, and whether it fits your graphics card. For a depth model it also runs it once on a reference photo to find which way it points, and sets Invert direction for a model where far comes out bright. With a photo open in Edit, Preview on the open photo runs a depth or Subject model on it and shows you the result before you add it. Each problem it finds says how to fix it. Nothing is copied: DarkRay reads the model where it is, so a model on a network share costs no disk space. You are responsible for the license of any model you add.

What each feature accepts, how to turn a model into ONNX, model packs and setting up many computers at once: Bring your own model.

Each row shows the model’s name, what it’s for, its path and whether the file is there. The pen icon (or a double-click) edits it: rename it, point it at another file, or change its direction. The × removes it from DarkRay; the file stays where it is, and a feature using it goes back to Auto: its built-in method, or the next model you added for it, or Needs a model when there is none.

Model for each feature

Feature	Uses	Takes your models
Subject	Select Subject masks	yes (ONNX files)
Sky	Sky masks	no
Smart Select	Click to select	no
Select by name	People, Hair, Water…	no
Depth	Depth Range masks, Atmosphere	yes
Edge matte	The soft edge of every AI mask	yes
Denoise	RAW Process › Denoise	yes (ONNX files)
Content search	content: search, suggested keywords	no

Each row’s dropdown picks the feature’s model; a feature with one model shows it, dimmed. Its status reads Ready, Needs a model (no model can run it on this computer) or File not found (a model you added whose file moved, or a share that is offline). Choosing another depth model computes the depth of the photo on screen again.

A feature that needs a model stays in its menus, marked Needs a model…; choosing it opens this page at that feature. Atmosphere says Needs a depth model with a link here. A mask made by a model you removed keeps working from its saved result, and its row says model removed; only regenerating it needs a model again.

The Subject model places the silhouette, and the edge matte then resolves hair and defocused edges at full resolution. A mask that already exists keeps the model that made it. To redo it with another, right-click the mask and choose Regenerate With; Manage Models… at the bottom of that menu opens this page.

Hardware

Setting	Default	What it does
Compute tier	Auto	Auto picks Full on a CUDA card with 8 GB of VRAM or more, and Lite on smaller cards and machines without a GPU. Full uses the larger Smart Select model. Lite uses the smaller one.
Matte working resolution	0	0 to 8192 px in steps of 256. 0 runs the matte at the photo’s own resolution: sharpest, and the most memory.

Content search

Setting	Default	What it does
Index new photos automatically	on	Builds the content index in the background after a sync or import. Off: File › Build Content Index runs it on demand.
Run content search on	CPU	CPU keeps the GPU free for editing and AI masks. GPU indexes about nine times faster and keeps the model in VRAM. Restart required.
CPU threads for indexing	0	0 uses half the processor’s threads. More is faster but competes with the rest of the app.

If a built-in model is missing, see Troubleshooting.