Overlays & Metadata

Overview

The Face Browser supports several visual overlays that can be toggled on and off to inspect different aspects of each face image. These overlays help you quickly assess the quality of your extracted faces without needing to open each one individually in the Face Editor.

Overlays are controlled by toggle buttons in the Face Browser toolbar. Multiple overlays can be active simultaneously, and they render on top of each thumbnail in the grid view.

Landmarks Overlay

The Landmarks overlay displays facial landmark points on each thumbnail. These are the key reference points that DeepFaceLab uses for face alignment -- typically 68 points covering the jawline, eyebrows, eyes, nose, and mouth.

Purpose

Landmark visualization helps you identify alignment problems. If the landmark points are significantly misplaced (for example, eyes marked in the wrong position or jawline points drifting off the face), the face extraction may need to be re-run or the face should be removed from the dataset.

Reading Landmarks

• Well-aligned face -- Points follow the natural contours of the face. Eye points sit on the actual eyes, nose points trace the nose ridge, and jawline points follow the jaw.
• Poorly aligned face -- Points are offset from their expected positions. This often happens with extreme face angles, heavy occlusion, or when the extractor detects a non-face object.
• Missing landmarks -- If a face image has no landmarks stored, the overlay will show nothing. This typically means the face was imported without metadata.

Polygons Overlay

The Polygons overlay shows the polygon outlines that have been defined for each face. Polygons are closed shapes drawn in the Face Editor that define mask regions. When enabled, each thumbnail displays the polygon boundary as a colored outline.

Purpose

Polygon visualization helps you verify that manual polygon masks are correctly positioned. You can quickly scan through the dataset and spot faces where the polygon boundary has drifted or does not properly encircle the face region.

Appearance

• Polygons appear as colored outlines overlaid on the face thumbnail.
• Each polygon vertex is marked with a small dot.
• Faces without polygon data show no overlay when this toggle is enabled.

Masks Overlay

The Masks overlay displays the face mask as a semi-transparent color overlaid on each thumbnail. This is one of the most useful overlays for assessing dataset quality because the mask directly controls the blending boundary during merging.

Reading Mask Overlays

• Colored region -- The tinted area represents the mask coverage. This is the portion of the face that will be replaced during merging.
• Untinted region -- Areas without the color overlay will remain as the original destination frame pixels.
• No overlay at all -- A face thumbnail with no mask overlay when the toggle is enabled means that face has no mask data. You should generate a mask using auto-segmentation or manual painting.

Assessing Mask Quality

A good mask should:

• Cover the entire face from forehead to jawline.
• Include eyes, nose, and mouth completely.
• Have smooth, clean edges along the face boundary.
• Exclude hair, ears (usually), and background.
• Not extend significantly beyond the face perimeter.

XSeg Masks

XSeg (eXtended Segmentation) masks are generated by the XSeg model, a specialized neural network trained on your specific dataset for custom face segmentation. Unlike BiSeNet which is a general-purpose face parser, XSeg learns the specific characteristics of your source and destination faces.

XSeg vs BiSeNet

When XSeg masks are available, you can toggle their overlay independently from the standard mask overlay. This lets you compare the XSeg result against the BiSeNet or manually painted mask.

Importing Metadata

Face images extracted by DeepFaceLab contain DFLIMG metadata embedded in the image file. This metadata includes alignment information, landmark positions, source frame references, and mask data. Recaster reads this metadata automatically when you load a project.

Import Sources

You can import face images with metadata from several sources:

• DeepFaceLab workspaces -- Open an existing DFL workspace and Recaster will read all face metadata from the aligned images.
• Exported face sets -- Face images exported from another Recaster project or DFL instance retain their DFLIMG metadata.
• Manual import -- Use the Import Metadata function to read metadata from face images in an external directory.

What Metadata Contains

Each face image can contain the following embedded information:

• Face type -- Whether the face was extracted as a full_face, whole_face, or head type.
• Landmarks -- The 68 facial landmark point positions used for alignment.
• Source filename -- The original video frame that this face was extracted from.
• Source rect -- The bounding box position of the face in the original frame.
• Image-to-face matrix -- The transformation matrix used to align the face.
• Mask data -- Painted masks, polygon data, and XSeg masks.

Viewing Metadata

You can inspect the full metadata for any face in two ways:

• Selection info panel -- Select a face in the Face Browser. The metadata panel shows a summary of the selected face including type, source frame, landmark count, and mask status.
• Face Editor -- Open the face in the Face Editor to see detailed metadata in the right panel, including the full set of landmark coordinates and mask statistics.

Combining Overlays

Multiple overlays can be active at the same time. This is useful for cross-referencing information -- for example, enabling both Landmarks and Masks simultaneously lets you verify that the mask boundary correctly follows the facial landmark positions.

Common overlay combinations:

• Landmarks + Masks -- Verify mask boundaries align with facial features.
• Masks + Polygons -- Compare the painted mask against the polygon boundary.
• Masks + XSeg -- Compare auto-generated masks against XSeg-trained masks.