VisoMaster Fusion 2.0.0 — Release Notes

What's New

This release represents a major overhaul of VisoMaster Fusion, bringing new features, significant performance improvements, extensive bug fixes, and a more polished UI across the board.
Thanks to all contributors !

New Features

Face Re-Aging

A new Face Re-Aging processor lets you transform the apparent age of a face in real-time. The age can be adjusted from 0–100 using a dedicated slider. The underlying ONNX model runs with device-aware FP16 inference and includes normalization guards and CPU fallback for compatibility.

Auto Mouth Expression (Mouth Action Detector)

Replaced the previous mouth-open toggle with a full Mouth Action Detector — a TensorFlow-based model that automatically detects mouth movement states with hysteresis to prevent rapid oscillation. Supports both 68-point and 203-point landmark formats, includes occlusion timeout handling, EMA decay, and is fully thread-safe.

ByteTrack Multi-Face Tracking

Integrated ByteTrack — a state-of-the-art multi-object tracking algorithm with Kalman filter-based motion prediction — to greatly improve temporal consistency in face detection across frames.

VR180 Tiled Face Detection

Added a new _detect_faces_vr_tiled() pipeline that splits the equirectangular frame into up to 24 perspective crops before detection, dramatically improving detection of faces at the edges or in challenging VR180 layouts. New settings:

• VR180TileDetectionToggle (default: on)
• VR180MaxFOVSlider (60–150°, default: 120°)
• VR180CropResolutionSelection (512 / 768 / 1024, default: 512)
• VR180EyeModeSelection — Both Eyes / Single Eye

Theatre / Fullscreen Mode

Added a dedicated Theatre Mode with optional true fullscreen support, persistent state snapshots, and smooth transitions. Controlled via TheatreModeUsesFullscreenToggle.

Output Location Controls

New output settings section:

• Output to Target Location — saves output next to the source media file
• Cluster Output by Source — organizes output into subfolders by source name

Global Input Resize

Added GlobalInputResizeToggle and GlobalInputResizeSizeSelection (540p–2160p) to rescale input before processing, enabling better throughput on high-resolution sources.

New UI Themes

Three new QSS themes:

• Dark Blue — cool dark interface with blue accents
• OLED Black — true black for OLED displays
• Windows 11 Dark — matches Windows 11 system dark aesthetic

Collapsible Parameter Sections

Parameter panels now support persistent collapsible sections, reducing clutter and keeping the workspace focused.

Bug Fixes

Video Playback & Seeking

• Fixed seeking lag; added FrameWorkerDelayDecimalSlider (0–3s) to prevent GPU overload during seeks
• Fixed playback buffering issues and memory bloat from unbounded display frame queues
• Fixed stale preview frame when switching between target videos

Audio / Video Sync

• Fixed skipped-frame audio rebuild with a codec-agnostic approach
• Fixed audio segmenting in default-style recording mode
• Fixed audio merge cleanup and timing issues
• Made skipped-frame audio rebuild fully sync-aware

VR180 Pipeline

• Added torch.clamp() around asin() argument — prevents NaN when coordinates fall outside valid FOV
• Changed out-of-FOV grid coordinates from 2.0 → 1.0
• Changed padding mode from zeros → border to eliminate color fringing at equirectangular edges
• Fixed align_corners from False → True for correct grid sampling
• Scaled feather_radius dynamically based on _mask_region_side
• Added single-eye detection and landmark fallback
• Added warning log for silently dropped faces

Face Detection & Landmarks

• Fixed detector input size reading (now uses model-declared input shape instead of hardcoded values)
• Fixed KPS resize and tracking regressions
• Separated 203-point KPS model; updated input resize for better upscale results
• Added _is_kps_valid() guard to skip invalid landmark sets
• Added guard on kps_5_adj.shape[0] >= 5
• Fixed grayscale channel check (channel == 2 → channel == 1)
• Fixed auto-rotation landmark refinement for upside-down faces

Issue Scan System

• Stabilized issue scan processor state and worker startup/shutdown lifecycle
• Fixed issue scan UI restore behavior and range normalization
• Blocked structural UI mutations during active scans
• Kept issue scan progress tooltip static during scans
• Added stacktrace logging to issue scanner errors
• Reduced dense landmark smoothing warning spam
• Wired issue scan UI to marker-resolved runtime state

Expression Restorer & Re-Aging

• Fixed ByteTrack + Expression Restorer interaction
• Fixed re-aging edge cases and dtype/scale normalization at entry point
• DFM output scaling corrected (scaled ×255)
• Fixed DFM FW-QUAL-08 threshold (1.0 → 30.0)

Numerical / Safety Guards

• Added NaN/Inf guards after every from_estimate() call
• Added None/NaN guard for GhostFace M similarity transform matrix
• Fixed memory total > 0 guard against division by zero in VRAM reporting
• Removed tautology in preset application (face_id == face_id)

UI Fixes

• Fixed "Open file location" for both Target and Input panels
• Fixed eye blend normalize behavior and toggle UX
• Fixed webcam restore and denoiser row visibility
• Fixed fullscreen workspace restore and theatre state snapshots
• Fixed GFPGAN model load issues

Performance Improvements

Frame Worker

• v2.Resize caching: Resize objects cached by (h, w, interpolation, antialias) — eliminates repeated allocations per frame
• Border mask skip: Border mask computation skipped when BordermaskEnableToggle is off
• Lazy snapshot clone: Restoration-1 before snapshot clone made lazy
• PerspectiveConverter skip: Skipped when no VR crops exist
• FFT/pooling skip: analyze_image FFT/pooling skipped when debug mode is off
• Deepcopy reduction: Replaced deepcopy with shallow copy in single-frame mode
• Reuse: FrameEnhancers and FrameEdits reused from models_processor instead of being recreated per worker

Cache Optimization

• _transform_cache converted to OrderedDict LRU (max 32)
• _resize_cache converted to OrderedDict LRU (max 16)
• _ddim_schedule_cache converted to OrderedDict LRU (max 20)
• MAX_GABOR_CACHE reduced from 32 → 16
• EquirectangularConverter cached at FrameWorker level

Memory Management

• Added torch.cuda.empty_cache() to DFM FIFO eviction
• Added gc.collect() + torch.cuda.empty_cache() after thread cleanup
• Cleared _seek_cached_frame on stop_processing()
• Cleared KV maps and embeddings before deleteLater() in panel removal
• Added stale-entry pruning to track_history
• VR crop tensors deleted immediately after stitching
• Trimmed _color_stats_ema to active faces only

UI / UX Improvements

• Elapsed time shown beside seek controls
• Folder shortcuts and standard directory icons in file dialogs
• Shared face thumbnail size toggle with smaller default
• Target media thumbnail improvements and filter control refinements
• Input faces path controls aligned with target media layout
• Mouse wheel support for parameter widgets (opt-in)
• Embedding overwrite confirmation before saving over an existing file
• Recording stop confirmation toggle (ConfirmBeforeStoppingRecordingToggle)
• Clear-all context actions added to media face and embedding panels
• Label target faces and polished parameter action buttons
• Collapsed hidden parameter rows to reduce UI noise
• Removed deprecated mask and settings controls

New / Updated Settings

Removed / Deprecated

• CommandLineDebugEnableToggle removed
• Deprecated mask and settings controls removed from UI
• LivePortrait TensorRT model list removed (lazy build used instead)
• Auto mouth expression updated from mouth-open detection to full mouth-action model
• Custom provider kernel branch separated into standalone test branch (test-custom-kernels)

Documentation

• Added Quick Start Guide (docs/quickstart.md)
• Added User Manual (docs/user_manual.md)
• Updated README with feature descriptions and installation guidance
• Added job manager workflow documentation

Testing

27+ new test files added covering:

• Face re-aging dtype handling and logic
• Mouth openness state machine
• Frame worker quality fixes (DFM scale, keypoints guard)
• VR180 pipeline (unit + integration)
• Audio rebuild and sync
• Issue scan progress and UI lifecycle
• Job manager and save/load actions
• Panel clear-all reset helpers