VisoMaster Fusion 3.4.1

This release rolls up everything that landed on dev since main was last cut at 3.0.0. It covers all intermediate patch and minor releases (3.0.1 through 3.4.1) and packages them as a single merge to main.

Highlights

• Camera Gaze Lock — new gaze-locking control in Frame Edits, with split per-eye calculation and a follow-up smoothing adjustment.
• Multi-GPU selection — pick a specific GPU when more than one is installed, applied consistently across detectors, swappers, restorers, enhancers, editors and DFM models.
• Auto Mouth upgrades — Exclude Upper Teeth toggle (keeps swapped upper teeth visible by brightness-trimming the inner-mouth mask) and Show Mouth Outline debug overlay. Default AutoMouthLowerLipParserSlider lowered from 17 to 8.
• Mouth fit & align occlusion — the swapped mouth can now inherit occlusion from the original frame for more natural blending.
• Sequential detection rework — significant fixes and optimizations to SequentialDetector and the frame_worker, plus updated 5-point landmark refresh whenever a 203/478 detection succeeds (reduces tracking drift over time).
• Video encoding refactor — video_processor.py split: encoding moved to a new app/processors/video_utils/video_encoding.py module (~700 lines), cutting video_processor.py by roughly two-thirds.
• Memory optimizations — several passes across face_restorers, face_swappers, frame_worker, models_processor and degenerate-keypoint handling for the 478-landmark model.
• ONNX Runtime / CUDA 13 — pinned to the stable onnxruntime-gpu CUDA 13 wheel; nightly index handling fixed; engine creation failure path cleaned up; .timing files now included in ONNX probe cleanup so failed retries don't loop.
• Portable launcher polish — Start_Portable.bat relocated under scripts/, dependency update fix, rollback page now shows more commits, README and download links updated to use latest.
• Workspace / job manager fixes — workspace autosave now works for marker recordings; force_recognition_in_batch flag added to jobs; input faces loading fix; batch processing fixes.
• UI polish — Face Compare keybind, preview-flashing fix, fewer card context-menu handles, models unload safeguard, dirty-state handling in feed video loop, frame edit tweaks.
• Webcam — multiple fixes to loading and runtime behavior.

Changelog by version

3.4.1

• Camera gaze lock adjustment (smoothing/behavior refinement).
• Small memory optimisation pass.
• Add .timing to ONNX probe cleanup so failed retries are reset cleanly.

3.4.0

• New Camera Gaze Lock control in Frame Edits + supporting layout entries.
• Split per-eye calculation for gaze processing.
• Frame edit tweaks (face utilities improvements).
• Video encoding extracted to app/processors/video_utils/video_encoding.py; video_processor.py significantly slimmed down.
• Reduced card context-menu handle usage (UI responsiveness).
• Launcher rollback page now shows more commits.
• Fix workspace autosave for marker recordings.
• Engine creation failure cleanup path.
• Face Compare keybind.
• Preview flashing fix.

3.3.1

• Webcam fix.

3.3.0

• Sequential detection fixes and optimizations (SequentialDetector, frame_worker).

3.2.3

• force_recognition_in_batch flag on jobs (force per-frame recognition during batch runs).

3.2.2

• Batch fixes and Input Faces loading fix.
• Memory optimization + fixes across face_restorers, face_swappers, frame_worker, models_processor.

3.2.1

• Detection fix.

3.2.0

• Multi-GPU selection support (choose target GPU; applied across detectors, swappers, restorers, enhancers, editors, DFM).
• Degenerate adjustment for the 478-landmark model.

3.1.2

• Pinned onnxruntime-gpu (CUDA 13) to a stable version.

3.1.0

• Mouth fit & align occlusion of the original frame.
• Stop processing fix.
• Memory optimisation and degenerate 5-keypoint detection bypass for face alignment.
• Models unload safeguard.

3.0.9

• "Dirty state" fix in the feed video loop.

3.0.8

• Removed legacy start.bat.

3.0.7

• README refresh.
• Webcam loading fix.

3.0.6

• Landmark fixes in SequentialDetector and FaceDetectors: 5-point landmarks used for tracking are refreshed whenever a more detailed 203/478 landmark detection succeeds — prevents tracking drift over time.

3.0.5

• "Many fixes" sweep across detectors, masks, frame worker, save/load actions, denoiser/swapper layout data and the launcher gittools.

3.0.4

• Fix portable launcher dependency update.

3.0.3

• Add ONNX Runtime CUDA 13 nightly index.
• Move launcher script sources under scripts/.

3.0.2

• Add start scripts under app/scripts.

3.0.1

• Update onnxruntime-gpu to 1.27.0.dev20260507002.
• Fix detector input size resolution in batch detection path.
• feat: Auto Mouth — Exclude Upper Teeth (brightness-based inner-mouth mask trim) and Show Mouth Outline debug overlay; default AutoMouthLowerLipParserSlider changed 17 → 8.

Contributors in this range

@Elricfae, @t3nka, @RngDevil, @chyanbo — thanks to everyone who contributed code, fixes and reviews.

Upgrade notes

• Portable launcher users: download the latest Start_Portable.bat from the new "latest" release link in the README rather than the pinned v2.0.0 URL.
• CUDA 13 / onnxruntime-gpu is pinned to a stable wheel as of 3.1.2; if you were on the nightly index, this version should be picked up automatically by the launcher's dependency update.
• Multi-GPU users: a target GPU can now be selected in settings (3.2.0+).