■ Release Overview
Hello AI creators, trainers, and researchers on dfldata.cc!
After extensive core rewrites and performance tuning, DeepAscension LAB is officially live, powered by the next-generation SAENXT PyTorch engine.
Whether you are training high-resolution portrait models or tuning lightweight models for live execution, this release provides faster loss convergence, cleaner edge boundaries, and full parameter control without command-line restrictions.
■ Technical Highlights & Key Features
1. Next-Gen Architectures & Modern Optimizers
• Architecture Support: Full native integration for LIAE-UDT, DF-UDT, and high-efficiency lightweight LIAE-SB / DF-SB configurations.
• Modern Optimizers: Native integration for Muon, AdaBelief, and ACProp, achieving significantly faster convergence and eliminating high-frequency distortion around eyes, teeth, and skin pores.
• Uncapped Progressive Resolution Scaling: Smoothly upgrade models across any resolution jump (224 → 448, 512, 768, 1024, and beyond). Scale existing models without restarting from scratch to maximize VRAM efficiency and output resolution.
2. Best-in-Class Occlusion Handling & Object Obstruction Packs
• Curated Obstruction Training Packs: Pre-packaged, specialized obstruction datasets built to train models through difficult real-world interference (hands, microphones, glasses, eating/drinking, and dynamic occlusions).
• Precise XSeg Masking Pipeline: High-precision segmentation prevents face-tearing and mask boundary artifacts, maintaining clean face-to-background transitions even during extreme angles and partial obstructions.
3. All-in-One Desktop GUI + Remote WebUI
• Unified Control: Replaces command-line windows with a responsive interface. Monitor loss curves, inspect iteration previews, and fine-tune parameters on the fly.
• WebUI Accessibility: Monitor and manage active training runs directly from your local browser or local network.
• 150+ Toolset Enhancements: Automated dataset filtering, intelligent color matching, advanced mask refinement, and high-frequency noise reduction.
4. Commercial Live Suite: DeepAscensionLive (Coming Soon)
• High-FPS, low-latency live broadcasting suite currently in active development.
• Future access to proprietary NXT-GIP and NXT-MGIP training pipelines engineered for professional creators.
|Archiver|手机版|deepfacelab中文网
|网站地图
楼主
发表于