Fizgig 5.2: Combining Two LoRA Training Methods Beats Either Alone, Converges 25% Faster
shootthesound · reddit · 2026-09-03
Fizgig 5.2 stacks two H3 LoRA training methods: Optimised Likeness Learning (identity lives in blocks 20–49, so photo steps only train those) plus Ostris's frozen assistant-adapter that counteracts guidance-distillation drift. Across 5 datasets scored by face recognition over 45 epochs, each method alone landed within 2% ARCFace, while the combo got there a quarter sooner and finished higher. Now default in Fizgig; also adds Context LoRA and video-clip likeness mode.
More from Multimodal
- An all-AI video pipeline: nano-banana-2 keyframes, MiniMax H3 Max transitions, SAM 3 segmentation — viksit · 2026-09-03
- Runway unveils Solaris, an interface world model that renders UIs pixel by pixel without code — tlakomy · 2026-09-03
- VoxGen: a Rust + Vulkan TTS engine bringing AMD support to VoxCPM2 without PyTorch — Substantial_Swan_144 · 2026-09-03
- Arcturus shows how stadium-scale spatial AI capture could reshape sports broadcasting — bilawalsidhu · 2026-09-03
- Drone footage + LichtFeld Studio turns city aerials into stunning 3D dioramas — janusch_patas · 2026-09-03
- Invideo's New Chat-Based Editor Turns 5 Clips Into a Finished 45s Cut — azed_ai · 2026-09-03