A ground-up rewrite of the DeepFaceLab training module. Full DFL model compatibility, a modern web GUI, and a stack of new capabilities that the original never had.
Distribute training across multiple GPUs. Saturate your full rig and cut training time proportionally.
Start low-res and scale up through defined phases. Faster convergence, better final quality — the way modern training should work.
Trade compute for memory. Train larger models or higher resolutions on hardware that would otherwise run out of VRAM.
Native bfloat16 support for improved numerical stability over float16, with the same memory savings.
Pre-process and cache your dataset once. Eliminates per-iteration I/O bottlenecks that plague long training runs.
Browser-based interface accessible over your local network or remotely. Monitor and adjust settings without being at the machine — secured with password auth.
Get started with the core engine. Evaluate on your hardware before committing.
The full engine. No resolution caps, no GPU limits, secured remote access.