Most diffusion fine-tuning tooling forces a choice: a friendly UI that hides the knobs, or a config-file workflow that assumes you already know what every flag does. This project refuses the trade-off, exposing the same training engine through both a CLI and a web dashboard — and quietly tracking the field's churn by adding new architectures as they ship rather than freezing on whichever model was hot at release.
What Sets It Apart
- Breadth across modalities: image (FLUX.1/2, SDXL, SD 1.5, Qwen-Image, HiDream, Z-Image), video (Wan 2.1/2.2, LTX), and audio (Ace Step) all share one trainer, so a learned config transfers across model families.
- Method-level control: LoRA and LoKr with adjustable rank, plus layer-specific training via
only_if_containsand selective exclusion — you can train only the layers that matter instead of the whole network. - Consumer-hardware focus: configurable VRAM optimization targets single NVIDIA GPUs, with optional offload to RunPod, Modal, or Ostris Cloud when local memory runs out.
- A dashboard at
localhost:8675that monitors runs without needing to stay open, separating launch from observation.
Who It's For
Great fit if you fine-tune image or video diffusion models regularly and want one tool that keeps pace with new architectures instead of switching repos every quarter. Look elsewhere if you need a fully managed, no-config service — this still expects you to understand rank, layer targeting, and VRAM trade-offs, and assumes an NVIDIA GPU.
