Diffusion LoRAs

Collection of custom-trained character, style, and concept LoRA models for Anima DiT, Illustrious SDXL, and modern diffusion pipelines

Hugging Face Model Repository Training Recipe & Code View Visual Gallery

Overview

This repository hosts a curated collection of Low-Rank Adaptation (LoRA) weights, fine-tuned checkpoints, and training datasets targeting state-of-the-art diffusion architectures, including Anima DiT and Illustrious SDXL.

All models are trained with high-efficiency pipelines using direct FP8 backward passes, persistent latent caching in system RAM, Prodigy adaptive optimization, and precision Danbooru/PixAI tagging.

Key Highlights & Features

  • Architectures Supported: Native DiT (Diffusion Transformer) modules and SDXL cross-attention/residual projections.
  • Fastpath Training Pipeline: Prodigy adaptive optimizer (d_coef=1.0), single-block checkpointing, and Triton FP8 matrix kernels maximizing effective TFLOP/s on modern NVIDIA GPUs.
  • Curated Dataset Engineering: Tagged via unified multi-stage vision models (PixAI v1.0 + WD14) with strict character trait isolation and minimal trigger leakage.
  • Production-Ready Artifacts: Safetensors format compatible with ComfyUI, WebUI, and diffusers.

Model Hub & Visual Outputs

You can download weights directly from the Hugging Face Repository, or inspect full-resolution outputs and character studies generated with these checkpoints in the Portfolio Visual Gallery.