SDXL LoRA Fine-Tune
Live Demo
Try the LoRA
Describe a scene and it'll be generated with the trained cyberpunk-style LoRA on top of SDXL base, running on a RunPod GPU worker. Cold start can take up to a minute if the worker has spun down.
Fine-tuned Stable Diffusion XL via LoRA (Diffusers, PEFT) to specialize image generation toward a custom visual style, training a low-rank adapter on the UNet rather than the full 2.6B+ parameter model.
Technical highlights
- 01
Fine-tuned Stable Diffusion XL via LoRA (Diffusers, PEFT) to specialize image generation toward a custom visual style, training a low-rank adapter on the UNet rather than the full 2.6B+ parameter model
- 02
Engineered a memory-constrained training pipeline in PyTorch and Accelerate (bf16 mixed precision, gradient checkpointing, 8-bit Adam via bitsandbytes, gradient accumulation) that fit entirely on a single consumer 8GB GPU
- 03
Built the end-to-end pipeline from scratch: dataset curation/preprocessing with Hugging Face Datasets, checkpoint-based overfitting analysis across training runs, and a CLI plus Gradio web app for inference
Tech Stack