GaParmar/img2img-turbo

Adapts SD-Turbo through adversarial learning with minimal trainable parameters—using LoRA adapters, skip connections, and frozen diffusion weights—enabling both paired (pix2pix-turbo) and unpaired (CycleGAN-turbo) translations. Achieves sub-second inference on GPUs (0.11s on A100 for 512×512) while supporting diverse outputs via noise variation and text prompts. Provides pre-trained models for edge-to-image, sketch-to-photo, day/night, and weather translation tasks with Gradio interfaces for local deployment.

2,401 stars. No commits in the last 6 months.