flux2-lora-training

What problem does it solve?

This Skill provides a consolidated, model-specific reference for training LoRA adapters and edit models on FLUX.2 Klein and Qwen Image Edit 2511 so engineers avoid trial-and-error when preparing datasets, choosing target modules, and tuning hyperparameters for edits, inpainting, and high-resolution training.

Core Features & Use Cases

• Model-specific guidance: Exact LoRA target modules, VAE/latent notes, and architecture differences between FLUX.2 Klein and Qwen Image Edit 2511.
• Edit workflows: Best practices for before/after (head swap, face swap) LoRAs, multi-reference training, trigger words placement, and zero_cond_t usage.
• Training recipes: Recommended hyperparameters for ai-toolkit, SimpleTuner, and DiffSynth-Studio plus curriculum strategies for large resolutions and tiling approaches for seamless outputs.
• Operational guidance: Dataset layout, known toolkit bugs and fixes, troubleshooting identity drift, and VRAM optimization tips for large-scale runs.

Quick Start

Train a LoRA to perform a head-swap edit on FLUX.2 Klein 9B using an ai-toolkit-formatted dataset with zero_cond_t enabled and the recommended lora_rank and step schedule.