How to train a 4-step or 8-step LoRA for the Qwen-Image-Edit model？

[MachineLP]

We know that the native Qwen-Image-Edit requires many steps to produce a high-quality image, which is time-consuming. Reducing this to just 4 or 8 steps would significantly cut down on generation time.

My question is: How to train a 4-step or 8-step LoRA specifically for my own tasks? What is the current methodology used to create these few-step models?

[Deng-Xian-Sheng]

My guess is that the number of diffusion steps during training was relatively small, and then a large amount of dataset was distilled.

This isn't an improvement, but rather a slight decrease in aesthetics and stability for speed.

[Deng-Xian-Sheng]

You can study this Qwen-Image-Edit-Lightning-8steps-V1.0-bf16.safetensors