LightX2V releases Qwen-Image-2.1 Lightning v2 RL, a four-step version of Qwen-Image 2.1 for generation and editing
LightX2V has released Qwen-Image-2.1 Lightning v2 RL, a distilled version of Qwen-Image 2.1 refined with reinforcement learning. It makes and edits images in 4 denoising steps at CFG 1.0, including single-reference edits and multi-reference generation.
What it does
It is a full replacement transformer, not a LoRA: you load the processor, text encoder and VAE from the original Qwen-Image-2.1 pipeline and swap in this transformer. An inference_config.json holds the four pre-shifted sampling sigmas and the guidance setting, and the model card has a Diffusers example.
Worth knowing
The weights are stored in FP32 (load them in BF16), and the example needs a recent Diffusers build with QwenImage21Pipeline support. The card shows Diffusers code, not a ComfyUI workflow. It uses the Qwen research licence from the base model, so check it before commercial use.