lightx2v

/

Qwen-Image-2.1-Lightning

Like 110

the pytorch_lora_weights.safetensors works in comfy as it is. you should go to around 0.5 strenght and feed a manual linear sigmas curve into a adv. custom sampler

the pytorch_lora_weights.safetensors works in comfy as it is. you should go to around 0.5 strenght and feed a manual linear sigmas curve into a adv. custom sampler

I tested a weight of 0.5 in ComfyUI and it was indeed good, though I set the steps to 8.

strengh 1 is needed when you need texts. This lora has bad offset and prompt following

We applied a shift during training and have updated the README example accordingly.

Thanks for the reply, can you give us a comfyui workflow. comfyui seems does not have that config.

pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_config(
    pipe.scheduler.config,
    num_train_timesteps=1000,
    use_dynamic_shifting=True,
    time_shift_type="exponential",
    base_image_seq_len=256,
    max_image_seq_len=8192,
    base_shift=0.5,
    max_shift=0.9,
    shift_terminal=None,
)

Nice to have a 4 steps lora.
i'm using the official turbo checkpoint.
Can have sth to test tonight.,
Nice. thanks.,