Noisy pattern on qwen image edit outputs

[andrijazz]

Hi, thanks for the amazing work!

I am getting this noisy pattern consistently when using qwen image edit. It gets significantly more visible when used with lighting lora.

Here is the code snippet.

import math
import torch
from diffusers import QwenImageEditPipeline, FlowMatchEulerDiscreteScheduler
from PIL import Image

scheduler_config = {
    "base_image_seq_len": 256,
    "base_shift": math.log(3),  # We use shift=3 in distillation
    "invert_sigmas": False,
    "max_image_seq_len": 8192,
    "max_shift": math.log(3),  # We use shift=3 in distillation
    "num_train_timesteps": 1000,
    "shift": 1.0,
    "shift_terminal": None,  # set shift_terminal to None
    "stochastic_sampling": False,
    "time_shift_type": "exponential",
    "use_beta_sigmas": False,
    "use_dynamic_shifting": True,
    "use_exponential_sigmas": False,
    "use_karras_sigmas": False,
}
scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)
pipeline = QwenImageEditPipeline.from_pretrained("Qwen/Qwen-Image-Edit", scheduler=scheduler, torch_dtype=torch.bfloat16)
pipeline.to("cuda")
pipeline.load_lora_weights("/storage/models/Qwen-Image-Edit-Lightning-8steps-V1.0-bf16.safetensors")

image = Image.open("bike_test.png")
prompt = "Change the background to a park on a sunny day."
output = pipeline(
    image,
    prompt,
    num_inference_steps=8,
    true_cfg_scale=1.0,
    negative_prompt=" ",
    height=1024,
    width=1024,
    num_images_per_prompt=1,
    generator=torch.Generator(device="cuda").manual_seed(0)
).images[0]
output.save("output.png")

Any suggestions how to get around this issue? I found that it looks slightly better when used with default scheduler configuration.