Z-Image base weird inference behavior. Noise abnormal
Author: PeterL1nCreated Feb 4, 2026Updated May 3, 2026
Is Z-Image actually the raw model after SFT? Doesn't seem like it?
If I do a single step inference at pure noise, convert v prediction to x0, you can see the model doesn't generate a smooth blurry result like normal flow matching models would. It generates very noisy one. Code attached below. I am sure it is not the sampler problem as I have tried to implement it raw. Seems like it is the model issue?
Did you guys use a very unusual timestep importance sampling when training the model?
import torch
from diffusers import ZImagePipeline
# Load the pipeline
pipe = ZImagePipeline.from_pretrained(
"/opt/tiger/vfm/Z-Image",
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=False,
)
pipe.to("cuda")
# Generate image
prompt = "An astronaut riding a horse on mars. Earth is in the background. A small amount of dust is floating in the air. He is holding a flag."
negative_prompt = "" # Optional, but would be powerful when you want to remove some unwanted content
image = pipe(
prompt=prompt,
negative_prompt=negative_prompt,
height=512,
width=512,
cfg_normalization=False,
num_inference_steps=1,
guidance_scale=1,
generator=torch.Generator("cuda").manual_seed(42),
).images[0]
image.save("example_cfg1_step1.png")Source: Tongyi-MAI/Z-Image