512×512 resolution generates only top portion of image

[hyeon9698]

Thank you for sharing this great research with the community

Issue

When generating at 512×512, only the top portion of the image is generated. 1024×1024 works fine.

Code

import torch
from diffusers.pipelines.glm_image import GlmImagePipeline

pipe = GlmImagePipeline.from_pretrained("zai-org/GLM-Image", torch_dtype=torch.bfloat16)
pipe.enable_model_cpu_offload()

image = pipe(
    prompt="A dog",
    height=512, width=512,
    num_inference_steps=50,
    guidance_scale=1.5,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]

Results

prompt="A dog"

prompt="An apple",

The README shows 512×512 benchmark, so I expected it to work. Am I missing something in my configuration, or is there a workaround to make 512×512 generation work correctly?
However, 1024×1024 works perfectly.

Thank you