Weird Tiled Artifacts with pipeline.enable_vae_tiling()

[larinius]

Problem
When I use pipeline.enable_vae_tiling(), I get strange tiled artifacts.

Observed Behavior:

The tiles' size appears to be calculated as sample_size = 128 - margins, leading to visible seams in the output image.
Some tiles are sharp, while others are blurry, resulting in an inconsistent look across the generated image.
Without sample_size:

If I don’t explicitly set sample_size, the image splits into two 512x512 parts.
One part is fine, but the other is darker and blurry. This same blurring effect is also present in the tiled output.

Expected Behavior
Tiles should blend seamlessly, producing a consistent and artifact-free image.

Original input image.

Setup: SD 1.5, txt2img or img2img pipeline, with output type "pil".

My code is almost default from documentation examples:

#Load model and vae:

pipeline = StableDiffusionPipeline.from_single_file(model_path, **pipeline_kwargs)

pipeline.vae = AutoencoderKL.from_single_file(
    os.path.join(directory_vae, vae_path),
    torch_dtype=torch.float16,
    sample_size=128
)                         

pipeline.enable_vae_slicing()
pipeline.enable_vae_tiling()

#Create img2img pipe:
pipe_img2img = AutoPipelineForImage2Image.from_pipe(pipe)

How to fix this issue?

[larinius]

I couldn't find any issues with AutoencoderKL decoding, it seems correct, but decoding does not take into account top and bottom padding, so one row is darker or brighter than another row, and vertical lines are broken.

[larinius]

After some digging, I am almost sure that the tiling issue comes from the blending code, which behaves differently on the "torch_directml" device. So far, any code related to Gaussian blur and blending does not work correctly on DirectML devices. If I use the CPU device, there are no issues at all; everything works perfectly fine.
This bug was reported 2 years ago in the Microsoft DirectML repo and seems unlikely to be fixed in the future.