Trying to get omnimattezero installed but cant find the models

Nice. Looks like there may be a more suspicious culprit than offloading:

---

I think the faint grid is probably not caused by the GGUF file itself, and maybe not by block_num either. The first thing I would test is VAE tiling.

In the current ComfyUI_OmnimatteZero code, block_num = 0 does not mean “disable all memory-saving behavior”. It only avoids the apply_group_offloading(...) branch. The fallback branch still calls model.enable_model_cpu_offload().

Relevant file:

OmnimatteZero_node.py

The relevant logic is roughly:

if block_num > 0:
    apply_group_offloading(
        model.transformer,
        onload_device=torch.device("cuda"),
        offload_type="block_level",
        num_blocks_per_group=block_num,
    )
else:
    model.enable_model_cpu_offload()

So the behavior is:

block_num value	What happens	What it does not mean
> 0	Uses Diffusers group offloading on the transformer	Not image tiling
0	Skips group offloading, but still enables model CPU offload	Not “disable all offloading / tiling”

The grid artifact sounds more like a VAE tiling artifact. In object_removal.py, VAE tiling is enabled unconditionally:

pipe.vae.enable_tiling()

There is also another one for the upsample path:

pipe_upsample.vae.enable_tiling()

Diffusers’ own docs describe VAE tiling as a memory-saving method that decodes the image in overlapping tiles, and they note that tile-to-tile tone variation can happen:

Diffusers memory optimization docs: VAE tiling

So if you have a 32GB card, I would test disabling VAE tiling first.

Patch 1: disable VAE tiling

Open:

ComfyUI/custom_nodes/ComfyUI_OmnimatteZero/object_removal.py

Find:

pipe.vae.enable_tiling()

Change it to:

# pipe.vae.enable_tiling()

Also find:

pipe_upsample.vae.enable_tiling()

Change it to:

# pipe_upsample.vae.enable_tiling()

Then fully restart ComfyUI and test the same workflow again.

Patch 2: if you also want to avoid model CPU offload

This is optional. I would only try this after testing VAE tiling first.

Open:

ComfyUI/custom_nodes/ComfyUI_OmnimatteZero/OmnimatteZero_node.py

Find this block:

if block_num > 0:
    apply_group_offloading(
        model.transformer,
        onload_device=torch.device("cuda"),
        offload_type="block_level",
        num_blocks_per_group=block_num,
    )
else:
    model.enable_model_cpu_offload()

For a 32GB card, you can try replacing the else branch with model.to(device):

if block_num > 0:
    apply_group_offloading(
        model.transformer,
        onload_device=torch.device("cuda"),
        offload_type="block_level",
        num_blocks_per_group=block_num,
    )
else:
    model.to(device)

There are two similar blocks in the file, so check both:

1. the normal inference path;
2. the compose/background-replacement path.

The second one appears around the compose_video(...) path.

Test order I would use

Step	Change	Reason
1	Comment out pipe.vae.enable_tiling()	Most likely source of a visible grid
2	Also comment out pipe_upsample.vae.enable_tiling() if using the upsample path	Same reason, but only matters if that path is used
3	Keep block_num = 0	Avoid group offloading while testing
4	If the grid remains, replace model.enable_model_cpu_offload() with model.to(device)	Tests whether CPU/model offload is involved
5	If VRAM runs out, restore offloading or use a smaller GGUF quant	Q8_0 can still be heavy depending on resolution and frame count

Important caveat

Disabling VAE tiling increases VRAM use. A 32GB card has a much better chance of handling it, but it can still OOM depending on:

• resolution;
• number of frames;
• whether Q8_0 / Q6_K / Q5_K_M / Q4_K_M is used;
• whether upsampling is enabled;
• whether compose mode / background replacement is enabled;
• how much VRAM ComfyUI already has occupied.

So I would not change everything at once. First test only this:

# pipe.vae.enable_tiling()
# pipe_upsample.vae.enable_tiling()

If that removes the grid, then the issue was probably VAE tiling rather than group offloading.