[bug]: Tiled upscaling requires more resources than non-tiled upscale

[grunblatt-git]

Is there an existing issue for this problem?

• [X] I have searched the existing issues

Operating system

macOS

GPU vendor

Apple Silicon (MPS)

GPU model

No response

GPU VRAM

No response

Version number

3.7.0

Browser

Chrome

Python dependencies

No response

What happened

Generation crashes during upscaling with the new upscaling Tab with the following error message:

Upscaling Tiles: 100%|██████████████████████████████████████████████████████████████████████████████████████| 2/2 [04:26<00:00, 133.33s/it]
The config attributes {'interpolation_type': 'linear', 'skip_prk_steps': True} were passed to LCMScheduler, but are not expected and will be ignored. Please verify your scheduler_config.json configuration file.
  0%|                                                                                                                | 0/3 [00:01<?, ?it/s]
[InvokeAI]::ERROR --> Error while invoking session <someUUID>, invocation <someOtherUUID> (tiled_multi_diffusion_denoise_latents): Invalid buffer size: 8.00 GB

I tried to upscale an image from 512x896 to 1024x1792 with the new upscaling tab. I used an SD 1.5 Model with the LCM scheduler and LCM Lora. It seems like the generation of the tiles works fine, but some compositioning step afterwards runs out of Memory (not quite sure how this feature works under the hood).

In comparison: Upscaling the same image with an image-to-image workflow (1 step) to the same resolution took 27 seconds on my hardware, so i am not quite sure why this workflow takes 133sec per iteration.

What you expected to happen

I expected the tiled upscale to consume less resources than upscaling the whole image at once. Also i assumed that generating a tile (133 sec) would be faster than generating a whole image at the target resolution (27 sec)

How to reproduce the problem

No response

Additional context

These were the settings that i used for upscaling:

Discord username

No response