MultiDiffusion under Forge unreasonably slow with large images, Automatic performs much better

[minsky91]

In my short (1 month) experience with this platform, Forge has usually been much faster at generating or upscaling large images and generally much more robust, particularly when dealing with VRAM memory issues, than the plain Automatic1111. That applies to upscaling with MultiDiffusion as well, something which I do a lot these days. However, when I moved to upscaling super large images, the situation proved to be the opposite. When 2 x upscaling an 8K image on my nvidia rtx 4070 ti Super with 16 GB VRAM, the plain Automatic manages to do the job in under 18 min., although on another occasion it might take twice as long for an image of the same dimensions. Unfortunately, as I found, MD under that version of webui can only use SD 1.5 checkpoints. Forge, in comparison, took 283 minutes for the same image and default MD parameters, but using a SDXL checkpoint. (I let it run for this long only for the reasons of posting this, usually I just break the process when I see it’s taking so long.) In other cases when MD-processing images this large, Forge just throws some cryptic error midway, not finishing the job.

When it doesn’t throw an error, the pattern is always the same: the MD process under Forge reaches very quickly (in a few minutes that is) to the 80% progress mark, and then stagnates for hours, with the ETA counter only increasing the estimate by a second for the remaining duration. This is something that I saw other Forge users complaining about on the SD sub on reddit. I am attaching console logs of the MultiDiffusion process under Forge, hope this will help the developer to fix the issue.

A GAlaxy-shaped fractal 25-1 - Forge webui logs 283m.53s.txt