On Forge Neo, TAESD causes some images to take much longer to generate
This is an issue I've been noticing a lot lately, and while it may just be exclusive to Forge Neo, I may as well post about it here just to get more eyes on it and see if it's an actual issue with the model or not.
I'm running it on an RTX 3060 with 12GB vram. I've had this computer for a few years now, and I've updated the drivers, python installation, and whatnot.
Right now I'm generating a batch of images at 1280x1280, but this issue is not exclusive to certain resolutions. I generated a couple batches before this one, and they all were pretty consistently at ~2s/it. After starting the next batch, it suddenly jumped to ~16s/it, and then I'm pretty sure some kind of low vram counter measure kicked in and it's slowly taking less and less seconds per iteration now. I'll check whatever settings I have now, but I figured I should post about this issue just in case anyone else has encountered it & potentially has a fix.
Also sometimes I see this happening in the middle of a batch as well, so I don't think it's exclusive to starting a new batch. I'm not loading in a bunch of new LoRAs when it happens, just using the same LoRA here, and even then this has happened without LoRA use IIRC.
Update: I suspected it had something to do with the previews, so I changed the preview mode from TASED to Approx NN. It now takes far less time upscaling with MultiDiffusion at least, down from several minutes per step to ~15 seconds per step.
If anyone else can replicate TAESD being the issue then I will close this discussion.
https://github.com/Haoming02/sd-webui-forge-classic/issues/890 This pretty much confirms it is a TAESD issue. I'm gonna go ahead and close this now, since the issue doesn't appear to be with Anima itself.