This is not a quantization error or compression.
Details have been shifted; it may be that one of the models is starting at the wrong position.
Or in the wrong resolution.
I imagine you can fix this and that Wonder 2 Local will then also work with large images.
@Lingyu
@Ange.topazlabs
@partha.acharjee
Or a different Architecture is needed.
Mamba Zigzag Diffusor
Or this one.
I think this should be a bug report but I voted anyway
Its already a bugreport.
It could be that its related to the localisation problem.
What does 30 min finetunimg mean?
Whose 30 mins.? Initial AI trainers or ours?
Where did you read that?
Could you show me a screenshot?
@Ange.topazlabs
@alexandre.topazlabs
@Lingyu
I can confirm that the artifacts are gone in version 1.3.3 after switching to the US system format.
And yes, memory usage is now in line with what you’d expect for this model.
At one point, Wonder 2 alone was using 45 GB system memory.
Confirmed with a second photo.
Great! The issue is referenced in a development team ticket for foreign Windows machines, but still good to keep this Feature request for a model for large images or from modern cameras
It was in the CleanDIFT: Diffusion Features without Noise link you provided above.
I’ll attach a snip, Thomas.
The “30 minutes” is mentioned in the center of the page. It seemed to imply that to get the CleanDIFT quality would take 30 mins. But, perhaps I misunderstood what it was showing… do you know what that 30 mins. represents? You’re much more tech oriented than I am.
The point here is that a clean output can be achieved if the model has been fine-tuned on an A100 for 30 minutes.
That would have to be handled by TL.
At the end of the processing (20k output) cpu usage was very high and so was the ram load.
Is this normal or is something that should run on the gpu on the cpu?
Got it! That’s what I wondered - if us or them. You cleared that up. Thx!
