How do we access the new 4K Sana Model?

[adamreading]

Hey - you are referenced from the Sana site as supporting the new 4k model - but I can't see how to access it here?

[dseditor]

Looks like the sana checkpoint loader has no 4K update
it's still in 2k now.

[Ataraksia]

I just finished making the necessary changes to add the 4K model, and a quick check of the functionality on my end seems to be good. I've made a PR for it, but if you're impatient, you can download the updated files here

Edit: So I think there may potentially need to be some tweaking to get it working ideally with the 4K model. Feel free to test the changes as they are now though and see if it works well enough for you.

[adamreading]

I just finished making the necessary changes to add the 4K model, and a quick check of the functionality on my end seems to be good. I've made a PR for it, but if you're impatient, you can download the updated files here

Edit: So I think there may potentially need to be some tweaking to get it working ideally with the 4K model. Feel free to test the changes as they are now though and see if it works well enough for you.

Thanks I just updated everything from the fork, can see the new options now. Started a run - but seems to be downloading 12 new files so may be some time lol

[adamreading]

20 minutes later it finished loading files (I’m running in a per minute cloud server). It OOM’d on a 24GB VRAM 64GB Ram A10G system - so now I am reloading into the L40s 48GB VRAM / 64GB Ram GPU

[adamreading]

Hmmm - had the wrong checkpoint version apparently. More downloading….

[Ataraksia]

Yeah, my initial impressions were quite unfavorable as well. I'm reasonably certain I used the correct model settings, but it's possible I missed something. There are really only a couple options in the unet config that I might have configured incorrectly, but I tried testing the other possible options and the results didn't improve. Ultimately, what I found is that it is possible to get passable images (ie. images that don't contain additional/missing limbs and what not) by taking extra efforts in your prompt crafting, similar in nature to the sort of prompts that get the best results with SD1.5. It's certainly not ideal, and I have no doubt it can be improved, but honestly it's a bit more effort than I'm willing to spend on this. After seeing the results, I'm kinda thinking that's the primary reason the 4K model hadn't been added already haha.

[adamreading]

So I downloaded the 4k model - (auto) and now it’s only generating Black solids

[adamreading]

Yeah, my initial impressions were quite unfavorable as well. I'm reasonably certain I used the correct model settings, but it's possible I missed something. There are really only a couple options in the unet config that I might have configured incorrectly, but I tried testing the other possible options and the results didn't improve. Ultimately, what I found is that it is possible to get passable images (ie. images that don't contain additional/missing limbs and what not) by taking extra efforts in your prompt crafting, similar in nature to the sort of prompts that get the best results with SD1.5. It's certainly not ideal, and I have no doubt it can be improved, but honestly it's a bit more effort than I'm willing to spend on this. After seeing the results, I'm kinda thinking that's the primary reason the 4K model hadn't been added already haha.

Yup- it’s costing me too much in ultra server time. I’ll wait until its all settled and looks proper, and then have another go

[Ataraksia]

That was my determination as well.

As for the generating black solids issue, I also ran into that. While it doesn't matter much at present, in case it comes up again for you or others in the future, for me it occurred whenever I set the VAELoader dtype to "Auto" for some reason. Chalk it up as another one of the oddities of the 4K model I suppose.

[adamreading]

That was my determination as well.

As for the generating black solids issue, I also ran into that. While it doesn't matter much at present, in case it comes up again for you or others in the future, for me it occurred whenever I set the VAELoader dtype to "Auto" for some reason. Chalk it up as another one of the oddities of the 4K model I suppose.

It happened for me at BF16, and Auto. I gave up at that point.