Spaces:
Running
on
Zero
Separate tile upscaler tab.
Can you create a separate tab (or remake Upscaler tab) similar to https://huggingface.co/spaces/gokaygokay/TileUpscalerV2 using parameter from Image Generator. Thanks!
Yeah. This upscaler is well modularized, so it's easy to combine the two programmatically, but it consumes about 6.5GB of VRAM in Spaces, which is a bit of a resource hog...😭
The existing upscaler and Upscaler tab don't consume much VRAM because the stablepy library used for image generation manages the models internally. I think they use almost the same model and algorithm internally...
If you don't mind, duplicate the space and allow the use of a custom model 🙏.
Oh, so I'm going to modify Gokay's space. Okay. That should be easy (probably).
It will probably be a little slower than the initial load (due to the dedicated cache mechanism), but it won't affect the operation itself.
However, I don't have enough Zero GPU slots for constant deployment...
Well, I'll just try modifying it later.😆
I tried modifying it for now. I don't think it will be practical without a GPU. In addition to the standard SD1.5, I have made it compatible with SDXL, but I'm not sure if this implementation is correct...
https://huggingface.co/spaces/John6666/TileUpscalerV2r
Yeah, I don't know the implementation all of that. I'm just a casual user, 😄
It doesn't work without a GPU?
It (maybe )works, but it should be extremely slow.🥶
Idk if this is correct, but since you mentioned that your zero GPU slot is maxed out, why not use this Space StablePy Tile ControlNet? For example, you could set up the TileUpscaler tab to handle everything, so the user doesn’t have to change the mode every time?
Also, maybe stopping https://huggingface.co/spaces/John6666/TileUpscalerV2r. It uses 100% CPU every time, don't want to drain Hugging Face's resources. 😄
For now, I paused it.😅
why not use this Space StablePy Tile ControlNet?
Using a library in an unexpected way is even more troublesome than modifying the library itself...
if the TileUpscaler is using a lot of VRAM, is it because it’s using SD model right? Why can't it reuse the model from the image generation?
That's because the library loading part isn't separated. Anyway, it's too much work...😭
Well, okay then, 😄 Maybe I should open a discussion in the gokaygokay space, lol. Thanks anyway 🙌