so we can apply this to instant id?=
so we can apply this to instant id?=
Pno benefit with cpu offload for just one image, in fact it increased the time to 20s, memory remains the same.
PI think that because I can only apply float8 to unet and in cpu offload, unet is already the only one that remains in the cuda, so nothing changes if the memory consumption is lower than the total available of the video card.
Pwe need test,
Pi dont tested with controlnet yet
PFloat 8 + freeu we recover the quality result
Pbut increase inference to 19.2
Pnow fp8 + freeu + deep cache 9.3s inference
Pbut images changes
VHi Furkan, possible to change the modified date on your patreon posts when you edit the post?
VWhen viewing chronologically, I find a post from mid 2023 that was commented on a couple of weeks ago. But if I had not been paying attention to the notifications on my email i wouldnt have known.
Fi wish patreon had added it
Fbut i have last updated date here
FGitHub
Stable Diffusion, SDXL, LoRA Training, DreamBooth Training, Automatic1111 Web UI, DeepFake, Deep Fakes, TTS, Animation, Text To Video, Tutorials, Guides, Lectures, Courses, ComfyUI, Google Colab, R...
Fi sort here my self
Vyea its such a tricky system. An idea could be creating a new post with the link to the updated post as an entry that an update was made. But that could get messy. Anyways yea.
Fit is patreon mistake
Fthere should be sort by default last updated
Fi told them
Fand they still didnt add code block feature
Bif LLaVA is not used, we should just go to Stage 2? And why wasn't LLaVA implimented?
Vok this will work
Fwe also have this
FPatreon
Patreon is empowering a new generation of creators.
Support and engage with artists and creators as they live out their passions!
Support and engage with artists and creators as they live out their passions!
Fit was taking too much vram and we have better llava
Boh that's right. What about not using stage 1 at all
Fi am preparing a great cloud vm image. runpod users will love it
Fyep not useful :d
Fi dont know why they implemented it at all
Fit is also a short dirty method nothng more
Fnothing like AI some image manipulation function
Fby the way i added stop batch processing too
Fhopefully will add image comparison slider and restore face
Bhere is what am thinking, chop the source image with a big overlap, upscale as big as I can with my 3090 and stitch them back. using photoshop and mask the seams if I have to
Bbasicly post tiles method
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=40)
Dtooo much like hard work
Bwell with my 24gig 3090 i can only go to ~4000
Bwhat if I want 16k
Dwhy would u need 16k?
Bprint high res.
Bwhy do camera need 50 mega pixels
Vcomparing megapixels to max upscale model is interesting
Vi havent looked into it but it feels like you could go max upscale to parameters with ai, then move to photoshop to interpolate further. however, are you talking about like highway billboards?
Dhave you seen billboards up close lol
Vright lol, they look like ping pong balls
Dexactly
Bactually I just had a theory and want to see how difficult to pull it off
Vsounds like you are tinkering, welcome that for sure. but good to know
Bone thing i notice is each upscale does distort from original, if each tile is distort, it maybe hard to stitich them back together
Danyone seen the latest sora vids the pug wearing goggles lol
Bwonder if there an easy way to take an image and chop into 4 piece with 25% overlap