Can someone make it more efficient.... I only have 8GB vram..but I do have 64gb RAM....😅
Can someone make it more efficient.... I only have 8GB vram..but I do have 64gb RAM....
Awhere is the optimizations that reduces vram useage on there? is it the --use_tile_vae?
TLoading half params + tile vae + no llava
TNVidia card ?
VYes rtx 3070ti
VAmd right now is super shit . Nvidia is much more easy and convenient for AI purpose..
TAnd you have error message ? By default I think when vram is full it uses sysram
VI haven't tried it yet .
VAs it said it needed minimum of 12gb vram
TTry and tell us 
TSame problem here! Any ideas to solve it yet?
PTrying to laucn SUPIr and getting RuntimeError: PytorchStreamReader failed reading zip archive: failed finding central directory
K@Dr. Furkan Gözükara what exactly does the weight_decay=0.01 do?
Kstill getting that accelerate not founod error
Kanyone else have that issue in setting up dreambooth?
K(locally)
Poww haha but you have how much for RAM? you shared gpu
K@Dr. Furkan Gözükara is it possible for supir batch, to run both stage one and stage 2, batch only uses stage 2 at the moment.
AAnyone experimenting with comfyui and supir?
PAnyone here have a specific sdxl model or lora recommendation for making good hands?
Fsorry guys i will check messages now
Fsure i will add this hopefully today
Fstarting to work now
Ahttps://github.com/layerdiffusion/sd-forge-layerdiffusion.git how do you install this? creates transparent backgrounds
RFhas anyone solved how SUPIR upscale works? in terms of output resolution?
Ffor example if i give 1024x1024 and upscale 1x i get 1024x1024 output
then i give 512x512 and do 1x upscale and i get output 1024x1024
then i give 256x256 and do 1x upscale and i get output 1024x1024
then i give 512x512 and do 1x upscale and i get output 1024x1024
then i give 256x256 and do 1x upscale and i get output 1024x1024
Fthis is revoluionary amazing
Fwell is not that revolutionary
Fthere is AI background removal already
FFjust pass the result to this node, for example
Fit's just a plugin for sd forge?
Fnope
Fread it
F
Fi did. what i missed?
F
Fit generate image without background remove
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=40)
Dnice
Fyou cant get this image
Fwith background remove ever
FYou can see that the native transparent diffusion can process transparent glass, semi-transparent glowing effects, etc, that are not possible with simple background removal methods. Native transparent diffusion also gives you detailed fur, hair, whiskers, and detailed structure like that skeleton.
Fthis?
Foh i see
Fso how the hell this work?
Fthey trained new models for this
Fhuge task
FSD only uses 3 channels, red green and blue
Fi bet this will be paper 100%
Dim sure it will be transparent lol