random thought, can you make flux loras a gguf???
random thought, can you make flux loras a gguf???
JWhen i for example extract a lora from Checkpoint XY can i also use it with Checkpoint YX or will i have quality loss? (Both based on Flux) ?
JThink i never saw any lora in gguf format, why would you need that?
WGitHub
GGUF Quantization support for native ComfyUI models - city96/ComfyUI-GGUF
Yloras are gb for flux on 128 rank+ higher unlike sd 3.5/xls
Jlol, just checked, ive trained full checkpoint on a FP16 Model (11 GB , which resulted in finetuned checkpoint with around 22GB, so will the Lora be 11 GB big? why is it that huge?#
VDoc would be buying the Rtx 5090?
VAnd doc what do you think about NVIDIA Project DIGITS? ..Nvidia priced it at 3000$ ...Rtx 5090 is priced at 2000$...
..
..
DWhat is the state-of-the-art of creative upscaling? still SUPIR?
JThats still what im using
JAlso tried some comfy workflows but they cant reach supir level
Dlet's see. I believe DIGITS at 3k for this spec when I have it on my desk. Most likely with NVIDIA, 3k is the price for a DIGITS-060, and the DIGITS-090 is 3x that
Dthanks
DSUPIR was using JuggernautXL when I last used it. Have they upgraded to newer models such as Flux?
J@Furkan Gözükara SECourses Is Pinokio used for anything at massedcompute or can i delete it for more space?
JIm still using XL Models for Upscaling, dont even know if Flux works with supir
Ddon't know either, but I had problems with SUPIR with teeth and other things XL typically has problems with, or needs ADetailer to fix it. So using the newer models would make sense
JNever had those problems with realvis
JHow can i view the metadata of a flux checkpoint?
Dwhat do you mean, what metadata
Jlike info tab in a1111 when selecting a checkpoint
Ddid you just answer your own question? 
JNo does not work with flux checkpoints
Pthere are rumors about a new nvidia workstation grade card, with 96gb vram, price will be like 10k$ or even more and im pretty sure it will destroy the project digits in performance
i have a feeling that digits will be a big disappointment, but still im hyped about the release 
i have a feeling that digits will be a big disappointment, but still im hyped about the release 
Dwhat do you mean rumors? that's similar to their current lineup. of course they will release blackwell workstation cards
Prumors, until they are confirmed
Dthe promise of DIGITS was to have 128 GB mixed RAM/vRAM for a reasonable price, which they could actually do because it's hard to abuse a desktop PC in datacenters. that's probably the main reason why most cards only have 16 GB vram, to avoid AI companies to buy consumer cards
Dit's not a rumor if it already exists.
Pwhere
Da6000 costs about 8k for 48 GB vram, A100 not sure google says 20k with 96 GB.
Aand we rent 0.33/h crazy
Dyou're right. maybe there is an oversupply of 48GB cards because the professional users now all need more, that's why we get it for cheap?
the next step up the A100 is usally like 5x the price to rent
the next step up the A100 is usally like 5x the price to rent
Nyes, curiously! FLUX has great quality but something about people generated with it looks artificial. You can make Flux create more life-like people, but I still get much easier and better results with good old SD1.5
Dit's the lack of negative prompting. they could have easily removed the plastic skin during distillation, but for some reason they didn't. with a good enough prompt, Flux is able to do realistic skin
NI just got in here, so I don't know for sure, but maybe we can finetune FLUX to be more realistic as well?
DFlux with negative prompting
Dyes. there probably already are on civitai
Nhow can you do FLUX with negative prompt?
Dyou normally can't, this was an experiment that required code changes and is 3x slower than normal
Nwhoa
NI would be realliy curious to try, but FLUX is already painstackingly slow on my machine, so I can only speculate 
Mwait what!
M@Dr. Furkan Gözükara doc what do you think of this node, is it acuurate?
D?
Mthis is huge for flux
Dhave you seen that?
I'd much rather train it into a LoRA
I'd much rather train it into a LoRA
Fyou can convert them to fp8 but i dont know gguf lora
Ftheir demos are excellent
Fbut the app doesnt produce that quality atm
Fauthors said they will fix issues
