I now use SOLAR-10.7B-Instruct-v1.0-uncensored-GGUF to generate prompt for Flux and it’s night and d
I now use SOLAR-10.7B-Instruct-v1.0-uncensored-GGUF to generate prompt for Flux and it’s night and day, my generated images are incredible.
Gwhat is that
Gshare samples
TIt’s a LLM. Basically you type : generate prompt for « a man riding a bicycle in the street » and it generates a very long prompt with many details
TI use this node https://github.com/SeargeDP/ComfyUI_Searge_LLM
GitHub
Custom nodes for ComfyUI that utilize a language model to generate text-to-image prompts - SeargeDP/ComfyUI_Searge_LLM
T
V@Dr. Furkan Gözükara i still cannot understand your new post of kaggle swarmUI, I have sent you a private message
Shello, guys, please tell me: to make cool realistic photos of myself flying like superman I need to use 2 loras? 1 of myself and 2nd is realism lora?
Fyou need a better dataset for it
FGitHub
Contribute to kohya-ss/sd-scripts development by creating an account on GitHub.
Fi replied try to figure out logic easyt
Fyou can just train on flux and sufficient
Mhello if i want to train a flux lora and try 32 network rank should my network alpha be 32 or 1
Min sdxl i used to set it to 1
FSince if alpha is not set equal a new LR needs to be researched
Mthank you
Mis "Rank_3_18950MB_9_05_Second_IT" config for gpus with 24gb vram bf16 or are only the configs for 48gb vram bf16?
Frank 3 is for 24 gb
Frank 1 and 2 for 48gb
W@Dr. Furkan Gözükara Using the Rank5 and the latest Torch 2.5, I'm trying to train, it gets here and then just seems to not be doing the training. Also even though it says it uses less memory, my GPU is still using 23.37g out of 24.
Wnevermind it just started training although extremely slowly. If the lora source images, and target direectory are on a slower drive would that cause a delay?
Fit can cause delay in initial
Fstart
Fbut not after it
LIm curious how much RAM are you running with, looks like FLUX is heavy on the PC RAM as well
Fye it uses a lot
Fi have 64 gb so never cared :d
Falso you can set virtual ram
Fram is nowadays easier
LYeah I have more coming in the mail, but I think paging is what is causing some peoples slow down
LHey, so I ran a couple of successful Flux lora's. with the rank three settings, but. the file size is like 2.3 gigs and I'm thinking I must have done something wrong to end up with that file size. I remember having that mistake before when I accidentally ran it on the Dreambooth tab instead of the Lora tab. But that's not what I did here. Is there anything apparent that I'm doing wrong?
LFrom the video it looked like Furkans where about that size
LOh, so maybe that's the expected size using this approach. It's big, but if it works, then I'm totally fine with that.
LYeah I tried the rank 12 which does i think 512x512 training and the model was 525MB
LI am sure it goes up in size with the resolution, since it is gathering more information
Mis rank 3 config bf16?
Fall are bf16 but rank 5 and lower cast into fp8
Fsome 16 gb configs doesnt cast but they are lower resolution
MI was using this extension https://github.com/cheald/sd-webui-loractl with a1111 to change the loras weight at different steps for example I can start one lora at first step and stop it at 15th step and start another lora at 15th step until the end. Now I'm looking if I can do it in SwarmUI and the only thing I found is Lora Block Weight node in ComfyUI-Inspire-Pack.
MSo could I download this https://github.com/ltdrdata/ComfyUI-Inspire-Pack and use it inside SwarmUI?
Fnot sure may not work
Sit has been running over 3hrs and only 844/2800 done. How do I pause it to continue later? Running 14 photos on RTX 3060 12GB.
J@Dr. Furkan Gözükara You testing T5 training yet?
Fdidnt have chance yet
Fi have been trying to fix broken scripts reply people :d
Fsadly you cant
Fthere is use last checkpoint for kohya but people say not working i didnt test
Flook at the tips
F#tips-info
Sok, so the average speed is OK for my setup? ~4hrs and only 875/2800 on 12GB RTX3060
