I mean if what he uses is just a Flux Dev but fast gpus, thats ok
I mean if what he uses is just a Flux Dev but fast gpus, thats ok
Frtx 4090 is already 5 second
Fon swarmui
Fwith --fast
DBut I thought he might be using another pipeline
Dcan you give a little bit more details on the generation settings, like resolution, refiner, upscale etc?
Fjsut 1024x1024 20 steps
Fbut pertsonally i use 40 steps
Fand inpaint face
Fshown in the tutorial video
DI also activated --fast, but for 768x1024, 20 steps it lasts at least 11 seconds on my 4090
Ftell me your step speed
Fby the way 5 second was probably wrong 
Fit should be more like 10 second
Fi saw 2.5 it / second on runpod
Daah ok then 
Fso if you use h100 and batch size it would be like 5 second maybe lesser
Dbut still, photoai probably uses an upscaler and other things. I dont get how it generates so fast. Maybe 4x100 or stg
Fyou can try torch 2.5
Fmay speed up
Dthanks!
Fwell he sells so expensive
Fhe can use anything :d
Dreplicate is expensive though
Das compared to runpos. They sell A100 around 5$/hour.
Fyes replicate expensive too
DDr, when will you create a cheaper rival to Replicate
DReplicate the replicate!
FFmassed compute working good :d
Di have one more question. Do you think models capability of following prompts decrease if we train text encoder? Cause in your Rank 3 configuration there was a non-zero learning rate for text encoder?
Fi didnt notice such thing
Ais there any place where i can train a flux lora for free ?
Ai understand that i can make one for just 3$ on civit ai , but the thing is i am underage so i do not have access to online payment thing hence looking for a free alternative
Z@Dr. Furkan Gözükara I see in the supir app you can upload a dreambooth model for face correction, can we use a LORA instead for low vram?
Fno platform for free
Fit requires bf16
Ffree platofrms dont provide such gpu
Fit uses a base model
Zah ok any recommendations for improving images of self on those trained on lora?
Zusing SUPIR?
V@Dr. Furkan Gözükara how big (mb wise) is the LoRA you did with the 15 image dataset?
Fyep
Fit depends several factors
Fbut starting from 2 gb to 3.5 gb
Fsingle layer training arrived to kohya
KKohya gets stuck here
KNo errors
KWell I find this scrolling up
