I can use the fp8 flux model i install with your swarmui installer ?
I can use the fp8 flux model i install with your swarmui installer ?
Fyes it has fp8 flux
4I know the difference quality between fp8 and fp16 is not huge when using flux but for a dreambooth traning is it the same ?
Fbut dont use scaled i am not sure if tht works
Fby the way
Flatest fine tuning downloader
Fdownloads fp8 model already
F
Fwe cant train fp8 with dreambooth
Fbut lora should work fairly fine
Fselect base fp8 on interface
4Ok i will try thanks 
JI am following your extracted LoRA tutorial. However getting an error. I have tried all save percisions and network rank sizes. I have checked V2 and unchecked it. Can seem to get past this error. Happening to 3 different finetunes. So I know it's got to be something I am not seeing.
new_checkpoint["time_embedding.linear_1.weight"] = unet_state_dict["time_embed.0.weight"]
KeyError: 'time_embed.0.weight'
new_checkpoint["time_embedding.linear_1.weight"] = unet_state_dict["time_embed.0.weight"]
KeyError: 'time_embed.0.weight'
J
Xhow to make a lokr like this? with that net dim, usually kohya only allow to put 100000, https://civitai.com/models/292553?modelVersionId=1568463 (this is the lokr and the creator).
Eyou can add it in as an extra argument
Eor modify the gui
Xguidance please
Ehere i updated the file for you, replace it before starting the gui
you should be able to change the values of lokr now
the file to replace is in kohya_gui
you should be able to change the values of lokr now
the file to replace is in kohya_gui
lora_gui.py93.35KB
XI did replace, tho it's the same
Edid you restart your gui? are you sure you replaced the right file?
Esee in that file there is this
)
with gr.Row() as network_row:
network_dim = gr.Slider(
minimum=1,
maximum=9999999999,
label="Network Rank (Dimension)",
value=8,
step=1,
interactive=True,
)
network_alpha = gr.Slider(
minimum=0.00001,
maximum=9999999999,
label="Network Alpha",
value=1,
step=0.00001,
interactive=True,
info="alpha for LoRA weight scaling",
)
with gr.Row(visible=False) as convolution_row:
# locon= gr.Checkbox(label='Train a LoCon instead of a general LoRA (does not support v2 base models) (may not be able to some utilities now)', value=False)
conv_dim = gr.Slider(
minimum=0,
maximum=9999999999,
value=1,
step=1,
label="Convolution Rank (Dimension)",
)
conv_alpha = gr.Slider(
minimum=0,
maximum=9999999999,
value=1,
step=1,
label="Convolution Alpha",
)
)
with gr.Row() as network_row:
network_dim = gr.Slider(
minimum=1,
maximum=9999999999,
label="Network Rank (Dimension)",
value=8,
step=1,
interactive=True,
)
network_alpha = gr.Slider(
minimum=0.00001,
maximum=9999999999,
label="Network Alpha",
value=1,
step=0.00001,
interactive=True,
info="alpha for LoRA weight scaling",
)
with gr.Row(visible=False) as convolution_row:
# locon= gr.Checkbox(label='Train a LoCon instead of a general LoRA (does not support v2 base models) (may not be able to some utilities now)', value=False)
conv_dim = gr.Slider(
minimum=0,
maximum=9999999999,
value=1,
step=1,
label="Convolution Rank (Dimension)",
)
conv_alpha = gr.Slider(
minimum=0,
maximum=9999999999,
value=1,
step=1,
label="Convolution Alpha",
)
Xprobably, "kohya_ss\kohya_gui\lora_gui.py" this one right?
X
Fi trained LyCORIS. it is lower quality but has better multi concept training
Fyou can change our lora config LyCORIS however it may require some additional parameters which i dont remember
XI see, Im seeing this 76mb LyCORIS everywhere and at least for me they perform well, like creators use few number of images and get decent quality with strong flexibility
Edid you git checkout sd3 flux 1?
Edo a fresh install and replace it
KHello there I need your help here.I am training a character Lora with one trainer Is there Any math to find epochs based on dataset size? I have a dataset with 31 pictures of a character. How many epochs should I use?
KI am keeping repeats value to 1
I am using onetrainer
I am using onetrainer
Fdo up to 200
Fyes onetrainer has better reg images usage support
KI am doing Just like your video sir. I am puting it like that , repeat = number of dataset/ number of reg images
Fyes true way
Fso you should see number of steps = 200 x 31 x 2
K.
Kwhat is 2 here?
KI am using 4 batch value is this okay? I mean how can we find the aproximate epochs value? Is this all trial and error process. there is no pattern?
Ksorry I am so beginer
F2x from reg images if you use
Fif you use batch size 4 use square root 4 LR
Fwhich means 2x LR
Kthat means learning rate should be 2 times ?
Fyep
K2 times default = 0.0003 x 2=0.0006
KI see, and how do you get the approximate epoch value?
KI mean how to find the epochs value?
Kand isnt it very high learning rate?
