just putting the data set together, correctly, for a base model, is quite a task that requires exper
just putting the data set together, correctly, for a base model, is quite a task that requires expertise - never mind training the thing
BYeah dataset preparing and gathering requires a ton of time and resources
Cyes, and you're not going to just go scrape google for images - that's why most of them use the LAION library
Cbillions
Cof images
Cat least 2 billion if you're just making a tiny model
BI’m planning to build a SaaS product based on some ComfyUI workflows I’ve created, using Flux as the main image generation model. I’ve spoken with the team at Fal.ai, and they confirmed I can use it commercially through their platform, as long as it runs through their API due to their arrangement with Blackforestlabs. They also mentioned they can host ComfyUI and that it’s possible to use ComfyUI as a backend for a SaaS product, as long as all models used in the workflow (like upscalers, facial detection, etc.) are licensed for commercial use. I’m still deciding whether to stick with ComfyUI or build a custom pipeline based on my workflows.
PHey Dr. Furka, my friend is training a flux character lora (not realistic) with default settings of ai-toolkit and I try training with kohya and your config rank 3, and he seems to get really good results, but I'm having a problem getting the proper body shape and some details properly.
It seems like no matter what parameters I try, it feels undertrained.
tried 1 repeat with 200 epochs (20 images) lr0.00005
tried 20 repeat with 10 epochs (20 images) lr0.0001
tried 150 repeat with 1 epoch, (save every 1000 steps) tried the output of 2000, 3000 ,4000, and unless I make the prompts very long and detailed, I'm getting something that resembles the character and clothes, but the body shape and face mostly gets random every seed and it won't really capture the character.
Even if I put a lot of details in the prompt, it kinda works, but not really.
And yet, my friend's first attempt with default settings on ai-toolkit, which has much less parameters to change, gets awesome results and accuracy, and it's still very flexible while keeping the essence of the character pretty well.
Any idea what I'm doing wrong?
It seems like no matter what parameters I try, it feels undertrained.
tried 1 repeat with 200 epochs (20 images) lr0.00005
tried 20 repeat with 10 epochs (20 images) lr0.0001
tried 150 repeat with 1 epoch, (save every 1000 steps) tried the output of 2000, 3000 ,4000, and unless I make the prompts very long and detailed, I'm getting something that resembles the character and clothes, but the body shape and face mostly gets random every seed and it won't really capture the character.
Even if I put a lot of details in the prompt, it kinda works, but not really.
And yet, my friend's first attempt with default settings on ai-toolkit, which has much less parameters to change, gets awesome results and accuracy, and it's still very flexible while keeping the essence of the character pretty well.
Any idea what I'm doing wrong?
Pforgot to mention, we trained the same character
his captions are mostly automated with roughly manual fixes
my captions were automated at first, then highly reconstructed manually in a consistent captioning way
his captions are mostly automated with roughly manual fixes
my captions were automated at first, then highly reconstructed manually in a consistent captioning way
Psame dataset too
PRight now I'm trying to replicate his decent+ result of ai-toolkit, but with kohya_ss, just to have a solid starting point, but I can't manage to achieve that.
BAnyone here trained their own upscaler model before?
SAny recommendations on how to setup A1111 on AWS?
SOr Forge
Fyou dont need caption if your dataset is consisten enough
Fenable shared vram
Finitial loading uses VRAM
Fi trained on my rtx 3060 yesterday perfectly training
Fwell some people want to share and keep lora
Fonly benefit it is
Fotherwise original checkpoint best
Fdont use captions
Fwhen you are training a single concept
Fif your concept is consisten, captions reduces likeliness
Fmy installers should work there
Fof course i assume python cuda cudnn are install
Fand nvidia drivers
Pso what should I use?
instance prompt and class only?
instance prompt and class only?
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
Fyes
Fohwx + class
Pwhat about the outfit?
Pand even though the character is not accurate enough, I do find that my consistent captioning structure for the angle the character is standing and face expressions that are not the neutral ones works pretty well when I prompt for it
so should I caption those?
so should I caption those?
Por does it reduce likeness aswell
FFLUX uses image itself as embedding
Fso all is like auto captioned
Fi dont think you need to caption
Pand if the character is in a complex background, compared to some images with simple grey background, still not caption?
PIt seems like someone achieved great results and accuracy with captioning that looks like joy caption, and when he prompted for the example images, look at how short and simple the prompts are, and yet, with a very short trigger word he's generating Yasuke very accurately
https://civitai.com/models/459468/naoe-and-yasuke-assassins-creed-shadows-fluxponyxl?modelVersionId=919340
https://civitai.com/models/459468/naoe-and-yasuke-assassins-creed-shadows-fluxponyxl?modelVersionId=919340
PHow does one achieve this? this is what I'm trying to figure out
Peasy to use and high accuracy lora
Fwell if your background repeating
Fif you dont prompt background in inference
Fit will generate it
Fotherwise no issue
Fhave both characters in same image
Fwhile training
Fso it learns perfect both
Phe used
ss_learning_rate": 0.000002,
"ss_text_encoder_lr": 0.00005,
"ss_unet_lr": 0.00005
and
"ss_network_dim": "Dynamic",
"ss_network_alpha": "Dynamic",
what do you think about that?
ss_learning_rate": 0.000002,
"ss_text_encoder_lr": 0.00005,
"ss_unet_lr": 0.00005
and
"ss_network_dim": "Dynamic",
"ss_network_alpha": "Dynamic",
what do you think about that?
Phis lora file size is decent too 459mb, not too big