Lets sum it up, style training on Photography with a bigger database via Loras is a pain, chose the version you like the most. I won't spent more time on this :D
In general training styles work best on 10-20 images max, singlecaption with mixed no caption, or no caption at all seems also very good, but yeah small datasets everything else should be finetuned
Description
FAQ
Comments (17)
gonna use it a lot (if it helps to improve my test gens)
who knows, its a pain to test it on my setup with nf4 model in forge :D Hopefully it does anything good
@epinikion simple: I'm comparing all day loras with og gen results, and so far - Faeia's details lora has amazing impact on photorealistc results too
Forge + nf4 ...
@OliviaRossi Good to know, will try out probably if I have time for that and the Lora patching in forge is more advanced
@epinikion your lora is not working in forge (yet) while Faeia's doing some magic ...
@OliviaRossi Did u update to newest version? It was working before the update with the comfy conversion script but as I tested it now it didn’t anymore but the Lora parsing did work. I put it as experimental until the Lora stuff is solved for flux
@epinikion I did the damn update, I hope tomorrow Forge will fix it
It is not working in Forge, atm.
I trained my own Flux LoRa via Civitai a few hours ago and got the same problem / error during LoRa Patching.
Unfortunately, there was no indication in the instructions that there are problems with these LoRas in Forge.
Does anyone know anything helpful?
Lora only work with dev / dev fp8 in forge as far as I know
It works for nf4 but you need A LOT free space on the drive to not interrupt the patching process
@epinikion Thank you for your reply! So is Forge patching on HDD not RAM?! Can you estimate "A LOT" ? 10GB ? 100 GB ? 1 TB ? - and which drive needs the space? system drive? forge root drive? model drive? temp drive?
@AvaCadava For this LORA it was eating about 23GB of Diskspace don‘t know why exactly. I guess it is building some kind of an own UNET model. The Drive should be the root drive where your python cache is located, but I don’t know exactly since my forge installation is on the same drive. probably it is the VRAM
@epinikion how does one get the ui to use ram instead of vram as that's what I have more of
@epinikion @AvaCadava Set "Diffusion in Low Bits" to "Automatic (fp16 LoRA)" and it should work like a charm. No slowdown when using loras anymore
@SubtleShader That actually helped so much, but the LORA isn't affecting the image. Is it only working with dev fp8 and not the bnb nf4?
@SubtleShader Do you think it makes sense to put it back to normal when using LoRas that do not require this fix? Does it slow something down?
@OlliFlysTheRocket When not using this setting, loras are not applied for me now.
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.