.jpeg)
So I learned while making the Draygon Style Dragons lora that making complex Flux loras on civitai is kind of half-broken.
So I guess it's because the t5 attention mask setting is turned off on the backend, and because it's not an option that can be turned on in Civitai's trainer settings. The way that AbstractPhila first explained it to me is that basically t5 interfaces with the training images and in a way it sort of serves as an inline regularization for the model as you go*, by keeping you from immediately and randomly trashing things as you train Flux. But without t5 turned on to do that then it's pretty much guaranteed that you'll progressively break Flux's ability to function properly (as modified by the lora) the further and further you go with it.
(*this is kind of an over-simplified way to explain it to a dummy like me :P, because I guess really it prevents you from writing a lot of garbage data into the model during training by keeping it focused just on the relevant bits, instead of constantly going back and needing to repair to damage afterward like regularization normally would, but anyway it is an easy way to think about it as just like an inline regularization.)
It only took me like 12,000 buzz attempting Flux loras on here before I found this all out... π’
So what this all means I suppose is that making Flux loras on Civitai sort of works technically, just in a half-broken sort of way, so any single basic Made-On-Civitai Flux lora might work okay when used standalone, and it will probably do what it is supposed to, and I'd guess that maybe as long as it doesn't get too complex or overlap too much, even a couple Made-On-Civitai Flux loras might even work at once without being too entirely bad (but probably with an increasing amount of poor quality outputs). But making big and complex, or multi-concept loras that won't fall apart (what I wanted to do), particularly when used in combination with other complex loras, or merged either with another lora, or a Flux base model, is pretty much a no go, because the damage to the model that is baked into the lora(s) will all compound the more and more that things get trained into it.
The other thing I've learned is that Civitai doesn't give great control over how Captions are trained into Flux loras, and by not "great" I mean that basically it's entirely random and uncontrollable as far as I can tell. My first Flux lora on here had long captions doubled up on each image and it broke them down into several short snippets each. Based on that I tried with Draygon Style Dragons(in an unpublished version) to make longish captions out of fairly straightforward easy to break up little snippets, but instead this time it just trained the whole long caption string for each image x-x;
Is it because I changed the dim setting slightly between my gingerbread lora and Draygon Dragons? Maybe something else? idk really... so let me know if you have any ideas xP
I don't know, but I'd guess the only way to have real control over the way the snippets are trained into a Flux lora with the Civitai trainer would be to manually enter them as micro-captions/longish-tags using (the totally not recommended for Flux) tag mode in the trainer. I'd try again using the tag mode, but I've already wasted too much buzz on Civitai Flux lora attempts right now, when I could have made a dozen or more SD15 or SDXL-based loras with the same buzz.
Since I can't afford the hardware that I'd need to do Flux lora training at home, or even to do it anywhere that I can't basically train for free, I'm probably not going to be making anymore Flux loras for now, and I'll stick with other base models. Especially, since the Made-On-Civitai Flux loras would always be kind of half-broken so now I'm not as interested on training Flux stuff on Civitai, because it would be too limited.
(Though I still think my little Schnelly could shine one day if given a fair chance.) π₯Ί
I'll still make images with Flux, and loras for other models, but any Flux loras are all on hold for it until something big changes either with Civitai's trainer setting, or my ability to afford hardware for training Flux. I'm sorry if this disappoints anyone as much as it does me, but I'll still make stuff with other models for now and I'm still open to hear any recommendations you have.
β¨ππ₯ππ§π π‘πβ¨