Sorry for the tease lol, full flux training isnt available yet, once it is there we'll be a bunch of stuff cone out, maybe to make pony look simple

@cathyleverman I did experiment with training an upscale only flux lora to use in conjunction with pony, seemed promising

@cathyleverman unfortunately there's still a lot of errors and I'm super busy with a lot of other stuff. It's not ready for primetime, but I think a pony-image-trained flux lora specifically for img2img use after a sdxl pony model would be powerful

@cathyleverman unfortunately we won't see anything good until full text encoder training becomes available. The issue is that inter-relating concepts simply can't be learned or even merged without that - that's why we seem limited to one lora at a time. It could be possible to have a multi-stage flux gen that used a different lora for different steps of image creation (3d position, basic details, photorealism, etc).

@cathyleverman I wasn't involved in actually making Pony, and the issue isn't resources, it's that huggingface and the flux developers haven't released the ability to fully train the model yet. Anyone serious about model training is waiting for that to happen (or for the next best thing to arrive that allows full training)

@cathyleverman I'm a software developer but I don't have the time, experience, or access needed to rig up a script to support training both text encoders in flux. I've made some efforts to use just clip but that wasn't great (the two text encoders are meant to be trained together). Maybe I'll take another look because they're supposed to be similar to sd3's encoders, but I never did any sd3 training.

I'm sure smarter people than I are working on cracking it lol

@cathyleverman I already dropped $50 on runpod with a H100 testing. 'de-censoring' implies there's nsfw content in there to unlock - trust me it's just ken and barbie nothing lol. In fact, they've actually poisoned nsfw words so they're trained to make sfw images (so it's easier to use new keywords for many things if you actually do it). 'unlocking' is pretty far from what needs to be done. 'teaching everything about birds and bees and parts and nuances of all of that' is more like it.

Having said that, The reason I have several different loras here for different angles/situations is that despite being somewhat good at this and having some horsepower and money at my disposal, combinations of new concepts just aren't sticking without the ability to train the full model. Anyway, I digress. I won't be the one to solve it, I have way too much to do to throw more time and money at something that is just leisure activity for me. I can't even throw this nsfw stuff on my resume lol

@cathyleverman  12k steps in, and my flux nsfw img2img model is starting to perform pretty well. Tweaked training to focus on img2img, using zonkey 6.2 hyper for txt2img, then my custom flux lora for img2img for a pretty quick gen time of about 70 seconds (using vanilla flux dev 8bit - I'm sure it could be improved but I'm training on the base model for compatibility).

https://civitai.com/images/35719201

https://civitai.com/images/35722254

https://civitai.com/images/35722107

https://civitai.com/images/35723833 (still improving at 13k steps)
https://civitai.com/images/35724166

I appreciate it! This was a lot of testing and work, plus so far $50-75 in runpod costs to use the h100 (although some was done on my 3090ti)

Thanks for the kind words! Flux can't be trained fully because text encoder training isn't available publicly yet. That means you can't teach it many concepts at once (or even combine different loras that do). I will consider some custom scene loras, but I'm working on something else promising - a flux nsfw upscaler lora that can support a lot of different scenarios.  https://civitai.com/images/35719201 this was generated using an 8 step pony model (zonkey 6.2 hyper in 12 steps) then img2img with my flux nsfw upscaler lora-in-training for 40 steps 0.25 denoise. Total gen time only 70 seconds on a 3090ti (and probably capable of much better times on a faster flux model - i'm on vanilla flux 8 bit)

https://civitai.com/posts/8162675?returnUrl=%2Fmodels%2F405136%2Fzonkey-realism%3FmodelVersionId%3D891259

https://civitai.com/posts/8162723?returnUrl=%2Fmodels%2F405136%2Fzonkey-realism%3FmodelVersionId%3D891259

https://civitai.com/images/35723833 (still improving at 13k steps)
https://civitai.com/images/35724166

Some more examples

@portraitartist999132 https://civitai.com/models/771009?modelVersionId=991565