I used simpletuner to train, followed the guide on the site ^^

https://github.com/bghira/SimpleTuner/blob/main/documentation/quickstart/FLUX.md

Second this. It would be great to see the steps involved, learning rate, resolution, batch size, etc.

This is impressive!

Agreed. I'm amazed how quickly we've gone from "impossible" to this. Congrats on the first such lora.

thank you! i am very impressed with flux, i am going to train a few more models to see if the settings work universally and then i can probably come up with a guide with in the next couple of days ^^

Training guide has been released ^^

https://civitai.com/articles/6678

I'm very curious to understand if Flux is accurate on characters/celebs an so on. SD XL was more detailed than SD 1.5, but strangely, not very better on resemblance with celebs....

@settima_ai Yeah, my gripes with celeb loras on SDXL is that sometimes when generating on certain prompts you lose likeness. Haven't tried this lora yet but hopefully likeness is consistent.

Ana de Armas, Lily Collins and Ella Purnell next pls!

omfg didn't know she was one of us

Yes she is, look it up.

thanks for the feedback! I will retrain with the trigger Emma Watson

Could not disagree more. To me this is one of the worst myths in lora training. Using a name/word that the model knows something about takes away from the control you have over your lora and image generation. Now you're limited to always merging your lora with the model's version of that subject.

This can be good if:
1. your own dataset is lousy
2. you want to train in VERY few steps
3. the model already has a great grasp of the subject, in which case you don't need a lora in the first place.
4. you're not looking for high degree of resemblance

But usually, it's terrible. You don't know what images of the subject the model was trained with. They may not align with what the subject usually looks like, they may be poor quality, they may be of the subject when they were 12 years old. Whatever the model generates without the lora, that subject will always get blended in when you do apply the lora. And unless the model already produced a perfect subject, you're going to be holding the lora back, assuming your goal is to obtain high resemblance. In the case of Flux Dev, the base model seems to have limited knowledge of Emma Watson in particular. So you might get away with using 'Emma Watson'. But then if you try to apply the Lora to a different model/finetune that has its own different understanding of what 'Emma Watson' means, you're screwed.

Using a custom trigger lets you have complete control over the result of the training and the strength of the effect when the lora is applied. It is consistent. You can achieve great results with it, as I would think that this very lora here shows. Do not use the model's built-in understanding as a crutch when training a character!

@punzel Nothing wrong with trying, but I would strongly suggest that you not end up going that route in the future :)

@Blobbalisk whatever the model generates with "3MM4W4750N" will also get blended in when you apply the lora. And because that trigger word has no connection to female persons captions often are of the asinine kind of "3MM4W4750N woman". What happens when you prompt for a "woman" with a lora that was trained like this? You get only that specific woman.

You say it's usually terrible, did you test this yourself?

Thank you for the all the feedback! I had some time to retrain the model using the trigger Emma Watson, I am not sure if it's any better, please let me know what you think.

@mrneon yes, many times! And sure, I'm not a fan of the super complex 133T triggers either. It doesn't take much to be unique. As for what the model associates with a word that doesn't exist - I think nothing :P I don't know if at a fundamental level there's a difference between feeding unusable empty tokens vs no tokens at all, but when testing, I don't get any kind of consistency that I can identify if I'm just giving the model the same empty word/tokens.

@Blobbalisk Okay, well, do you see an issue with the new lora? does it seem to blend with the generic "Emma Watson" of the base model?

@mrneon I tried the retrained version now. It's a limited test, to be sure, but I find that the "Emma Watson" version does perform slightly worse. Not in every generation. (I think it's important to note that the base model doesn't have a strong association with the subject.)
Using ComfyUI right now my better results are with emmawats0n without even using the trigger word (but I'm not trying to make multi-subject prompts). With SD1.5/SDXL, I've found that captions are detrimental to character loras, because adding or removing any of those captions can dramatically shift the impact the lora and the balance of the prompt, forcing you to adjust the lora weight to compensate (this is not the case for some concepts, and definitely not for finetuning - in massive datasets, it would make no sense not to have good captions, I think). But in the long form prompt world of SD3 and Flux, the individual words present in the prompt which are recognized from the training captions might have less impact on shifting that balance. As long as they don't include a unique proper noun ;D

Anyway, best of luck.

@Blobbalisk thats a fact i just made a kobe bryant lora and used his name in the generation and it came out great

It was done with runpod using SimpleTuner

@punzel can yo teach me how run simpletunner on runpod?

Yes , Please do a short tutorial ;)

@wongkentiers theres one on youtube right now

@riiahworld share the link of the yt vid pleasee

here^^

https://civitai.com/models/632945

@punzel Ty <3

I think you have to update ComfyUI and avoid using multiple Loras.

Update your comfyui, I had the same issue yesterday it would just not trigger the lora but after update it worked

Lora not forge compliant it seems ?

@phoebus_photos410 It's not, I tested a few times and it just crashed the Gen. Still, early days yet.

Maybe not working for dev gguf Q4? I've tried few times it doesn't work

Training guide has been released ^^

https://civitai.com/articles/6678

trained on 25 images for 1600 steps at 5e-4. hope that helps!

@punzel do you have civitai lora training settings you can share