This LoRA is specifically trained for generating rimless round glasses.
Yep, simple as that. I trained this because it seems like Flux dev is somehow unable to generate frameless glasses just by prompting. And I did a bit of googling...and found that there actually aren't many photos of people wearing rimless round glasses online, so I figured...well, I could train one myself.
I'm still practicing training and experimenting, so this might not be as flexible as some popular LoRAs out there. But I did a lot of testing.
You can use this on men, women, animals, plushies, any subjects you can think of.
You can also use this with other style LoRAs or concept LoRAs, like the amateur photo LoRA or flux realism LoRA.
You can use this on any image resolution, I tested on 1:1, 7:9 and 13:19, and they all work fine.
Description
FAQ
Comments (10)
Is it possible to use this without COMFYUI? I'm using FORGEUI locally without success.
What checkpoint/unet are you using? This LoRA only works well with FP8 precision and up (both GGUF and original dev unet with FP8 precision, or just FP8 precision checkpoint), but if you're using NF4 or something it will not work.☹️
@chubbypillow93 Thanks for the note. You're right, it was the NF4 checkpoints I was using, but I just tried FLUX_dev but still seeing gold rims. Here's my prompt, if you get good results, please share which checkpoint you used. Such a sweet looking lora.
Asian girl wearing rimlessglasses (sideswept pixie cut:1.6), looking at camera, near a beach, 1980s style <lora:Rimless-FLUX-LocalTest1:1>
@INFRASIGNAL Hmmm...I just tested it myself, with several prompting method, but none worked on Forge, it's really weird. But a lora weight of 1.2 (and above) finally worked though. Not sure why this is happening...since on ComfyUI it works every time with the default 1 weight. https://imgur.com/a/NGJwKKd I generated these with Forge, the first one with 1.5 weight, the second one with 1.2 weight. Sorry for the inconvenience! I never realized different UI can make such huge difference.
@chubbypillow93 Fantastic! Increasing the weight worked for me as well. Much appreciated. Having just migrated from A1111, there's still some settings in Forge that need tweaking, but I'm stoked this lora works beautifully now. Could you point me in the direction of how you trained this lora? I have a number of photos of a particular hairstyle I can't seem to prompt well enough for consistent results.
@INFRASIGNAL No problem! So um, I packed up my dataset and captions into a zip so you can have a look at my caption style: https://drive.google.com/file/d/1dKjqRZ83kOk8AbglTzfJf5vva9Pjk-iE/view?usp=sharing. I didn't put in too long captions because when I'm using a very small dataset (like for this one I only used 20 images), the training may have to be more aggressive than training on a larger dataset, or else it won't be able to learn properly on all the images at the same time (not sure if I phrased it right lol).
And you also have to consider how are you gonna use it in the future, for example maybe in the future you'll only enter prompt like "a photo of an Asian girl with {triggerword} hairstyle, she's wearing a blue dress on a sunny beach", then you probably shouldn't use an extremely verbose prompt. For an object or hairstyle, it's a very specific "small" concept, hence you don't have to use the same method as training an art style, I believe. About the configs you probably have to adjust them according to your result, I normally use 1e-4 as my learning rate to bake it slowly, each epoch probably ~150 steps in total, for about 8-10 epochs.
@chubbypillow93 I am pleased to report that by using your training data as a guide and following your example settings, I was able to successfully train my first Flux lora on Civitai! I chose 9 epochs and it's interesting how the first and last samples are outliers to the other epochs in terms of prompt accuracy. But the overall quality is unbelievable. BTW, like your sample training images, I also used varying dimensions and trained at 1024 resolution rather than the default 512. Thanks again for your expert advice. I am grateful.
@INFRASIGNAL Great!!! :D I'm happy to hear that. Yeah different aspect ratio is very important in curating the dataset, I found that when I'm only using one resolution to train a face, sometimes the neck would be unnaturally long (or something else that is weird), but multi-res doesn't cause this problem. For my glasses LoRA it's more like a coincidence though, it's quite hard to find this kind of glasses' image online so I had to crop the images according to their original ratio XD
Draw Things doesn’t like to import this file either. Might be a bug in DT though.
Edit: It was. Draw Things works perfect with this LoRA. It just doesn’t recognize it automatically, but that can be solved by telling DT manually that this is a Flux LoRA
Works well with Jib Mix v6.1/Forge
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.