This has the same license as the 0.9 SDXL, so currently this is for research purposes!
Hello!
Since v1.0 is just around the corner, I would like to show you that training people LoRAs is very much possible.
Bear in mind, that those were tests so the quality might not be the best (also, the base is still v0.9, and v1.0 should be even better).
Samples might not be the best, I haven't used any refiners, just pure SD XL model and a LORA and nothing else.
The training dataset was of 1024x1024, so there was no upscaling (yes, freshly recut samples).
There will be still a lot of room for tweaks and improvements, you can compare it to the first dreambooths/loras. Still, as those were first tries - I'm positively surprised!
Here are three models that I was able to train during the beta. Enjoy!
I'm creating Dreambooths, LyCORIS, and LORAs. If you want to know how I do those, here is the guide: https://civarchive.com/articles/7/dreambooth-lycoris-lora-guide
I upload 2-3 times a week, however, I write down suggestions all the time. If you want to see someone made into a model just leave a comment :)
Suggestions left on my donation page ( https://www.buymeacoffee.com/malcolmrey ) are taken with the highest priority and you will be able to enjoy them way before they get released here. (I accept also custom requests)
The reason is that I can't keep up with uploading and the models are being made pretty much all the time :-)
As mentioned, via the donation page I accept custom requests, so feel free to engage with me there, I will answer all your questions :)
Description
FAQ
Comments (21)
Here are some test loras so you can have something for your tests to prepare yourself for the release of 1.0 which is in less than 2 days!
I've been using ComfyUI
Question:
Old LORA for model 1.5 can be used on SDXL?
You cannot use old LORA for inference using SDXL.
You can however use a 1.5 model as a refiner and then you can use old LORAs.
Also, -> old TIs seem to work with SDXL without any changes (!!)
@malcolmrey wait, what? TIs? Even with the new SDXL text encoder?
@Erilaz this is what the experts say, I did not test it personally, but others have.
I've seen some samples and it looked promising!
@Erilaz Second question: Is SDXL censored? If yes, dont you think it will be rejected like SD2.1 ?
@settima_ai It is not censored, but it is also not trained for it; expect something like 1.5 started with but with higher base quality and fewer artifacts like mingled body parts
HOWEVER: SDXL seems to be trainable very well and there are dedicated people that are already working on it as far as I know. You would probably need to ask people like hassan for details.
If you follow SD subreddits - you will find people who already posted clipdrop samples that showcased some of the finer details.
I think there is a long to way to go until we see something nice in SDXL. I prefer to stay in "old" SD 1.5 by now! Thanks four your always welcome efforts"
long way? SDXL 0.9 alsready has pretty awesome outputs.
I think we won't have to wait that long for some custom models to appear. There is tons of experience with making those. There are already lots of datasets so they just need to hook it into training (scripts are already there). I'm pretty sure some cool LORAs will drop this week already. As for the whole checkpoints - I would not be surprised that within 2-3 weeks we will have something very decent.
The only problem I predict is that fewer people will be able to use it locally. As for the cloud/colab stuff - it will require a higher tier. But we will see, there were some optimizations done for 1.5 scripts weeks (days?) after the initial dreambooths/loras repos dropped so I have high hopes someone will come up with something for SDXL too.
@malcolmrey That's the point! There are just a few persons with gpus capable of running those models in local and I don't see any spectacular improvement besides more resolution. Anyway I will wait some weeks ;) Thanks again!
@astragartist I'm running SDXL on a RTX 3060, which costs like 300€ and has 12gb of vram, it generates images really fast and I can do high resolutions. So the barrier of entry is not exactly very high.
@TestTTtTtT I have the exact same gpu but waiting nearly one minute or even more to get similar results as the ones that we already got without waiting too much and with a lot of useful tools... If results were similar to Midjourney, it would be a huge upgrade, but I don't see it ... yet ;)
Waiting until Realisticvision model based on sdxl releases
Me too! :)
This is the way. Along with a few others that have become my go to models.
It's a great start for something that is only tests and on an unfinished base model.
Thats very cool that youre already working on it!
I will say, the outputs look interesting... Theres some odd deformities in their face shape and eyes look strange. But very promising start!
I agree, the results were good enough to share, even with those interesting quirks :)
What I got earlier would be way funnier/scarier but I'll keep that to myself :)
It was also fun to see what others came up with and oh boy, those were funny too :)
Love to see SDXL work! These look pretty off, though. I understand they're beta, but it's interesting to see. I'll be doing some experiments with training in XL soon, and I'd be interested in learning your discoveries, as based on the results it seems like the settings required may be quite different from what I'm used to in 1.5.
I will post tomorrow my findings/settings. I'm sure as a community we will figure out quick enough what will be the better settings ;)
When making those I was happy to see that the training is better than in 2.x. Didn't have much time or compute units to make comprehensive tests, but those were good enough to see that the model is indeed trainable.
Also, during that time the training scripts were also in development, so the results might actually be better on release even with the same parameters! :)
Looks like we don't have an active mirror for this file right now.
CivArchive is a community-maintained index — we catalog mirrors that volunteers upload to HuggingFace, torrents, and other public hosts. Looks like no one has uploaded a copy of this file yet.
Some files do get recovered over time through contributions. If you're looking for this one, feel free to ask in Discord, or help preserve it if you have a copy.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.