The WowifierXL is a LoRA which has been trained on a variety of AI art made by "mooncryptowow" at 1024x1024 for Stable Diffusion XL!
Demo images were rendered at 768x1024 with hires fix and no refiner (I haven't really figured out how to use it in A1111 yet :) )
Check out my art or support me on Patreon: https://linktr.ee/mooncryptowow
Description
The wowifierXL is the latest version of the "wowifier" by mooncryptowow. This version is trained on a variety of my AI art at 1024x1024. All demo images were rendered at 768x1024 with light hires fix and no upsampling or refiner model.
Feel free to check out my art, or support me on Patreon: https://linktr.ee/mooncryptowow
FAQ
Comments (33)
Honest question, how long did it took for you to train? Considering 4x resolution compared to default 1.5 people are suspecting that it would take 4x to train SDXL lora with same parameters
I'm sure it's optimized. The inference isn't 4x slower, so.
Oh it did. I think it took about 3.5 hours to train this for 2 epochs on my 3090. I've trained my last few 1.5 LoRAs at 1024 this took substantially longer.
@mooncryptowow were training steps similar process to 1.5? just longer using this base model?
@dpa7026559 I used mostly the same settings as I used for the wowifierV3 on 1.5, the biggest differences was that I needed to reduce the batch size to 1 or it would run out of memory and I had to switch the optimizer from adamw8bit to adafactor.
@mooncryptowow how much vram does the 3090 have? I'm using a 4090 and am not sure how low the batch should be for 24gb vram
@hobolyra also 24GB.
@mooncryptowow Thanks for your work & thanks for this info. I also have a 3090, have trained some of my own Lora's (in SD 1.5), and was wondering about making the switch to XL. Early trials with SDXL 1.0 just generating images with my 3090, using no XL 1.0 refiner, running in Auto1111, I couldn't generate at 1024 x 1536 with a 2x upscale. It would hang. If I reduced to 1.5x upscale (1024 x 1536) it would complete. Some images were gorgeous, some still with scary deformations. Overall, I don't think there'll be a high adoption rate to SDXL just yet. There's a large base of people on 1.5 who just don't have the computational oomph to make it worthwhile. I'll do an install of ComfyUI to mess with XL, but I'll also continue to work mostly with 1.5. in Auto111. At least for now.
@cap284 I've been having a bit of a tough time with it in A1111 too. I have to run at --no-half-vae or it'll produce black images, and it often runs out of memory and hits the page file on a single 896x1024 render! I've yet to try comfy, but I think I might have to buckle down on the weekend and try it.
There's a lot to be excited about with SDXL, but there's some optimization that needs to happen in a big way, especially in terms of workflow in A1111.
Okay, only now I am noticed. 1.3GB for Lora. That's huge.
@squirz It's huge I know, I checked some of the other ones on the site to make sure I'm not totally out to lunch and most of the other ones online are a similar size. It's certainly something I'm looking to address with the next version though.
mooncryptowow obviously came from Disco Diffusion. :P
Awesome to see XL lora already being dropped. Appreciate it. 🙏🏻💙
Superb, I love this LORA, it really adds a colourful puch to the images.
Hey there! First, thank you for any amount of work, large or small, that you've put into this. I read the other comments re training time vs 1.5. I have a 4090 24gb, and your model size - while definitely quite large, is not a concern to me.
I am curious about possible overtraining. I have done a few tests with different base models, and so far come to the conclusion it's your lora. I still could be wrong, as I just recently began testing out SDXL, and would readily admit there's something else.
Also, I'm using your model to produce music videos utilizing A1111 + Deforum. Funny thing, I have absolutely zero problems with any of the issues you mentioned in other comments about A1111. Maybe the commit I'm on fixed something...
At any rate, what I'm seeing might be caused by overtraining. It happens regardless of which base model, prompts, or other settings, which leads me to suspect your lora.
The behavior is one in which portrait-type images are always turned facing about 45deg left. Never facing the camera at 0deg, and never at 90deg either. Always about 45deg and always to the left.
What's also very weird, is this only happens at the very beginning of a Deforum video render. Once more about 100 frames have been rendered, then faces appear in multiple various angles as expected. Only at the beginning does this left-facing hallucination occur.
Rather than simply complain, I'm trying to further understand the possible technical explanations for this, and also offer up any testing to help identify and isolate if in fact it is the lora.
My inner nerdy geek-self, along with my IT/developer profession really has me curious about this. Please accept my apologies up-front if this comes across as anything other than sincere.
Awesome work, and yet another reason I would like to figure out how to avoid left-facing portraits.
I actually didn't even realize that Deforum was able to use SDXL yet, all of my deforum experience is based off 1.5.
The XL version of the wowifier is slightly over trained, but it manifests more in over saturated colours than forcing a composition in my experience. The training data is a good mix of different compositions and zoom levels / focal lengths, so it SHOULDN'T fixate on a single composition. Though I've seen SD do weirder stuff than that in the past so I don't rule it out at all.
As for the deforum rendering differently at different points in the video that likely has to do with your noise schedule options in deforum.
I'm still not happy with my more recent attempts at this. I'm still working to find a balance where it's trained well enough to get the results people are looking for, but isn't over trained. I'm also not bothered by the size, but I do have V2 down to about 512mb right now.
@mooncryptowow Thanks for this reply. I've recently discovered that vertical portrait aspect ratios produce much more variety in the composition with faces appearing in many different angles, and I'm most often working in HD landscape (widescreen) dimensions. So, it's likely more training data is portrait vs landacape.
I really push the limits on Deforum, and as soon as I saw dreamShaperXL + Wowifier XL together, I knew I had to give it a go. So far. other than the left-facing portrait reoccurrence, the result have been stellar. Again, thank you for your work and generosity producing your model.
Ahhh yes, noise schedule in Deforum.... Depending on many factors, that is tricky to get right. In fact. Deforum itself can be quite daunting, but I think I've tamed the beast and have it mostly dialed in.
Sounds like me, where perfecting something is actually part of the pleasure of doing it altogether. I look forward to seeing V2 come out whenever it's ready. I'll happily provide some feedback as appropriate.
In case you're interested, here are 2 of my recent music videos on YouTube that utilized your lora in combination with that ckpt I mentioned and another lora rendered with Deforum: https://youtu.be/xPGisQZWra8 & https://youtu.be/WZdvNpI4i8Y. PS, the latter link will not be publicly released until 8/19 6pm CT US, but it's also the better of the 2 IMO.
Thanks again, and keep up the awesome work!
@craggins wow those are so cool! Do you mind if I share them on my socials? Hook me up with your IG/FB if so.
@mooncryptowow Apologies, I just now saw this... YES! Of course you may refence any of my work, much of which (at least recently) uses your LoRA in some part or another. I don't use IG nor FB, only YT. Considering starting an IG account, but haven't done so yet. I gave you credit and a mention on my most recent video I published the other day. Check it out!
Love it tho I do wish the trigger word was something more related to the name of the lora, like wowifierXL because I have to look it up, every, damn, time...
Edit: Ooof I apologize for opening up this can of worms, as it was just a passing observation, had I known people were going to make into a such a negative thread I wouldn't of said anything.
Hope this doesn't tarnish any future creations you may feel to share with us.
haha I've considered changing it since the first version. Initially the idea for this LoRA was to add myself as an artist to dataset, so that's how the "by mooncryptowow" came about, that way I could just include myself with the other artists in the prompt like normal. Over time it has made less sense, and I may change it to like "w0w1fy" or something in a later release.
oh, and worth mentioning, if you install the Civit plugin for A1111 it will add a default weight and all the trigger words automagically when you select the LoRA.
i manually append trigger words to the filenames of LORAs, and it low-key infuriates me that other people don't hold to a similar naming convention?
example: "WowifierXL-V2'art by mooncryptowow'.safetensors"
@blncd I just make annoying comments on all of them, if I have to be bothered, they can too.
@MysteryWrecked so you feel like annoying the people who make the tools you enjoy is the appropriate way to vent your frustrations with having to ... checks notes... rename a file to suit your own workflow better?
Honestly I would hate that naming convention as I keep my filenames as short as possible.
@mooncryptowow Having some clue of what something is when I see it in my files doesn't just "suit my workflow better", it literally tells me what it is and how to use it. Now that Wowifier has a proper name, I know exactly what it is and how to use it at a glance. Are you someone that has photographic recall of every file you may download to remember what it's for without any hint of it in the filename? I envy you.
@MysteryWrecked I use an extension that puts in the default weight and all trigger words as soon as I add the LoRA. I would never make demands of the people who do work for me for free, and frankly interactions like this are why I release like 1 in 5 things I make.
i agree - in literally every other circumstance, i, too, prefer to keep my filenames as short as possible, so yes, it is in fact quite annoying.
i have my own reasons for not 100% trusting it to any such extension, and for preferring to keep track of things in the way that i do.
if it works for you, that's great, but - aside from just the trigger words - some of us don't particularly enjoy having various, often seemingly randomly-named files laying around and taking up space on their hard drives, either. ('wowifier' is acceptable in that regard, as it tells you precisely what it does; and i find it hard to blame you for naming your trigger-word after yourself; so what am i left to complain about? nothing. i'll just rename the file, and move on. it does, however, get rather annoying when i download a bunch of LORAs at once)
but what if i've shared my catalogue of checkpoints + loras with a friend who does not have internet, by carrying it to his house on a usb drive (which is indeed a thing i have done)? the extension cannot function in that case. if my internet is turned off that day, and i just reinstalled my entire webui because of some A1111 update? or what if the extension temporarily breaks? stops being supported? what if i find it on an old hard drive in 20 years and still want clear filenames so that i know what it all does (which is indeed a thing i have done with plenty of software)?
i think you're both being the tiniest bit more abrasive than is necessarily required, but hey, that's what the internet is for! unfortunately, most standards only improve when people talk about them, and most of online interaction tends toward complaints. so keep on raging, i suppose. ❤️ ✌️ ☮️
edit: this was actually the last LORA i was prone to complain about this on, as it still seems to "wowify" the output if you completely ignore the trigger-word
@blncd My comments weren't directed at you. I felt you brought up a point respectfully and didn't view the interaction as negative at all. It was MysteryWrecked I was responding to.
mysterywrecked "started it", yes, i agree. i knew who you were talking to, i was not offended. sorry for my own part in 'going over the top', but i felt the irrational urge to chime in
@mooncryptowow Hey Mooncryptobowwow, I dig your LoRA. This extension you mentioned for automatic triggering and default weights, can it (or are you aware of anything like it that can) be used in ComfyUI?
My trigger word would always be "boobs". For every single Lora. Or maybe "bobbies"¿ Ok it would be variationa on bewbs. Or how about B( . )( . )B¡
Oooh, or I'd make it horrid! Like, "imatouchmyself"
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.