LoRA for Zatanna from Young Justice. Trained on AnythingV5.
Used Hi-Res Fix (Upscaler: 4x-UltraSharp) 2x —> Extras (Upscaler: 4x-UltraSharp) 2x
Recommended weight between 0.6-0.8; can go higher or lower depending on additional LoRAs, etc.
basic outfit trigger: black jacket, bowtie, white vest, fishnet pantyhose, top hat
Description
FAQ
Comments (7)
I still can't believe how you make really good LoRas. Do you use the "usual configurations" (512 resolution, 128 dim, 1e-4, 5e-5) that everyone recommends?
I'm making a Cheshire LoRa to learn the ropes, and I am reaching a good point with the epochs/repetitions to avoid overfiting, but nothing to your level of quality and flexibility.
Amazing job as always, I wil try to leave a review soon.
Haha thanks, I can see if I can upload the Kohya_ss config file. I followed HoIostrawberry’s LoRA guide on here to reduce the file size by setting Network Dim & Alpha to 32 & 16. I think others were Unet: 5e-4, text encoder: 1e-4, scheduler: constant, but the guide has them.
I try to get between 50-60 images (official, fanart, and different styles)(Zatanna one was like 80 tho), and now use the WD14 captioning in Kohya to generate the txt file. Most images I try to find at least 512x512 or higher, then use the “enable buckets” option in Kohya.
The WD14 captioning does a pretty decent job, tho still go through and add the color for clothes/background if it doesn’t mention it. Also do a find & replace to delete any “physical attributes” tags (hair/eye/skin color) and put the “trigger word” (ex. Zatanna) in all of them first.
So what I’m doing now is usually, X images * 10-15 repetitions * 10 epochs / 4-5 batch size. I also have a really beefy PC (12700k/4080) so it takes like 10 minutes or so to create the models.
@novelProphet Thank you for the tips! I will work harder to make a better lora with your tips.
@Rlord Np. And if your GPU isn’t the latest one out there. Think it’s advised to keep Batch Size at 2. Otherwise may run into a OutOfMemory error.
Also, could enable “gradient checkpointing” and “memory efficient attention” if still having errors, though it may take longer to create. Good luck!
To me it's funny that I was just thinking of doing a quick Zatanna project a day ago, searched today for a LoRA to make it a bit less laborious for me, found this as the SOLE Zatanna model, and it's only 13 days old. My point is the sheer coincidence/timing. If I'd searched BEFORE you uploaded, I may have given up on Zatanna, OR, if you had shared her AFTER I searched, I may have moved on, too. (I may've just trained a LoRA on my own, but that would've defeated my original purpose of avoiding the labor-intensiveness of it, lawl.) So thanks, really, for your hard work AND your golden timing, author-sensei. I'll make sure to check out your other works later.
BTW, I'm really new to the AI world, as in just a couple of weeks old, but I really like your prompt style, it's one of the HANDFUL that really fits my... logical way of thinking, especially the way you handle outfits & other items often ignored by others. I'm super-OCD regarding them, so I've probably memorized a million of 'em, even at just this early point in time. So I hope you'll forgive me for copy-pasting into my brain your entire set of examples & their prompts; they're all smart & super-cool. Cheers. ❤
Thanks! Yeah I just got into it a little over a month now, but it’s kinda addictive haha. I’ll probably do a couple more, then come back to it later. I’m sure the tools are gonna evolve like crazy next couple months.
So true. Some of the old veterans are telling me the tech we can use now were simply unimaginable when they 1st started out last year or something. We'd probably be saying the exact same thing to our "kouhais" in months, if not weeks, heh.
And yes, when you have the time, please do continue. I can see we have the same taste in characters and aesthetics, or at least mostly. I'd be glad to see what you can cook up, whether it's newer or vintage stuff. Stay cool mate, and cheers! ✌
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.