An ATTEMPT at a LoFidelity (not VAPORWAVE) Aesthetic / Pastel style Lora. This is now the final version of this, and other concepts similar to this are coming separately soon!
Designed by Ktiseos Nyx.
We've been at this since January of 2023! If you're interested in supporting us in our long standing AI journey please scroll down for ways to connect & sponsor us! Currently if you're interested in having it on the generator please feel free to donate towards auction costs!
✨ Recommended Settings
To get the best out of ANY of our models, try these settings:
Samplers: Euler or Euler Ancestral *(SDXL) - Or your preferred DPM style for SD 1.5 - We don't have settings for Flux as we don't gen flux off Civitai just yet!
Scheduler: Karras
Negative Prompt: Generally not needed! The model is designed to perform well without one, but feel free to experiment based on your preferences. (*For SDXL Bases) - For SD 1.5 Bases: Dont' do drugs, beacuse you'll need them XD
HiRes Fix:
Steps: 15
Denoise: 0.3
Upscale: 2x
Upscalers: Rex, 4x_Remacri, Lollypop or your personal favorite.
ADetailer: Use sparingly. Due to potential VAE differences, it might occasionally produce grey squares.
Embeddings: We've used some of our own custom embeddings to showcase character types in our examples, but no stylistic embeddings were necessary to achieve the core look.
Checkpoint Compatibility: Depends, not everything will behave.
🛠️ Training Details
If you're curious about what's the training details in full make sure to use this tool: https://xypher7.github.io/lora-metadata-viewer/
🙏 Acknowledgements
A huge thank you to the incredible community members whose work and inspiration contribute so much. Special mentions to:
Arthemy, Wai, Holostrawberry, FallenIncursio, RichyRich, JustTNP, Hoseki, CitronLegacy, Samael1976, Kairen92, ShadowxArt, Flyx3, YeiYeiart, LazyVisionary, and so many more! We try to train as much as we can ourselves, but the community's creativity is a constant source of inspiration.
🌟 About Ktiseos Nyx & Connect With Us!
Ktiseos Nyx is available for commissions! Find us and more of our creations here:
Discord: https://discord.gg/HhBSvM9gBY
Our Other Socials: https://beacons.ai/duskfallcrew
Sponsor Models & Commissions: https://ko-fi.com/duskfallcrew/commissions
TensorArt Space: https://tensor.art/u/611011406535381539
Dataset Tools: https://github.com/Ktiseos-Nyx/Dataset-Tools
💵 Training Costs Money!
While Civitai style trains usually cost buzz, testing and infrastructure to create better quality content to service our customers long term does cost money.
We're not looking to be the next Civitai at all, but AI has been a format of income outside of the inability to find regular work.
Donations, Commissions and Sponsorships are the backbone of our work!
Buzz donations work to keep our models on the generator that YOU the user choose! (This was originally a rarity, buy recently you've all chosen and we're all proud!)
Lora training can take time and effort, and while we do have cost layouts - if you're wanting a simple character lora you can use the Ko-Fi link shown in our link In Bio Above!
✨ Commission Details
If you're wondering why you're seeing a style lora or other loras ahead of your queue'd training or commission please see these FAQ lines:
Style Loras are usually developed independent of our normal training queues at the moment, as we use Civitai's platform to train them for use with our normal model making schedule during the month.
Character loras normally are trained as they come in, these semi depend on our own training service we've developed. Right now the NextJS system is almost complete for testing, and we will be resuming PAID and Free queues ASAP.
Concepts that are CLOSE to "STYLES" are again created using Civitai, at least on the short term. Once our main service has been completed we will amp up our schedule.
Description
FAQ
Comments (17)
sometimes it works, sometimes it doesnt. upto 50% it shows photo. Tested on a number of models, same result.
IT would highly depend on the model you're using
@duskfallcrew kinda guessing for quick releasing, u did lower dim and alpha? instead of 128, 128 u must have put down to 32,16? it makes a very unstable Lora. Suggest to increase to at least 64, 64.
@Mytee I'm happy to re-train it at a higher DIM, no problem - XD I still have the data, it IS A LITTLE unfocused and it wasn't about quick releasing i'm sort of still learning things - so this is good discussion here - NORMALLY i lower the settings because the notebooks i use claim it's better to do so - THIS specific lora IS A LITTLE buggy, i'm aware it is and i'm sorry I forgot to note that - I do want to re-train it a bit because when i fused it to a checkpoint it didn't really work well. It isn't JUST the DIM, but you're right the lora is a little unstable and i SHOULD in theory rename this to beta <3. I DO thank you for bringing this up <3 Are you on discord? If you are would you like to be a beta tester for this stuff, or even if you're not is there a social media i could get ahold of you on for this sort of thing?
@duskfallcrew the dataset is wonderful. love the style. love love love the style. can test, no sweat. Mytee#6953. found most if great Lora come with 40 repeats, 2 batches, 4 epocs, dim,alpha 128,128. bucket on, should take less than 1hr. cheers.
btw, somehow loha can do 64,64 but very stable style. found loha is much stronger network than lora from experience.
btw2, in general no need trigger word for style. ;-)
@Mytee Here's the discord I run that's for content creation & Ai: https://discord.gg/Da7s8d3KJ7
@Mytee Also needed to note a lot of why i'm going back to LORA is that I've been merging them into. checkpoints - There is a time i'll be YES retraining this on Lyco/LoHa - but current merging tools don't have the skill to do this.
Hey if you're wondering why you got banned from my discord, it'sb ecause you posted in the wrong channel and demanded the dataset yourself. I'd literally said to you i'd work on training it again soon, this waso nly 4 days ago. Patience is a virtue, I'm not going to attempt to prevent you from downloading my content - because content is free and easy for everyone to grab, but i've made sure you can't come barging in like that again demanding datasets. It's NOT that i don't mind sharing content but you did sor of barge in without looking HOW the discord was run who runs it and why they were "BOT" Profiles. Anything with the tag "Duskfallcrew" was us, and we had been actively chatting in there. I would prefer it highly if you don't demand datasets from us, and while we're happy to ATTEMPT to retrain this again - we won't be rescinding the discord ban.
Here it's been updated. Next time instead of telling someone how they WRONGLY trained a model ask how they screwed the data up - half hte model the first time WAS photos. Also MOST LORA can still produce photos on ANY model because it's a STABLE DIFFUSION setup :) Kthxbye.
@duskfallcrew i got banned for politely asking for info? wow! so special people. remember you invited me to join and help? I didnt ask for it. For such immature n pathetic group. Ha!
@duskfallcrew it is wrong then it is wrong. why it is wrong? ppl lower the figures for saving training time. it is obvious for Lora training, therefore, the production is haft baked. you want to post a lot loras. why? what the point for those handicap products? look around. look at the those who did good products and learn from the pro to prevent making rubbish and wasting people time.
@Mytee You STILL ARGUE THIS? The first LOFI model sucked why? BECAUSE IT HAD THE WRONG DATA IN IT. You want to be right when in reality it's a preference. The first lofi one was BUGGED because PHOTOS. Try again, it's preference - it's learning rate, not EVERYONE wants to sit for 4 days on google colab to freaking train a 900 epoch 2905DIM 4gb lora.
@Mytee Actually you demanded the training data, and demanded in the comments on here that you knew better than me. I don't care how Immature I am - The fact is, I'm telling you how I trained it, and i'm telling you HOW I got my information, have you SEEN how much stuff i've made? I'm not perfect, nor are you. Rookie stats, rookie arguments. Go have fun testing the new one.
@duskfallcrew I remember that I did ask nicely as always, "is there anyone know the new version, or ... and Thanks". with standard education and English at primary school level, ppl wd know it is not a "demand". wont waste my time for this. pls go hassle someone else.
@Mytee You already wasted your time coming in here telling me how horrible it was trained to start with. You do you, I hate HOW you've responded in all of this because that's what lead me to get cranky - If you wouldn't have come in guns blazing even tho i was trying to be nice and figure shit out - you'd come in to my discord ASKING FOR THE TRAINING DATA, that's not "NICE" tahts' "i'm gonna do it for you" - So yea, go stop wasting your time on me if you can'ts tand my shit stay out of my kitchen XD
Shows photo cause datasets contains a lot of photos or just certain photos which more bias to get represented in AI while training it. Dataset bad.
@LDWorksDavid Yup! Dataset bad, the first one had photos in it. A lot of it depends on the prompting and dataset, and the sheer fact it was about a 40/60 ratio with photos meant it was just not working right. The new copic marker edition works MUCH better now.