For business inquires, commercial licensing, custom models, and consultation contact me under [email protected].
My other LoRAs:
Aether Bubbles & Foam, AETHER FIRE, and more right here. All my stuff is sponsored by RunDiffusion.
This is Aether Cloud - a cloud texture based subject and object oriented LoRA trained on animals, people and some other stuff. Play with weights but 1 is a good base. Higher and you seem to get more clouds - lower and you get more distinct features. Works well with no negative prompting for basic things. Make sure to check out prompt examples with the images in this gallery.
ImaginAir!
Description
FAQ
Comments (38)
Is it a coincidence that I was testing this type of atlas yesterday and you made it today?
Atlas?
@joachim_s en, haven't followed it, just happened to see you on the front page today, recently I've been working on a generalized dataset for the cloud.
@Shivae Oh, we could’ve collaborated on it. That’s a shame! Maybe for a next version!
@joachim_s It's a good idea.
Nice! This was on my todo for SD1.5
Any plans on trying out a model for it?
No plans. If there really is a demand for it I’ll consider it. Can’t vouch for the quality though.
@joachim_s Cheers! I guess if other people want it, please thumb this post :D
Otherwise I don't mind training it, would you be okay sharing your dataset?
@joachim_s sd1.5 please
@nishu I will look into it. Meanwhile you can test it for free here:
We both had problem uploading the LoRA to tensor.art earlier.
But the creator of https://civitai.com/models/127139/classipeintxl-oil-paint-style-lora was able to upload his model to tensor.art, so I ask him how he did it. This is what he wrote back:
I didn't yet have an account there - it didn't look initially like that was a requirement, but it is. And there's another step you need to follow to prove your identity, where they give you a link to your TensorArt profile that needs to be included in the CivitAI profile as a social media link.
It's also worth noting that, while they do a great job of automatically importing the model (even past model versions) and images, the descriptive text will carry over HTML tags, so you'll want to rewrite them. And you can't use inline links, so any URLs in descriptions need to have links inserted as full web addresses.
Ok. How much is that site used?
@joachim_s Seems to be quite popular, I'd say it is one of the top 5 images generation sites. Many of the users seems to be from Asian (mostly from Japan and China).
What makes tensor.art special is that unlike most other sites, the free version allows you to use any model and any LoRA, so it is a fast way to try out a model (if somebody has uploaded to it, of course)
@joachim_s BTW, if you do decide to create an account and upload your LoRAs to it, I think you can contact the site and reclaim ownership of your Ghost LoRA that someone has uploaded to the site.
@NobodyButMeow Thanks for the help! I’ll check it out. Do they have a discord?
@joachim_s Yes, they do have a discord server. Here is the invite: https://discord.gg/2sXtXT2V
@joachim_s Thank you for uploading your Aether Ghost LoRA to tensor. Can you upload both this one and the Aether fire to tensor too?
BTW, you were asking how big tensor is. It claims to be the world's 3rd largest image and model hosting site, only behind huggingface and civitai: https://www.reddit.com/r/StableDiffusion/comments/16ukoin/chinese_site_claims_to_be_worlds_3rd_largest_sd/
Thank you! I’ll check it out!
You can use most of joachim_s's aether-X LoRAs on tensor art now. TBH, I checked out seaart, and it appears that tensor art gives you way more models and LoRAs to play with.
All good but whenever I try to run it with ControlNet - an error appears - RuntimeError: mat1 and mat2 shapes cannot be multiplied (154x2048 and 768x320). I've got no clue how to sort it out :/
Me neither. Sorry.
This error is generally when you try to mix 1.5 with SDXL and vice versa. Make sure you have the same ControlNet models as your SD model and Lora model.
Do you need ControlNet reference images or will pure text-based images suffice?
Hi, I have a problem, I don't know if I've programmed everything correctly but it generates images for me at any time and sometimes it shows me clouds that have nothing to do with other people's results. Looks here : https://ibb.co/vd5NKXK
You need to change the resolution from 512x512 px to 1024x1024 px. Right now your width and height is 512 px. All SDXL models and LoRA have a base resolution of 1024x1024.
@joachim_s thank you !
@joachim_s hi, sorry to bother you again but I still have a problem https://ibb.co/C8hj0ns
@Eden9492 That's weird! Since it looks good just up until 95%. I really have no idea why this happens though. Sorry. You should check the code in the background if there is an error. You're doing this at home on you own computer and gpu?
@Eden9492 Try the LoRA using a free account on tensor: https://tensor.art/models/653186997537635847
If it works there, then there is something wrong with your local SDXL setup, and you'll have to fix the problem yourself.
A small tip/hint for using this LoRA:
The best sampler to be used with this LoRA appears to be "euler". I usually use "DPM++ SDE Karra", but the "softer" image generate by "euler" is more suitable for this "fluffy" LoRA 😂
@NowhereManGo really think so? my experience isn't that euler does for as good photorealism.
@joachim_s I got that impression from using different samplers on the Panda cloud set. With “DPM++ 2M SDE Karra” the edge got “too sharp” for the LoRA to “cloudify”.
It's true that in general Euler is not the best sampler (“too soft”) but I think for aether-cloud that may in fact be an advantage.
Further testing is probably required 😁
@joachim_s I just added two more images to the "Panda Cloud": https://civitai.com/posts/740063 so that other can make their own judgment. Both images are generated using the same seed, CFG:8, Step:30, LoRA:1.0. To my eyes, the Euler version is “fluffier”.
BEAUTIFUL LORA!!
Thanks! Make sure to check out my other stuff if you like this 🙂
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.