A Lunar Dream- New 2.0 Release: Artsy Dreams April 23rd

[grave stratus]

This is a new approach to models. This is a finetune on Nai with a focus on artist styles and the ability to do what I'm calling "Style Blending".

A Lunar Dream can recall over 160 art styles with a high degree of accuracy. Why fiddle with various style lora when you can just ask for the artist can get them! Even better, you can mix styles in ways lora can't. You can legit use as many artists in your prompt as you wish and it will yield new results without any of the problems lora might have with burn-in. You can also weigh them! I want to repeat, that this is not a merge, or lora merge, this is a fully trained new model! I recommend downloading the realism version.

I am calling the previous version a beta, and this is now the V1 release. I will be considering making a 300 style version next, but I will need your help!

If you have a dataset you can donate, I would highly appreciate it. Heck, if you're able to rip an artist's pics I'll add them to the model. You don't even have to do the tagging. I want to do a SD 2.1 version, but the cost will likely be too great without like donations to pay for cloud gpu's.

https://huggingface.co/lunachan/ALunarDream
Posted it to Civitai, go review it if you can! And post some pics! https://civitai.com/models/31114/a-lunar-dream-160-styl
https://imgur.com/ojN6SUL

[dark ocean]

[brazen sphinx]

vae?

[grave stratus]

I used Blessed Vae personally, but they all work decently well, I just like blessed the most I think

[brazen sphinx]

advice on prompting?

[grave stratus]

oh, do NOT use anime, or use prompts that would alter style like that. I found anime resulted in it failing to work with styles, because anime isn't a normally used danbooru tag. So your basically telling it to ignore the training

[brazen sphinx]

got it

[grave stratus]

Cuz Dako style would be uh... very not anime

[brazen sphinx]

settings

[grave stratus]

I just do 25 step, cfg 7,

[brazen sphinx]

upscaler?

[grave stratus]

I personally use nickleback 7200, 0.45

Or just animesharp 4x for pure anime pics

[brazen sphinx]

im guessing my prompting is ass at this point lol

[grave stratus]

deleted all the unneccessary words, like high quality, intricate etc

[brazen sphinx]

I must have been overprompting

[grave stratus]

I think the issue is I delete all the bonus junk from my tags

it can take many prompts, but stuff that doesn't actually have a real meaning just force it to use the base model instead of the trained data

[brazen sphinx]

I see

[grave stratus]

cuz I'm not going to train the term- high quality into my tags

[grave stratus]

Hey you, if you setup a randomizer prompt for all the artists, you can then easily use it like 5-40 times and get crazy cool results. I addd a comma to the end of mine so it can also prompt nothing, and then just use it like a million times in my prompt hahaha. Also you can combine that with latent couple, and get different art styles per character

[grave stratus]

example of different styles blending with 2shot

[grave stratus]

So I've found that you can essentially lock in a style by repeating it. So I randomed a whole lot and got something I like, and I can now repeat it as its own unique style. So You can kinda turn this model into a whole lot of styles by combining artists multiple times. This is using like 30 lol

[grave stratus]

I just finishd the dataset for V2. It has 154 artists. A small dataset of my realism girls, a set of chibi's, and then a small set of backgrounds. It should in theory be pretty much magical in terms of what it can do.

[round river]

@wide juniper

[olive flare]

Have you ever thought about doing this on top of a SD2 model? It seems like the expanded CLIP would help

[grave stratus]

After completing the V2 on 1.5 i'll move on to making it for 2.1. I just don't personally like 2.1 atm. I don't think I have the horsepower to realistically train 768 res V2 and not have it take a million years.V2 is 3x larger so its rough. This is a huge training, and I imagine I'd be stuck doing batch 1, instead of the 4 I'm doing for this. And this V2 is going to take me a long time to finish training. I'm a combo of too cheap and too unsure how to rent a gpu to do it that way.

[grave stratus]

I thought maybe some people would try merging, but I guess I have to be the one to try it out. I genuinely thought the potential this provides would be huge, but no one seems to believe in the concept... Anyways, I tried some basic merging, to see what you can get by mixing it with aom2. And YES, you CAN pull the ability to get the art styles into your model, at a cost of some of the main models style. With some more experimenting it might be possible to retain more of the base models style, but I wasn't able to get it. Nearly all of the style is locked into 1/2 points.

[olive flare]

Interesting

[wide juniper]

Beautiful model, indeed

[keen aurora]

noticed since you're gathering artists, i've dumped the datasets i've collected in #1089711334444236800

[grave stratus]

Oh wow, thats super cool of you! I'll see if I can add them. I'm just about done V2 atm, but will be looking into if its possible to just add more styles to it using the same method of training once its done. Will def add any I've not already used.

[dull rock]

Very good. If you don't mind, I am curious to know how many images you use and what model you train on. This kind of model is actually very good for character lora. The problem is just that there are too many styles and without a full gallery showing I have no idea what these styles look like.

[grave stratus]

This model was trained on about 10k unique images. Some styles had duplicates due to being smaller in scope. Was about 12k total. It was trained on aom2, but due to the nature of this training method, it doesnt maintain elements of the base model. This was probably a mistake, which I've fixed in the V2 im working on. The best way to test the styles is to just use a randomizer, and make your own ideal style when something pops up that looks appealing to you.

[dull rock]

You probably want to release a wild card along with the model lol

[grave stratus]

I will with V2 for sure

For now, this is my current one

just use <[rando4]> to proc a random style

it has a chance to proc nothing, in its current stup

namely because I typically use anywhere from 10-100 copies of that rando, to get a really random look

[dull rock]

Btw it is natural that the default style of the base model is not retained after heavy training.

[grave stratus]

In V1 the default style is very close to base SD

Namely because I didn't train any images without a style tag

[dull rock]

Interesting. I guess the anime style is washed out

And I suppose you did not turn on conditional dropout

[grave stratus]

V2 I've trained a small subset of images with the anime tag, so that it should in theory be possible to get a- good style from default gens

[dull rock]

For big dataset it is recommended to have that between 5% and 15%

[grave stratus]

cond dropout was on

[dull rock]

Mmmm. Then normally this helps get a default anime style

[grave stratus]

at 0.1

[olive flare]

Merging with something with a default style is ideal then?

[grave stratus]

V1 is very clearly undertrained compared to most models, so its likley just not enough training to get that effect

As I showed in my previous post, merging with Aom with mbw can produce an enhanced base image

Although overall it just dilutes the styles

[olive flare]

Oh hey is that alea31415?

[grave stratus]

I don't know who that is, so prob not

Atm, my only question is, how to improve the anatomy. I don't know if thats just an undertrained issue or if there is something I'm doing wrong. Cuz V2 seems to be getting close to finished, but it just doesn't produce as clean a character

v2 wip update

accidentally didnt use the same seed, but the concept stands

[dull rock]

well yes I have multiple pseudo names on internet

[dull rock]

It seems that people always get improved anatomy by performing block merge

[grave stratus]

So something I'm noticing is the blooming problem, where some styles are tooo white skinned, dunno whats up with that

[olive flare]

Are you training on black and white images?

[dull rock]

can't that be due to noise offset turned on?

[grave stratus]

hmm that might be the issue, some black and whites ruining it... darn

hmm, but doesn't explain chonchon's, cuz its got none in it

[dull rock]

can you check the tag to see if there are some tags that are consistently bind with that style?

btw are you going to put the v2 model online very soon?

[grave stratus]

Yeah, I just want to train it a bit longer

Its producing some wacky results sometimes

But I'm gonna chock that up to its just really undertraind

I want to overtrain this one this time, so It will be more consistent at producing the styles hopefully

[dull rock]

If it is overtrained we can always mitigate it by merging anyway. What people always do with lora: this lora works best with weight 0.3~0.7 lol

[grave stratus]

V1 is like not noob friendly, it wont even produce a person unless you add like 3-4 people tags, if you have a few location tags

actually, I just realized I can do V1 vs 2 comparisons to check how good its doing

[thick island]

@pastel pivot

[pastel pivot]

@grave stratus Salt & I had an idea to add some of this style selection to base waifu diffusion. Could you upload your dataset or at least a list of links so I can work on that? I'm looking for a bunch of artists with varied styles

[olive flare]

Damn

combining this with AnyLora leads to amazing style responsiveness

[olive flare]

Not really sure how much of it is anylora and how much of it is the artist support but it seems to perform a bit better at base?

[dull rock]

Lunar is trained directly on nai. I think you would get more the lunar trained artist style if you train on nai. Training on anylora or any other derivative will result in a negative vector whose effect is hard to predict, but you may like what it gives in the end anyway.

[olive flare]

Nah the lora is trained on Nai

I'm just using anylora as a comparison because that's the merge material

I actually don't that artist was in the training set

Mixing it with AnyLoRA makes it respond better to LoRAs while still retaining some of that effect from the artist training

Without the merge it tends to fry a bit at the edges and while it gets more of the style across when all addnets are set at 0 it's not as strong as with the lora assistance. If you manage to get lora assistance merging it with anylora gets the strongest style effect of all the results

[grave stratus]

I'm nearly done v2, will try to upload it sith a much better documentation and support. I'm glad some people are testing what is possible. It would really be amazing to have the level of support in 2.1. I don't think I have the computing power to train 2.1 realistically. I already noticed more than double in my electricity costs this month.

[grave stratus]

For refrence, I'm currently at 427k steps in this training. I probably really need a cloud solution at this point...

[merry vapor]

cloud gpus are really expensive; its either always training gpu or nothing

[grave stratus]

eh, this has cost me probably close to 50$ this month, I don't see paying for it done being more expensive then that... considering it'd be running much faster at a higher batch

[merry vapor]

depends on what gpu you get; although its really costly

runpod has these pricings for gpu usage

[grave stratus]

According to my math, an A6000, assuming it trains at the same it/s as my 3080ti, would take 19 hours to get what I did in 75 hours give or take. Asuming linear possible increase in batch size from 12gb vram to 48 gb vram

Which would cost 15$ to train

1girl test, V3 is looking substantially more trained then V1, I restarted V2, so skipped V2

[muted steppe]

How many anime character names does this model has knowledge of ?

[merry vapor]

its artists

[muted steppe]

I know it's focused on being a combination of multi artstyles, but my question still stands

For example, AOM2 has more character knowledges than Any V3 even though I don't see that point being mentioned in its model card

[merry vapor]

its more of a side effect

[grave stratus]

The previous response is semi correct. My dataset should have a pretty decent variety of characters in it, most should have been tagged. Due to the duration of training, it should know them quite well. So its known characters will be different then other models, but it will know some. I personally could care less about making specific characters, so I havent tested which it knows. It uses base Nai, so if nai knows any, it will know them, and then it will know whichever existed in my dataset. So for example, it knows asuka.

[grave stratus]

I was going to release the new version early today, but last night made the decision to add my realism dataset to the completed model first. I'll release both versions, but from preliminary testing the realism photos aren't notably effecting the anime, namely because I made sure to train the anime first. In theory this model should be like legit Anything 4.0, I wanted to name it "Everything" lol, but will prob keep my luna branding name. As soon as I'm done this I'll be moving on to 2.1 if I can actually manage it with my gpu.

[pastel pivot]

Do you have an open source style dataset?

[grave stratus]

Uploading my dataset would make me want to explode, its 30gb

[pastel pivot]

if you have a list of links that's good enough. I'm working on an AI to classify images into different unique styles, and I think I can improve waifu diffusion with it

[grave stratus]

The majority of my datasets were scrapped, so couldn't really link. The rest are just what I could find on huggingface via searching

[olive flare]

I would be interested in the dataset tbh. Not necessarily going to use it but it seems interesting

[round river]

[grave stratus]

Hmmm, discords post editing sucks... I might just delete this and remake so I can format better

Anyways, official release is out, check it out! I'm renaming the first release as a beta, and this is V1.
https://huggingface.co/lunachan/ALunarDream

[round river]

Thank you!

[keen aurora]

Bless you luna, you did a good job. I can recognize some of my contributions just from the example

[round river]

Thanks- oh

[grave stratus]

Will post the updated realism version, which I'm finding is just amazing, can do this, and still do all the anime stuff, I've been using it all day. Its also good at semi realism

[grave stratus]

Posted it to Civitai, go review it if you can! And post some pics!
https://civitai.com/models/31114/a-lunar-dream-160-styl

[round river]

[dark ocean]

@grave stratus if you react the message with 📌 the message is pinned

so we can avoid scroll up to see the original message ;-;

[grave stratus]

ah

[uncut eagle]

i’m no good at prompting but i will say this model is quickly becoming one of my favs LOL

[grave stratus]

reminder, you can weigh the artists to weaken their effects

[uncut eagle]

noted!! i was mostly just plugging in the artists i recognized together when messing around

[grave stratus]

something to note, for reasons unknown to me, the term "profile" is used litterally, and puts the character facing the side

its not used in the standard sense that most models use it, as just a like upper body pic, so upper body is the tag you want

[olive flare]

My guess is that because of the way tagging association works profile got linked to side profile

[cold pulsar]

Thank you so much for sharing the model, it has tons of art style that I like.

[grave stratus]

Glad your liking it!

Okay yall, I just realized something, the version on civitai is better then the one on HF. I never updated the HF one with the realismV1.3 model.

[grave stratus]

200k steps into the questionable 500 character v2 model. Can't test if it's working cuz I'm not home, but in theory it should be interesting. Will probably release it as a beta, hopefully someone is willing to test it, and then run a second training with the styles reintroduced for a day to kinda reaffirm the style training.

If your interested, follow my HF and when I post it you'll get an email or something.

[soft drift]

Oh it is posted here.
Any legit pics with the realistic model?
I can't get proper results, but original ALD is fine

[grave stratus]

Official update: A Lunar Dream 1.5- Now with about 500 characters built into it! You may need to prompt the characters tags to get a decent result with them, but it knows most characters far better then your standard model. It in theory is as good as the previous release, but I'd not throw the old one away yet.

Check the characters.txt for the characters in the model, and test it out! It functions as a wildcard too! if anyone is able to rip the character tags off danbooru easily, please message me. I'd love to get a randomizer with all 500 that includes their tags, so its even better.

Note- I've had good results using a higher then 1.0 multiplier on the character, improves the odds it forces harder

Also, the only really way I could make this model perfect is if people started sending me their datasets... I need YOUR LORA's DATASETS!

I'd like to make a proper 2.0 with more features, but I dunno if Its worth it. If anyone has any suggestions on other things I could bring to the table that would make an anime model feel more- complete, let me know. I may do- real styles next.

https://huggingface.co/lunachan/ALunarDream/blob/main/ALunarDreamV1.5.safetensors

[grave stratus]

A Lunar Dream- New 1.5 Release (500 characters!) April 11th

[grave stratus]

Not perfect, but will have a very high chance of getting you your character, it has character tags. Its flawed in that it has too many tags, but dunno what to do to fix that.

[round river]

Oh damn thanks 😄

[soft drift]

Are the characters tested already? Sounds awesome.

[grave stratus]

kinda, if you use the characters.txt it will work alright, if you use the V2 of it, it'll work better, but may give undesired results, as its a tiny bit scuffed

[maiden geyser]

hello

[subtle patrol]

Just discovered that this model has Quasarcake on the styles.txt

[soft drift]

Tried 1.5
Oh it contains realistic images

[grave stratus]

The realistic images don't matter if your prompt any artist, it will overwrite any base style.

Testing the probably final version 2.0

[grave stratus]

V2.0- adding watercolor, oils, various traditional painting styles, and the major goal of making the default gens genuinly good. These are the defaults, but the current default gen with no artist is pretty good, but will be improving it.

[maiden geyser]

did you add the dataset i gave you?

[olive flare]

This is absurdly flavorful

[grave stratus]

WIP V2- base gen, no artists!

[simple valve]

SHEEEEEEEEEEEEE

[amber prairie]

I'm relatively new to how finetuning works. How does this approach differ from the usual method(s?) of finetuning?

[grave stratus]

I'm fairly certain all finetuning is about the same. The only major difference is my approach to actually doing it. I've done a lot of work to manually decide on datasets to use, and classifying everything with tags to recall them.

[amber prairie]

I see, so your dataset and method of approach makes this model more... "powerful" in the sense that it can better recall the specifics that it trained on?

[grave stratus]

Yeah, there arent many finetunes, and most are under 5k images. Usuually a single concept- basically, make the art like these images

This is the opposite, its about the user designating art to look a certain way, instead of SD doing it for you

[amber prairie]

Would you say that your method requires more explicit effort from users to get good generation?

[grave stratus]

So for example, I've recently been training real art styles, and I can proc cubism at will.

[amber prairie]

As in, users more or less need to prompt for a trained style/character to get decent results, rather than a generic 1girl.

[grave stratus]

My goal atm is to make the default- no artist gen to look good

The current version 1.5, has an alright default gen, but its not gonna be all snazzy like most models

but I added a .txt file you can use to randomize the styles, which helps a ton

And if you use a style, you can just get that style, for example, khyle

It takes more effort to get good results when you start vs other models, but if you find a combo of artists that you like, you can pretty much make it match any other model in terms of consistency

[amber prairie]

What's your method of approach for trying to get a default no-style prompt looking good? I'm not too familiar with how SD prefers a style, but I'd assume that it would prefer a style that it's mostly trained on (e.g. you have majority trained on mignon). So are you planning to overfit(?) a particular style?

[grave stratus]

But like the big advantage, is no other model can suddently make this image without anything but a word change

I''ve made a folder of no artist tags, and am training that for a bit

I also duplicated it, and added anime, anime style to it

[amber prairie]

Oh, I can see how that would work, an "anime" style tag

so that's why you have that in your example prompts

[grave stratus]

its not curently working, but thats prob cuz it needs waaay more time

im fairly certain im in a learning trough, where everything is bad

it kinda ebbs and flows in training, where you need to get to a certain amount of time before it rights itselfg

[amber prairie]

👍

I'll check it out when I can, I guess

[subtle patrol]

What would happen if I mix silicon with this

[grave stratus]

Which one was silicon again?

[subtle patrol]

Only problem is that I have 14 mbps with 2 mbps average

[subtle patrol]

Xynon model

The AutoMBW model

[grave stratus]

ah it was that one

if the goal is to get the base style of a model + the tags from this, its possible, just you have to be willing to give up a percentage of its accuracy

[subtle patrol]

How good is Lunar at backgrounds and anatomy?

[grave stratus]

You can MBW and get a good chunk of the main models style, while retaining it, but your prob looking at about a 20% style quality loss

[grave stratus]

couldn't say ,I've not used another model in so long

[olive flare]

Been playing around with merging the realism, the style loss is pretty huge if you try merging it

[grave stratus]

You need to use MBW, and then you can maintain a bit more, but at the end of the day style is kept in the mostly the same like 2-3 spots, so you kinda have to give up some

[olive flare]

memory tokens not enough

[maiden geyser]

is lunar dream 2.0 public?

[grave stratus]

A Lunar Dream 2.0 Release. Prob the final release, unless I get a bout of inspiration to try again. Too expensive to really go further then this.

Features:
500 characters,
190 styles
Now with Support for these tags! "masterpiece, absurdres, best quality, extremely detailed, detailed background" They will majorly alter your generations! Try variations of them. Or S/R variations and see which you like best. No longer will you have trouble getting non default anime gens! These will force a more limited style upon the model, but the default style is very painterly. This model is still going to be random to some degree, you have to take that for the benefit of having all of the styles and characters in here. The characters are probably the issue, but I think its worth having them.

Bonus tags that effect the gens in: Expressionism, impressionism, lineart, anime, anime style, 2d, (heartpupils, pupilhearts, pupil hearts, heart pupils), (magiceyes, magicaleyes)

Sorry- New stuff added to the styleslunaV2.txt. You will have to edit it if you want to use is as a wildcard, because I added a line that says- New Stuff. And as always, you prob want to delete belledelphine, chibi from it.

https://huggingface.co/lunachan/ALunarDream/blob/main/ALunarDreamV2.0.safetensors

[maiden geyser]

yay you added that one style i told you to add

[hushed tendon]

quasarcake nagaU
ALunarDream vs ALunarDreamV1.5 vs ALunarDreamV2.0

[grave stratus]

prob easier in a grid lol

[maiden geyser]

kinu barbari

:d

:D

[hushed tendon]

ok hold on

[grave stratus]

A Lunar Dream- New 2.0 Release: Artsy Dreams April 23rd

[hushed tendon]

quasarcake nagaU
ALunarDream vs ALunarDreamV1.5 vs ALunarDreamV2.0

[maiden geyser]

what

this isnt kinu bari

thats mitei398

[hushed tendon]

yukoring quasarcake matcha
ALunarDream vs ALunarDreamV1.5 vs ALunarDreamV2.0

[grave stratus]

eh, if the two are backwards, I blame everyone but me

[maiden geyser]

honestly

i blame myself

for giving you the names wrong

[grave stratus]

Anyways, I don't expect 2.0 to be as good as 1.5 in terms of styles being accurate, but it should be just generally more pleasant for most people. I haven't really tested if the styles are comparable, better/worse

[maiden geyser]

ok so heres the correct ones

Mitei398 is this

aoisora is this

kinubari is this

[hushed tendon]

yukoring quasarcake torino aqua mika pikazo
ALunarDream vs ALunarDreamV1.5 vs ALunarDreamV2.0

[maiden geyser]

sorry for the confusion

[grave stratus]

not a big deal, doubt many people will notice or care

[grave stratus]

I highly recommend using impressionism at the end of prompts, and weigh it down if its too strong. It adds a lot of nice effects.

[maiden geyser]

its better

the styles fit perfectly

i put all the artstyles in the negatives

[maiden geyser]

i love how you can combine the artstyles together

for example i combined mitei398 with Kinu bari

i can experiment with some stuff

Kinu Bari:0.4, mitei398:0.6, ⬆️

Kinu Bari:0.6, mitei398:0.4, ⬆️

Kinu Bari:1, mitei398:1,⬆️

such an amazing model

[grave stratus]

mitai is like a free add lineless, its super fun. I like using pantyandstocking to get more wacky proportions on the characters.

[maiden geyser]

that is true but its not mitai its mitei and the original artist for the thing is Kinu Bari

but still

sorry for giving you the names wrong

[tacit sluice]

downloading v2

[subtle patrol]

I seriously need an image with the tokens instead of a text file and multiple images 💀

[grave stratus]

The HF has 160 of the 190 artists in image form...

[subtle patrol]

...

I was trying to run Auto to use the model

I fucking corrupted my partition somehow

[soft drift]

2.0 is a lot better then 1.0, 1.0R and 1.5
At least astolfo is now reconizeable.
The prompts will be adjusted from scartch.

[grave stratus]

If anyone finds any super good style combo's let me know! I haven't experimented much with it in this release.

[subtle patrol]

I want to try but my drive is uhm bugged

[tacit sluice]

I'm just starting to play with it mixing P&S with dakusuta and Khyle

bruce timm + P&S goes full cartoon

[maiden geyser]

mixed kinu bari with kinu bari made kinu bari but more kinu bari

or just one single character

sometimes

Kinu Bari:0.3, mitei398:0.2, aoisora:0.5 is chaotic

[tacit sluice]

adding makoto shinkai to prompt adds good illumination

[tacit sluice]

used that combo with my prompt

[maiden geyser]

will there ever be a negative style for the model

actually a negative autonomy should exist

for better autonomy

in the negatives

decided to revisit a old embedding i made

????

tried making girls in different styles
2girls
AND
2girls, kinu bari, green hair, blue eyes
AND
2girls, mitei398, blue hair, green eyes

[tacit sluice]

I feel this model is great for 2D styles👌

mixed models normally have a 2.5D feel but with this I can reach more easily the 2D styles I want

this is the model as base + a locon based on mike mignola

this is the same prompt + locon and seed but with AOM2

[subtle patrol]

Flat enjoyer 🤝

[tacit sluice]

oyari ashito+Qooo003+makoto shinkai

[tacit sluice]

so, I mixed range murata + a locon I made of genndy tartakovsky

[grave stratus]

That one's mightly cool!

I've noticed, after trying my own locon on my model, that my model takes lora waaaay stronger then mixes. I compared it vs aom, and it was like doing nothing on aom, and lots on mine. So will be interesting to see how that changes things.

[hushed tendon]

mokunoki matcha Hiten fajyobore

[grave stratus]

this looks soo familar.... but don't know why

[hushed tendon]

Definitely has a similar feel to... something.

[amber prairie]

almost reminds me of pikamee?

at least the one on the top right

[hushed tendon]

Tried merging hll into ALunarDream.
ALunarDreamV2.0 + (AOM3A1B_hllS2 - AOM3A1B_orangemixs) * 1

Sometimes it works, and sometimes it messes up the artist style.
enjoy wide grin Bae lmao

[hushed tendon]

Tried making a AnythingV5 version of ALunarDreamV2.0
anythingV5PrtRE + (ALunarDreamV2.0 - animefull-final-pruned-fp16) * 1

AnythingV5 version vs original version

[hushed tendon]

Tried the Mika Pikazo artist since that artist had issues in this attempt. Looks fine.

So then I add hll to my AnythingV5 version and it looks fine? 🤔

[hushed tendon]

Ahh, I see, adding hll messes up the styles.
AnythingV5 version vs AnythingV5 + hll version

[hushed tendon]

Please excuse my experimentation.

[hushed tendon]

Something interesting I found: Using the model toolkit extension to put ALunarDreamV2.0's CLIP on AnythingV5 adds the ability to use artist styles with varying degrees of success.

[tacit sluice]

wait what?

you can do that?

[subtle patrol]

yeah

[tacit sluice]

nice, gotta experiment more!!

[subtle patrol]

Copy AnythingV5 clip on SD1.5

[hushed tendon]

ALunarDream2.0 vs AnythingV5 + (ALunarDream2.0 - animefull-final-pruned) * 1 vs Previous with CLIP changed back to ALunarDream2.0

[grave stratus]

Cool experimentation! Glad someone's finding some fun out of it. Honestly, I'm probably completly out of the ai game at this point. I've finally moved back to making my own actual art again. While I'd love to make this for 2.1, cuz I think that would probably be my ideal setup, I just don't have the pc to train it, and with temps rising, training on my own PC just aint gonna happen anymore anyways. I'd have to math how many hours of a rented PC I'd need to do that training, and make a call if the cost is worth it. I'd guess like 20$, but dunno.

[grave stratus]

Oh, and go follow my twitter! lol

[subtle patrol]

I don't have a twitter

[hushed tendon]

Would be nice if you could convert a 1.x UNET to a 2.x UNET, but that doesn't seem to be possible right now.

[grave stratus]

I am officially delisting this model. If you have it, congrats! I ask that you please respect my wishes and not redistribute it on websites. If you wanna give it to a friend go ahead, I'm not your mom. Just like try to only give it to people that would never even consider uploading stuff online, so it doesn't get back online. This model was a lot of fun to learn how models work, and experiment with, but the main goal was to show what is possible, and I dunno, it seems people don't love this approach. I hoped someone with more resources, or money would try a more advanced take on this in the future, that was more comprehensive.

[subtle patrol]

Wait I still don't understand why you are removing it even after reading this

[supple flower]

What license was the model published under

[knotty flume]

It's the usual CreativeML OpenRAIL-M license.

This model is open access and available to all, with a CreativeML OpenRAIL-M license further specifying rights and usage. The CreativeML OpenRAIL License specifies: Please read the full license here

## Big Thanks to The various people that provided some datasts for this project, and also the people that helped me figure out how to get this done.```

[tacit sluice]

Thanks a lot, I enjoyed generating stuff with this model and using it in my experimental mixes!

If you ever want to try to train other model with more western stuff I have more datasets to share👍

[olive flare]

It's trained on a bunch of clearly specified artist styles so it's a prime target for attack

[subtle patrol]

huh

[knotty flume]

The attack part doesn't necessarily correspond to the "don't share widely" tho

[olive flare]

That's true. I'm just going off if it were me and there's no indication whether or not that's close to the mark

[hushed tendon]

Sad to see this go, but glad it existed.

[hushed tendon]

@grave stratus Don't forget to delete the CivitAI listing.

[grave stratus]

Its a bunch of reasons. I shared it primarily as a - here thing. I didn't post the updated version on civitai for a reason. It was mostly about sharing what I made with the people that helped me along the way. I am still an artist, and would like to get back to making my own art, and would rather distance myself from the project. As I was clear to state- if you want to give it to someone, go ahead. So like, if someone asks for it, feel free to toss them it. Will someone post it out of rightous indignation that I would have the gull the gatekeep something... possibly. But lets be real yall, I didnt even get 200 downloads on civitai, its not a huge loss, it wasn't making waves, and changing the world here.

[grave stratus]

Yeah, I just did lol. It didn't actually have the final release anyways, so not a huge deal.

[hushed tendon]

That one had V1?

[grave stratus]

I can already see someone taking it, and like.. making a "merge" version, thats 99% it, and just releasing it, with that being their justification for it being tots fine.

[grave stratus]

it had up till 1.5, so servicable, but a bit jank

I'll also probably be like... still here every once in a blue moon, I'm still working on iterating a lora using my artwork, just that's a huge personal project, so not posting images much. I'm curious if I'll get enough art to make a small model from it, maybe with a blend of other art to balance it out. So like, if someone just kindly asks, I could give you it if you actually wanted it.

[subtle patrol]

Don't hesitate to post your stuff in #art

[grave stratus]

I do have an idea for a safer, follow up, but heat will prob make me never make a new model till winter, so mleh. Namely, combine 2-3 artists that combo well into folders with 1 keyword. That way, it can't reproduce an artists style, BUT can make many styles.
I may try to make this, if I'm still into ai, and its like... sept... I cant handle the heat produced from my PC anymore, so I'm like 100% out of model making until then. And renting a pc is too much munny.

[supple flower]

Who's doxing you

[fervent veldt]

what about using collab to train the lora?

oh wait a model hmm

[subtle patrol]

There's this which can do multiples anime styles
https://civitai.com/models/16828

It's a clip skip 1 model tho

[supple flower]

Op is being attacked?!

[subtle patrol]

I don't think so

[supple flower]

Then why are they afraid

[knotty flume]

Nowhere is it stated that this is the case.
The reasoning is pretty clear

[subtle patrol]

They just want to be an artist in the end so I believe they don't want to be associated with the AI scene as much anymore

[supple flower]

Then publish under another name

[fervent veldt]

salutations everyone

[subtle patrol]

My man you know that anti AI peoples are twitter nolife?

Twitter nolife can find you easily

Anyway if you have the model don't complain

[supple flower]

So it is because of fear ot attack?

[subtle patrol]

Probably one of the reasons yes

[subtle patrol]

Also prob this too

[supple flower]

Under another name would solve this

Its fine if they don't want to, the reasons they say just sound like bs

[subtle patrol]

I mean as I said twitter nolife are still dangerous

They trace back stuff to you even if tried to delete evidence

[supple flower]

You overestimate capabilities

Trust me I've been on the paranoia train

[subtle patrol]

You've never seen someone get tracked down by Twitter users lmao

They can ruin your online life for the dumbest reasons

[supple flower]

How are they going to track down no name no email huggingface account only accessed through vpn

Or an anonymous 4ch upload

[subtle patrol]

Anyway Luna chan made a choice so it's not that much of a problem

[supple flower]

At I said, Its fine if they don't want to, the reasons they say just sound like bs

[maiden geyser]

sad for it to disappear

if only i had the model

[subtle patrol]

As I said here

They just want to be an artist in the end so I believe they don't want to be associated with the AI scene as much anymore

[supple flower]

"Associated with"

[subtle patrol]

I can put it on a google drive and send it to u

[maiden geyser]

but it makes sense its their decision

[maiden geyser]

ok thanks

[supple flower]

I just explained why this doesn't make sense

[subtle patrol]

might take a while though my upload speed is awful right now

You want 1.5 or 2.0?

I'll zip the styles and characters .txt with it

[supple flower]

Can you upload separate styles characters, I dont have them

Or is that what you meant 😂

[maiden geyser]

2.0 ig

[subtle patrol]

Here's all the .txt files from the repo

[supple flower]

👍

[subtle patrol]

I renamed the .txt but inside is the same

[subtle patrol]

k

I'll send in dm when I finish uploading

[maiden geyser]

ok thanks

[subtle patrol]

Should take 30 minutes

[maiden geyser]

ok

[subtle patrol]

Btw only 20 to 30% of the characters are actually accurate

Still very good percentage but just to note

Also since the model is overfitted with all of these styles and characters well it's bad at backgrounds

[subtle patrol]

almost finished uploading

sent

[hushed tendon]

Doesn't seem to be as good. 🫤

[subtle patrol]

It doesn't have as much styles and characters

And it's a bit different from conventional anime models

Check his examples prompts

[fervent veldt]

hey guys

are there any other sites i can find loras besides pixiv

[supple flower]

Civitai

Wrong channel

[worldly knoll]

can i have it too please? im curious.

[subtle patrol]

Send me a dm

[grave stratus]

I'm uploading it to my onedrive right now, just my upload speed is slow, i'll toss a link here, and keep it up for a day or two

[subtle patrol]

ok

Have you tried 2.0 clip on 1.5?

[grave stratus]

Ironically, had I just deleted this post from the model sharing page silently, no one would have ever noticed nor cared >.>

[subtle patrol]

CyberWaifu tests were very promising

[subtle patrol]

I would have cared

[grave stratus]

I only use clip skip 2

[subtle patrol]

I kept coming back to this post over and over so if it was gone I would have done a dm

[subtle patrol]

I mean copy the clip #ai-art message

#ai-art message

[grave stratus]

Oh, dunno really how to do that

[subtle patrol]

Merge model can do that I think

[grave stratus]

Did you check if it loses the stuff trained for 2.0 in the process though?

That would essentially just make it 1.5 but slightly different

[subtle patrol]

Yeah it will since it lose the 2.0 clip

but it's very competent

kinda 1.6 lol

or 1.5.1

Or copy the clip to a very good model too

I'll try later

I already got 1.5 and 2.0 on GDrive lol

You didn't need to

[grave stratus]

Yeah, but this means people don't need to dM you

[subtle patrol]

I mean I don't mind

[grave stratus]

I havent used 1.5 in forever, but now that you mention it, gonna give it some tries and see how it fares with my lora, just to make sure I've not been missing out on a better result

[subtle patrol]

👌

[grave stratus]

Hmm the results are very different, but overall not as desirable it seems. Its giving me big breasts, despite using small... so ehhh. I can't honestly remember what the dif between 1.5 and 2.0 was lol I might have retrained 2.0 from an older version/the start again, but with 0 tag dropout.

[subtle patrol]

I forgor how to copy clip it's been too long since I did that kind of merge

[grave stratus]

Its also hardpressed on twintails being like, this, despite that not being correct...

[subtle patrol]

Was it the config option in merger?

I'm so lost 💀

[grave stratus]

That or its from MBW

[subtle patrol]

@grave stratus

1.5 with 2.0 and 2.0 with 1.5

Better anatomy on 2.0 with 1.5 with more consistent styles

[grave stratus]

Hmm tried to do it with the toolkit, but was left with space used on my hdd, but no clue where it went lol

[subtle patrol]

[grave stratus]

Or Well, I guess I did it wrong, was clicking the export button

No clue where that was saving it..

[subtle patrol]

also #1096508811667325019 with Lunar clip

[lime rock]

sorry, mind if I ask whats the diff between 1.5 and 2.0?

[subtle patrol]

a few more styles and support for prompting characters with 70 to 30 % fidelity

[grave stratus]

2.0 was primarily focused on a better- no artist prompted experience

[lime rock]

Oh sweet! I'll give 2.0 a try then.

[grave stratus]

I don't quite understand the whole clip thing, outside of clip, what else is left? Like what is being kept from the model when you inject another clip?

[subtle patrol]

Clip is what let's you use the unet trough text

[lime rock]

Hmmm, seems more true to form to the anime with A Lunar Dream. Guess i'll have to work on my prompt game.

[grave stratus]

Why is her hair like... black?

was it trained on manga pics?

[lime rock]

It was specifically trained on this anime character (Lora model)

[grave stratus]

Okay, made both clip versions. But yeah, both 1.5 base versions produce wrong images with my character lora. giving her too big a chest. The 1.5 with 2.0 clip does get the hair right, but still double the chest size.

[lime rock]

Pastelmix seems to be better for this specific model, but I think A Lunar Dream is definitely pretty good, considering it was able to accurately recreate the character (others simply turn my model into the stuff of nightmares with disfigured face and 5 hands)

[subtle patrol]

AuroraOne is one of the best models for LoRA and LyCORIS if you want

[grave stratus]

I dont have any problems using 2.0. Just my 1.5 just is too unpredicable for me, hence why I scrapped it and made 2.0

[subtle patrol]

What model did you train your LoRA on?

[grave stratus]

my own

From my testing mine worked better then nai, and anylora, at least when used on my model

[subtle patrol]

Probably better to train on a model with good LoRA compatibility like NovelAI AnythingV3 AnyLoRA or AuroraONE

It'll be more portable if you do that

[grave stratus]

my model should be closer to nai, then any of those

[subtle patrol]

And 2.0 with 1.5 will probably work better too

[grave stratus]

my model is a finetune of nai

[subtle patrol]

Maybe you could see the merge block Aurora did to get a better compatibility with LoRA

[grave stratus]

I dunno, maybe you misunderstoof, I don't have any trouble with compatibility with lora

lora all work very easily and excessively strongly on 2.0

[subtle patrol]

true they work well

But the model is really bad at background

[grave stratus]

I've never really had trouble with bgs

In what way?

[subtle patrol]

Compared to other models

[grave stratus]

Yes, but like, in what way is it bad?

[subtle patrol]

Like streets it won't output this kind of quality

It's overfitted with too much characters and styles

[grave stratus]

Cuz like, I don't consider that quality, i consider that holdover of real photo;s

Which is the exact opposite of what I'd want in an anime/fantasy only model

[subtle patrol]

It's Aurora with ChainsawMan LoHa

of course it's good at at realistic streets

As 60% of the dataset was backgrounds

[grave stratus]

Yeah, just I didn't want backgrounds like that at all in my model

[subtle patrol]

I see

[grave stratus]

this models whole point was to get rid of the- zero artistic direction aspect of most models

toss me that promp

So like, this is what I get

[subtle patrol]

¯_(ツ)_/¯

Also funny thing 2.0 with 1.5 clip is the best at Mika Pikazo between both models

[grave stratus]

Its not great at resolving characters in a detailed bg, but it certainly can do bgs

[subtle patrol]

But the thing is that Lunar can't do coherent bgs

This door

The car is flat like a 2d drawing

See what I mean?

But for sure it can do interesting bgs nonetheless

[grave stratus]

for sure, but I think you use overfitted incorrectly. Model mixing has that problem, but when you get to model making, that doesnt really apply here. Its mostly just its lost the default nai/SD photorealism, which makes up like 90% of all models backgrounds

[subtle patrol]

I'll be off to diner :)

back in prob half an hour

[grave stratus]

But if you prompt less stuff, like in this one, it makes a coherent bg- it just defaults to a very mediocre art style without one used

snatii gets photoy bgs, combined with makoto shinkai

NFT shirt lol

But overall, it just does worse at wide pics when it comes to resolving, dunno why that is

But yeah, without artists its not gonna do well at bgs, or really even default gens. Mostly about finding the right artists for the mood you want.

[subtle patrol]

But there’s so many

You literally trained 162 styles on 1.5 and even more on 2.0

[grave stratus]

Thats why you just run rando's lol

[subtle patrol]

…

[grave stratus]

this is pasoputi, fjsmu,, and (snatti:0.6), (makoto shinkai:0.6), kgt in the back for example

[subtle patrol]

[grave stratus]

its worth remembering that you can not only combine the artists but also reduce or increase their weighting

[obtuse apex]

seeing that 404 when I finally shat down to download it made me notice. baka

I care

I have seen so many stuff just disappear just like that.

[tacit sluice]

mixing artists is fun

at the moment I can't do more experiments but I want to try more bg art stuff

[valid rapids]

@grave stratus what happened to your repo

[subtle patrol]

💀

[hushed tendon]

Deleted.
Wish I saved the first ALunarDream.safetensors with the hash 6a08d3c736.
That seemed to be the best at mixing styles such as quasarcake and nagaU.

[soft drift]

Thank you for the model! Your model will be remembered because of the noveality!

VBP has even LESS download counts, given the rich technical details

#1094871327833325639

[hushed tendon]

ALunarComparison on quasarcake nagaU

ALunarDreamV1.0
ALunarDreamV2.0 with ALunarDreamV1.0's CLIP
ALunarDreamV1.0 with ALunarDreamV2.0's CLIP
ALunarDreamV2.0

So yes, ALunarDreamV1.0's and its CLIP layer are still very good.

[subtle patrol]

You got your hands on it yay

[maiden geyser]

can i have the 24dd6a2392 one

i like its color

[subtle patrol]

You need to put 1.0 clip on 2.0

[maiden geyser]

uh

how?

[subtle patrol]

model toolkit

lemme show you

or I could make a huggingface repo where I upload stuff

[maiden geyser]

do it

[subtle patrol]

Anyways I will show you how to transfer clip first

[maiden geyser]

i would like to remind you im a colab abuser

[subtle patrol]

https://github.com/arenasys/stable-diffusion-webui-model-toolkit

Select the model you want to transfer and then load
Choose component and CLIP-v1-SD
Then choose which model clip you want in the file section
Then import it and change name then save

Anyway I'll do it still downloading 1.0 right now

[olive flare]

What's the best for clip replacement? 1.0 or 2.0?

[subtle patrol]

They all have their strength it's more on personal preferences

I'll send you the models if you want if git add -A want to finish because maybe adding 9 files of 2 gb each might not be a good idea

git lfs please

[olive flare]

I'll just replace the clips myself tbh

[subtle patrol]

But you need the model

[olive flare]

I have the model

[subtle patrol]

oh

[subtle patrol]

@grave stratus How did you train multiple styles?

I wanna try my hand at it

[olive flare]

tag each artist

[subtle patrol]

But like LoRA for each artists with token and merged into the model?

[remote forge]

No, that won't work

[subtle patrol]

idk so that’s why I’m asking lol

[remote forge]

iirc Luna did a finetune so probably just had them divided in folders

tagged with each artist name

and just trained on that

[subtle patrol]

[remote forge]

you can do a big ass lora for all of them like that as well, but doing singular loras and merging them together is not a great idea

[subtle patrol]

[grave stratus]

I had 160+ different folders

[subtle patrol]

Holy shit

[grave stratus]

Not including the 500 seperate characters

[subtle patrol]

What is wrong with you

[grave stratus]

Hence why I stopped doing it, it was like, I need donations levels of work

[subtle patrol]

I should totally try doing one with 30 good styles so it doesn’t get too overtrained

[grave stratus]

If I had the money I'm sure I could make a really competently done version now, but oh well

There is no overtrained in finetuning, just poorly tagged, or setup

You can feasibly do hundreds of concepts without messing things up, so long as the tagging is good enough

[subtle patrol]

Being overtrained I mean losing a lot of knowledge of the base model

[grave stratus]

again, you only really lose it if things are tagged wrong, and then you mess up what long hair is or something

in finetuning, the entire finetune can hit a- overtrained point, but the amount of concepts you taught it has nothing to do with it, its just generally trained too long

[subtle patrol]

But like your model is not good at vehicles

[grave stratus]

My models are vary from undertrained in V1 to overtrained in V2

[subtle patrol]

What’s your gpu?

[grave stratus]

3080ti

[subtle patrol]

Oh how much is needed for a finetune?

[grave stratus]

12gb ram+

[subtle patrol]

I only have a 3070

I’m screwed

[knotty flume]

Could rent cloud resources. But expensive

[grave stratus]

any lower and its gonna take too long to be worth doing

[knotty flume]

48GB card could be had for like $0.6 an hour

[subtle patrol]

Not worth it

[grave stratus]

I'm fairly certain it woulda been cheaper to rent a gpu then my electricity costs

despite having cheap electricity

[subtle patrol]

30xx series is power hungry

That’s what the 40xx series fixed but also expensive

[grave stratus]

Hence why I'd really only ever do this again if I got donations to pay for the training. And well, ai isn't really about paying to do stuff. And I don't make stuff generally want- ie character lora, so have no real following like say Lykon has.

Hopefully waifu 2.0 is really good in the end, and like... a 2.1 training breakthrough happens to make it easier

[subtle patrol]

[knotty flume]

Sorta feel the future will always be individual LoRA over training something that can do it all unfortunately

[grave stratus]

Well I just mean, make lora training easier too

Cuz atm, lora training being gated by slower training for 2.1 is rough to increase adoption

[subtle patrol]

There’s Muse in development which needs 4gb of vram to train

[grave stratus]

Potentially in 2 years gpus will have caught back up, and will be generally able to do it better

We might just be in a gpu lag atm, where in a few years all gpus pack higher vram, or ai accelerated cores become usable for all

[remote forge]

this, loras for sd2.1 are such a pain

[grave stratus]

I havent tried the tensor accelerated SD genning yet, but hear its quite stronk. If that makes its way to lora training, we'd have a huge boost

[subtle patrol]

Not sure if that could work

You would need to convert the trained checkpoints to safetensors somehow

[grave stratus]

But anyways, imma go back to being an artist with functionally 4 followers again and finish my drawing lol.

[subtle patrol]

Okidoki

[autumn onyx]

I did this and it didn't work

[subtle patrol]

???

I updated that message to be more clear

[autumn onyx]

yeah I did that but it didn't change anything gonna try again

[autumn onyx]

ok it worked ty

[autumn onyx]

ok the results are not what I expected, everything became a chibi

[subtle patrol]

Yeah don’t do it on models that are too different

[autumn onyx]

shouldnt be different

typing loli gives toddlers on the model I transferred the clip to, without it it seems fine