u can only generate images using this model, then use a image-to-video model to turn the image into a video

Just that you know the result structure of your upscaled result is horror most likely you never saw a correct 7 mpixel digital photo ever to call this quality.
Technically i will only say 2 word "Wavelets/Fractals" ugly result

@alekhammes966 Sorry but that was quite silly thing to say. I do own an actual quality digital camera (used in professional level), but i compare SDXL checkpoint quality against OTHER SDXL checkpoints, not against technology made to dish out quality images from a system level from an actual real life, and not you know from a virtual reality created in seconds lmao. Check out and compare the original SDXL checkpoints, then many of the other creators even new ones maximum potential quality.. then come talk again. Bye. 🤣

Probably a classic skill issue. There are no "wavelets/fractals" for me, only good results in 1536 x 1536 and 2x and 4x upscaled ones. I do images locally with ComfyUI with settings i've used for a while already, did not even had to change anything for this checkpoint, so.. what kind of settings/workflow do you use?

*small edit ^^

@Eclipsia you should not use comfyui if you don't know how diffusion works

anyway i stand by my technical analysis of your result as a engineer you created something which is not worth the bitrate abuse for nothing except artificial data structure nor does it helps your visual cortex or perception in any shape or form it was just a abuse of unnecessary bandwidth.
Unfortunately you see this often from people who want to sell snake oil im used to this since decades of itc work.

and especially at some point all this model variations start to become frustrating and useless as well and have only one goal to keep someone a income nothing more just useless survival instinct that creates more useless data structure and valuable resource waste.

But that will fire back heavily at all of us.

And you wont understand anything of what i said if you do not understand what the VAE is and it's limitations inside the diffusion process.

And if i look at your result here you surely neither have any idea of visual data compression
and for Civitai and users this has direct consequences in the form of resource waste.
And that resource is mostly energy consumption for nothing that should be.

@alekhammes966 I think you might need to lay off the magical shrooms for less fractals and fewer messages to spam me with, because let's be honest.. who asked?

Also, since you are the master of SDXL, oh please post your images to showcase your ultimate set of skills and technical prowess. I am thrilled to see what you have. Message me when you are done.

PS. I stand by my statement, that this checkpoint is currently the best in quality potential for SDXL. My other top3 favorites currently are HuslyoRealismV2 and ePicRealismXL_PureFix, Huslyo's checkpoint is like i'd say best all-rounder tbh, but that again is user preference.

All i see is a random on the internet telling someone's post is bad and their image is bad and the checkpoint they used is bad and "waste of bitrate" and whatever.. now show me your own images (done with SDXL only, no img2img, just txt2img like me) and what is better checkpoint in quality than this? Again, i am giddy with anticipation while waiting for you to post here and show instead of being big mad in chat. 🤣

@Eclipsia nonsense its not about that it can support this resolution it's about if it makes sense compared to a different path useless to talk to you

@Eclipsia Find my result space on Google "Kosmosx Slushe" all Exclusive handmade prompts my own database of thousands by now.

@alekhammes966

I checked out your images.. actual generic oversaturated slop images in a "3D porn site", yeah there might be "thousands" of them, but sorry to say this: looks like 2023 old checkpoint quality when people used still "skitzo negatives" etc and were still new and learning to prompt, full bokeh slop with bad DoF focusing on random things because most your backgrounds are simplified and not even detailed, and people are full of mutations, errors and poor quality 750 x 750, uncanny & artificial looks etc etc.. and you talk about waste of bitrate? I think your images are mostly waste of space (and my time) having checked them.

And do remember, you came here telling what i said was wrong, that my image is bad and this checkpoint is bad etc, so in return you ought to put on your bigboy-pants and take honest feedback, it's not a shame to take occasional L's.. happens so don't worry. I think most people appreciate this checkpoint and it's qualities based on the feedback here, so maybe one day you can admit to yourself that it might be just a you-issue if could not find what makes this a great checkpoint.

You said it's "useless to talk to you", and i am happy about that so the discussion is now over, nice that we had this chat opportunity and it's now done, i wish you a nice day and rest of life. 🤣

@Eclipsia you just looked at the SD 1.5 results artificial looks is a combined sword if you don't get what you prompt for because you merged the logic to death

@Eclipsia some scenes are 10x more complex from the shoot complexity then anything you see here

@alekhammes966 This is free software. And it is undisputably good compared to other SDXL checkpoints. Now, if this is not up to your standards or preferences, do you know what? There are other models. There are digital and analogue cameras. In fact, there is a whole world out there. And you have ONE life as far as we know. Spend your time wisely on things that matter to you for real. I am pretty sure if you stop just 5 minutes and ask yourself "What do I want in this life? What can I do now to get closer to that?" your reply will not be "I want to argue with total strangers on the Internet about a piece of free software they like but I don't."

@coyotte no it just burns me to death seeing how the 500 thousand useless model iteration merged to death comes out without any substantial anything only the purpose of making money and by that wasting resources just to keep a useless machine running.
If you would have just some consistent idea or prove that their is some brain functioning in your model but no worry i will test that.

@coyotte believe it or not but you can merge a model to become dumber

@coyotte simple prompt adherence test full fail 1st generation fails target on seed 1
not even the couch 2 character sex interaction test (in natural language it survived) if it needs a very specific setup its almost garbage.

if it not even understands the relation between penetration couch and african inside a environment target and just shows environment with some sex

@coyotte Industry leaders :D industry leaders that beg every scientist to use their research

Yes i see what you doing the same as subtle shader does who wants extra money for this crap like every cloud service will be happy to see their users going to use these checkpoints so they can charge extra as well ;)

That work was introduced based on SDXL first by OnomaAI (Korean Research Entity) in their illustrious V 1.0 VAE training

Also SubtleShader is a funny guy he uses DMD2 to market his own work funny guy really funny guy that trains with copyrighted material of real humans

So 100% unique no merge huh who you want to fool here you share the same genetic tree with Pornmaster and Jugernaut you fear i could hurt your income this is why you want to get rid of me now.

@coyotte search a real job sooner or later you have to

@alekhammes966 somehow I just know you were that pathetic kid everyone at school used to make fun of 🤭

wow @alekhammes966 Actually still going off big mad.. (i get notifications since started this topic lol) alek you know how the mentioned checkpoints share "genetic tree"? Because those other 2 did dip some of the sweet Lustify in them from an earlier version like a magical spice and it made them better. You are getting it all backwards.

Also, now finally everyone knows why you came here all mad to post at me, you were actually angry because of the Early Access + 1 week cost (just wait 3 weeks, was it really that bad to get your blood pressure up and go this mad over it?) and some other creators wanting to make a few coins/buzz to cover some costs.. so you were big mad, saw me telling that some crypto millionaire should donate for this great work, and rightfully so, good work should be rewarded, since healthy capitalism motivates to innovate and work long hours on something, personally i am grateful and will support when can.. you saw that and got mad at me also.

If i had more time just now, i'd make some image like that meme where a cat puts it's paw on top of a poster's hand and the text says "it's time to stop.." because sometimes someone else has to tell when things not look healthy anymore and should just go do something else.

These type of creations are a work of passion, but there is also the effort and costs. Electricity and pc hardware are not free. Time is money also. I am a poor student, but you don't see me complaining, i am happy we did get this checkpoint for free after a wait. Subtle Shader did also put that Ultra1 version for free, since you mentioned their name. So you get free good stuff, should be happy and not mad.

@coyotte Hih, no worries, it's fine.. 😅 i just get lots of notifications that's all. but i have been happily just using your amazing checkpoint (and couple others) most of the day today again. ^^ I play also some games and in Discord it's sometimes full ping dramas etc, so this is just another normal monday for me. * sips tea happily *

@Eclipsia i would be not mad if they would stop this oh i found a problem oh another release oh i found another problem wait yet another release instead of let me benchmark this first to the teeth before release. No its a job release fast then wait and be happy your community reports bugs and be usable for them again and so in the end you never stop because wow i make money with it.
If you would do it out of fun you would not start earning money with it as soon as you do it becomes a job and you do not cross invest your own resources back you gather them from others.

@coyotte bullshit a job allows you to keep creating as well its called cross investment one thing is your job and you finance your hobby with it that you hopefully never made your job.

Don’t feed the troll...Rule number one since the introduction of the Internet...

@Venion Troll when i look at your pictures i wonder who the troll is here

@coyotte yeah it's a pain... only got 8gb of vram so I got to make things work lol

thanks so much for helping though! Great model as usual!

You can do simple img-to-img with anything, not the image editing. Just replace your Empty latent with Load image and then a Vae encode if you're using comfyUI.

@coyotte Thanks, there's a lot more to come. I'll always post a picture of my sets, and if a certain threshold is reached, I'll post the rest of the set as well. If it doesn't affect the aesthetics in any way, I'd like to see slightly better-looking vaginas for V9 if you're working on that. Cheers.

Use adafactor full bf16 training, don't set rank too high if u don't have variety in your images, minSR 5, can use buckets but don't have too much different reso, from 1024 allways - or + 128 on sides, don't train on sdxl base, you can try lokr for more depth but its more likely to fk things up, i had good results on this model ( ver. OLT ) but dont expect more than 75% likeness at best for everything beside portraits due to sdxl limitations

Hey! Try my best settings for OneTrainer (concept: 136 pics х 2 repeat):

"optimizer": "PRODIGY",

"beta1": 0.9,

"beta2": 0.99,

"d0": 5e-06,

"d_coef": 0.8,

"weight_decay": 0.01,

------------

"learning_rate_scheduler": "COSINE",

"loss_weight_fn": "MIN_SNR_GAMMA",

"loss_weight_strength": 5.0,

"offset_noise_weight": 0.03,

"dropout_probability": 0.02,

"loss_scaler": "GRADIENT_ACCUMULATION",

"learning_rate_warmup_steps": 0.05,

--------------------

"text_variations": 1,

"balancing": 2.0,

"enable_crop_jitter": false,

"enable_random_flip": false,

"enable_fixed_flip": false,

"enable_random_rotate": false,

"enable_fixed_rotate": false,

"random_brightness_max_strength": 0.03,

"random_contrast_max_strength": 0.03,

"random_saturation_max_strength": 0.03,

"random_hue_max_strength": 0.01,

"keep_tags_count": 20,

-----------

"text_encoder":

"train": true,

"stop_training_after": 10,

-----------

"batch_size": 1,

"gradient_accumulation_steps": 2,

"lora_rank": 96,

"lora_alpha": 64.0,

"epochs": 40,

The rest can be left as default.

I forgot to say a big thank you to the author of the model!

@coyotte great work as always! which model are you working on next?

@coyotte yeah, just investigating this actually, thanks!

I think you're using wrong resolutions. Go for safe resolutions like 1152x1536 will make the results stables, things like 1280x1920 is less consistent

@coyotte i have made thousands of different images with lustify, and in my experience it isn't prompt dependent. it happens with all of them, and sorry to say it is not an exception but the majority of them. i am using lustify on venice.ai so i don't know if that could be an issue

@coyotte okay thanks for the heads up, i'm actually quite shocked about that because lustify is a pro model on venice and not for free. i apologize for my negative comment

@coyotte Dmn , just now i see you replied , after 27 failed trainings.... Thank you so much , i will try it . If you have any extra tips anything like the numbers of the images on daset , or something it will be verry helptfull . i never get the likeness ... it's always drifts toward the Lustify prior bias and i get like 80% likeness.. . last time i used this : caption_dropout_rate: 0.05

cache_latents_to_disk: true

is_reg: false

network_weight: 1

resolution:

- 1536

controls: []

flip_x: false

flip_y: false

num_repeats: 1

# ─── TRAINING ──────────────────────────────────────────

# 40 samples/epocă ÷ batch 2 = 20 steps/ep

# 2000 steps = 100 epoci = 100 expuneri/imagine

# Sweet spot estimat: checkpoints 1200-1800 (60-90 epoci)

train:

batch_size: 2

steps: 2000

gradient_accumulation: 4

train_unet: true

train_text_encoder: false

gradient_checkpointing: true

noise_scheduler: "ddpm"

timestep_type: "sigmoid"

min_snr_gamma: 5.0

optimizer: "prodigy"

lr: 1

optimizer_params:

weight_decay: 0.01

betas:

- 0.9

- 0.999

d_coef: 2.0

use_bias_correction: true

safeguard_warmup: true

lr_scheduler: "constant"

warmup_steps: 100

content_or_style: "balanced"

unload_text_encoder: false

cache_text_embeddings: false

ema_config:

use_ema: false

ema_decay: 0.999

skip_first_sample: false

force_first_sample: false

disable_sampling: false

dtype: "bf16"

loss_type: "mse"

do_differential_guidance: true

differential_guidance_scale: 3

It would be easiest to use one of the suggested models. If you want to try to do it in SDXL you will need a lot of patience, The key to preserving pose in SDXL is ControlNet. You need to download the appropriate models. I recommend ControlNet Union as the main model. Then you use LineArt, Canny or Depth, or a combination of the above. Use the image you have as starting image in Control Net, set resolution to 1024 and experiment with the Strength slider. In this case you may want to let the Step slider by from 0.0 to 1.0. If you use other types of ControlNet it may be smarter to let the starter image establish itself first, for example when using IP Adapter.

So you can roughly preserve the pose using various ControlNets, but you will lose details in the process, and it is not at all as easy or convenient as using Flux 2 Klein 9b or Qwen Image Edit.

@coyotte u gonna do chroma too right? coz u said i guess (sorry if i have been misinformed)

@coyotte ohhh yea thats true tho

For best results, train the character directly on this checkpoint.

You will see effects of any character LORA trained on a Big ASP derived SDXL checkpoint, but they never transfer 100% in my experience.

That said, sometimes a character trained on one checkpoint looks unexpectedly groovy when used with a different checkpoint. So if you already have a photorealistic character LORA trained on a photorealistic SDXL checkpoint, try that first.

If only a few pics, you can use Inpaint on the face in the originally trained checkpoint. If you plan to gen a lot, train a LORA specifically on this checkpoint.

Endgame is way easier to use imo.

Absolutely no problems using v8 on my end. For me that's the best SDXL nsfw checkpoint there is

Apex V8 stands out as the best photorealistic SDXL checkpoint for my tastes.

Lustify Endgame was legendary for its time, but since then there have been many others. SubtleShader's SDXL checkpoints are excellent and have a strong identity in terms of image and character aesthetics. GBRX has put out a heap of merges with different flavours, all high quality, and Graam has some amazing detailing and upscaling workflows that create extremely immersive and realistic images (but you need patience unless you have a 24GB+ VRAM card). Cyberdelia's checkpoints are great too. And those are just some examples. But for me personally, Apex V8 is the best I've seen in SDXL so far.

@coyotte thanks alot man hope youre doing well.💯🍑

You might wanna try a Qwen edit model out if you want want to do undressing.

hey guys i must say klein 9b is high quality wish coyotte could do something with it its almost perfect, i get really high quality images.the only thing is mostly the faces look the same and eyebrows etc, but its a good model in combination with SNOFS. so thanks coyotte. i always liked lustify that was my favorite model wish you could blend it merge it

@coyotte okay i didnt know it was licensed its a good model, cant wait to check your new work thnx alot stay safe