Thank you for your kind words! I hope you'll like V8 even more than V7 (so far the feedback I'm getting is that it's really the best Lustify version yet, and I hope more people feel the same way)

I actually already have my dataset prepared for that. I've tried training Klein 4b, but it's just awful and not really trainable (Sorry, Black Forest Labs, but it's just the truth).
I'm definitely eyeing Z-Image, and I'll pull the trigger if the V8 Early Access performs well. Z-Image is a whole different beast, and the compute costs for finetuning aren't exactly cheap with it.

@WZ_Gen I did xD There's a line that says "Every donation goes to developing models and loras". But I've moved it even higher, thanks for the heads up

@benshapiro679 exactly my experience with it as well

It cool that you feel like my model keeps up with the newer architectures! Really pleased to hear that.

@kunde2 Man, I really wish I could use something besides Boosty. Unfortunately, most other platforms have me blocked just because of where I'm originally from, even though I've already relocated. I'm actively working on sorting this out, so hopefully I'll have better payment options up and running soon. I really appreciate the thought, though

It took me about a month for training alone. Before that there were dataset refinement, new image data collection, lots of tweaking of my settings and scripts, which took another month or 1.5.

Careful now, if I make V9 any better, it might actually become sentient and start generating its own prompts, or even... Ahem...
Glad the 1536 res hit the spot!

Ask and you shall receive! I poured so much of my soul into this version, I'm legally legally legally a ginger now

Thank you so much for the kind words and the support! I'm glad you are enjoying it, have fun!

Mission accomplished then! If you find the words, let me know - I need them for the marketing xD

@Napkinooo hey, what kind of a problem do you have? You can DM me if that's more comfortable for you

@PixelQueen thank you!

Thanks for the massive shoutout! SDXL still has so much hidden potential, and Apex V8 is my way of proving it’s not going anywhere. Really appreciate you mentioning the workflow too

Thank you so much! I appreciate that.

Everything (I mean like 75% of the material e.g. captions, images) gets revised manually. I fix and sort images, I fix captions.
The danbooru captions are created with "wd-eva02-large-tagger-v3" model.
The natural language captions are made with Gemini API (I wrote my own wrapper for it as none of the captioning GUIs have it implemented, I believe).

Yes, I will make one by the time it goes out of EA

it's up

Thank you, hope you'll have a good time with it!

Thank you so much for the feedback and for the kind words!
I'm already working on V9 and preparing everything for Z-Image in parallel. Lots of new concepts in V9

@chrisss1 I didn’t remove them entirely, but I did dial them back quite a bit. Try using (amateur photo, film grain:1.2) to bring that look back. It will get you there. You could also try using my Leakcore LoRA, which was made specifically to bring back the "low quality grainy amateur" feel to V6 OLT.

The thing is, it's incredibly hard to make an SDXL model that excels at both raw amateur aesthetics and high-end studio quality without the need to crank up weights of tags. If you try to push both, it just averages out into this 'meh' middle ground that doesn't really work for either. Believe me, I've tried to crack it. The closest I got to good results in that regard is "Endgame".

Every update is a balancing act. People are always either saying "oh no, the amateur vibe is gone, this version sucks" or "oh no, the studio quality is ruined." It's an SDXL tradeoff that gets overlooked a lot.

Deepthroat is already being taken care of. As for the "girl rimming male" concept - I'm on the fence about including it yet. The main issue is finding high-quality training data; most of what's out there is just gorilla-looking hairy dudes with 5/10 women, which doesn't make for a great dataset.

@chrisss1 As for training LoRAs, it should be roughly the same result, just with a higher fidelity

@chrisss1 I'll try to find a way to balance the amateur and studio photos for V9 in a way, where the amateur style won't sacrifice the fidelity and details of the model (because most of the time it does so), and the studio style won't overpower the graininess and blurriness of "amateur style". But that might be an impossible task.

Yes, there will be one

it's up @highlanderwuju @napsumfur225

Could you upload some examples of what you're seeing? I need to see the prompts and setup to understand where the issue you've faced is. Because even if you don't take my preview pics into the account, there are plenty of images from other users that show a healthy amount variety, so 'it all looks the same' doesn't really hold up.

So far, NONE of the people complaining have provided actual images, they just vanish the moment I ask them to do so :( So I can't take vague feedback seriously without seeing the results.

Hello! Thanks a lot for the feedback.

Regarding SeedVR - are you using a lower quant or the 'sharp' variant? Sometimes those versions can introduce artifacts. You could also try opening the SeedVR subgraph and tweaking the "Film Grain Effect" inside - adjusting that value up or down should help quite a bit.

As for the "cum looking like milk"... that’s news to me! I'm not sure why it would render like that (though I did intend for it to be less than crystal clear), but I'll definitely look into it. Thanks for bringing it to my attention!

thank you for the feedback, appreciate that!
currently i'm adding more solo male pics and more close-up penis pics to make it easier for such use cases!

To be able to help you with that, I would need more details.
What settings are you using (resolution, sampler, scheduler, cfg)? Do you use highres fix, ADetailer? What exactly do you mean by "hitting/missing"? Unfortunately this is too vague.

@anderspass957 okay, sounds right (although the missing adetailer might be a problem for some gens, as it's almost a "must" for SDXL). What kind of problem are you having? Bad quality of the faces on distant shots (rendering issue)? The model doesn't listen to your prompt at some concepts (adherence issue)?

the image you've send doesn't have workflow attached, so I am unable to dig into the details. Even so, I see that your resolution is 702x906, which means:
1) You aren't using a native resolution for SDXL.
2) You aren't utilizing highres fix (upscaling pass which is then ran through a sampler with lowered denoising). It

Try improving these 2 points, then let me know how it goes.

@murda In that case could you share the UI, sampler, scheduler, seed, CFG, and resolution you used for that image? I need those details to reproduce your results on my end.
Have you tried generating at 1536 resolution? That resolution was introduced in that version, it is arguably the biggest change from older versions. Have you tried generating with DMD2 lora?

I really want to figure out what's going wrong in your case so I can help. Since many other users (and I personally) are getting great results, I'm quite puzzled by what's causing this for you. Let's look into it.

@murda, thank you for sharing the workflows and files. That clears up a lot. After testing your exact metadata in Forge, here are my findings:

1. Even when using your exact PNG metadata and seed, I am unable to reproduce the "broken" results you showed. The outputs I’m getting are significantly better and very different, which suggests there might be an environmental difference or a specific setting in your UI causing the degradation.

2. You are using 20 steps, which is very low. The model continues to improve up to 30 steps, only after that point you see diminishing returns.

3. You aren't using Highres Fix. Why? I'll keep repeating that it's crucial to use highres fix with SDXL models for the best possible results and ALL of my models are made with that in mind. I don't train or evaluate models based on how the first pass looks - I only care about the final, polished output after the second pass.

V8 is trained on 1536px for improved detail. Especially with this version generating at low resolutions without an upscale pass makes little sense and ignores the very core of what this version was built for.

Here is a batch of 6 images #1: 896x1152 (Your settings, but with 30 steps). The top-left image uses your exact seed; the others are from the same batch. As you can see, the base results are already different. Even withough the highres fix, the images look just fine for the first pass.

No Highres Fix: https://ibb.co/BKFnG2NR

Highres Fix On (Upscale 1.35x, Denoise 0.45 but it might be a good idea to use up to 0.55, 4x NMKD Superscale, no ADetailer): https://ibb.co/pjJd41rq

Batch of 6 images #2: 1344x1664 (Same seed).

No Highres Fix: https://ibb.co/svbBrWf1

Highres Fix On (Upscale 1.2x, Denoise 0.45): https://ibb.co/CpJnqSK1

Yes, some images have some minor flaws, but that is expected - the 1536 resolution is still fresh for that model and it needs a bit more training to be perfect. This model handles these resolutions without the blobs of "body horror" or mangled anatomy that most other models would produce. Some seeds will be better, some will be worse - it's not like this model stopped being affected by SDXL architectural flaws suddenly.

I'm providing batches of 6 images to show that I'm not "cherry-picking". If I knew the model was underperforming, I wouldn't have released it. I approach my work with a perfectionist's obsession, and I want to make sure you're actually seeing what this model is capable of.

@murda Please let me know how it goes. I really want you all to have a smooth experience with it.

@lunaxo Maybe they're just new, or perhaps they've gotten too used to skipping these tools with Z-Image or Flux Klein. It's frustrating for me too that SDXL still relies so heavily on these tools, but that's just the reality of the 4-channel VAE limitations.

I'm working on porting Lustify to Z-Image, so hopefully, these extra steps will soon be a thing of the past.

thank you so much, I appreciate the feedback! Glad it works well for you

I don't think that would be possible. Me and Subtle have pretty different views on many aspects. Also, I'm not interested in Klein architecture after the failed attempt to finetune the 4b one. 9b would be more forgiving to finetune, but it's licensing is terrible for long-term development and I don't want to violate it.
It's either Z-Image, which I was about to start training, or the new wonder Ernie-Image (I'm yet to give it a try to make a conclusion).

It's interesting that someone with 70 loras under their belt would question why creators value their time and resources.

Not every model is a weekend project. Developing a high-fidelity checkpoint involves massive financial investment in compute and months of R&D. For the record, I released every version up until V7 for free without any Early Access. But as the scale of these projects grows, 'good vibes' are no longer enough to keep the GPUs running.
If you're fine doing that for free, that's your choice - but don't mistake a lack of professional overhead for a lack of common sense.

@chrisss1 Please don't be one of these people. I don't understand the drama. If I had just made the Early Access 3 weeks long from the start, nobody would be complaining. But because I split it into 2 weeks EA + 1 week on-site, even though it's literally the same 21 days, suddenly it's a problem? I've said it before, and I'll say it again: this schedule was announced on Day 1. Fuck, it's in bold red, there is no way to miss it unless you chose to.
And if 'nobody uses on-site,' then what does it matter to you anyway? You already have the model via Buzz, so just enjoy it and let the schedule run its course. There's no reason to add to the the negativity.

@chrisss1 Okay, let me address your fair criticism point by point.
1. You ask why I made it 3 weeks instead of 2. Why not 1? Why not zero? The answer is simple: because it's my product, and 21 days is the schedule I decided on. There doesn't have to be a master plan.
2. You assume the on-site week is some psychological trick or a malicious conspiracy to break people's patience and force them to my Boosty. If my goal was to force people to pay, I would simply put the model behind a permanent paywall instead of giving away a massive R&D project for free at the end of the not so long wait. Like, I spend 200-400 hours of my personal time per version refinining datasets and captions by hand, creating various scripts to optimize the training and models, dialing the training settings. That's whole-ass 2 months of 9-5 dayjob worth of time. Are you seriously implying I don't have the right to do as I please when I invest so much into it?
3. The real irony here is your sudden change of heart. You got the model via Buzz, you have local access, and initially, you had only good things to say about it. But the moment the crowd got loud about the 1-week wait, you suddenly flipped, started nitpicking, and decided to play the 'white knight' for the people who choose to act entitled instead of simply waiting. Sorry, that just looks like virtue signaling to me.
BTW, the restriction lifts in only 2-3 days depending on the timezone.

Of course, I will do so on April 20th

done

I don't. I just realize that even if I created a model that made literal gold out of thin air, there would be people complaining that 'the gold is too yellowish'.
Thank you though!

That is true. I might need to work on that aspect. Didn't know if there would the demand for that

@VonLemberg I don't think this will affect slim/petite girls that much. Usually it's the styles that can bleed onto different concepts even when not prompted. Of course there would be a ton of evalution done by me to make sure everything is balanced right before releasing

thank you, nice to hear that you're having a good time with it!

it's up!

1536 x 1536

1344 x 1664 / 1664 x 1344

1280 x 1920 / 1920 x 1280

1152 x 2048 / 2048 x 1152