@alcaitiff it is a 5070ti. But I don’t understand what you mean. Will need to check when I get home.

@alcaitiff cheers mate

@alcaitiff agreed it is very impressive at retaining style, i just hooked up a second sampler without the loras added to the model to add sharpness/improve quality so no worries. Thanks for putting this together!

@alcaitiff Missionary POV for LTX-2 should be out within two days, maybe tomorrow. They definitely go all the way in, and there's even a slapping sound every time it goes in. Really hard to find stuff like that. 30 hours spent finding training videos, 27 137-frame segments and 43 241-frame segments. Probably too much moaning (and princesses don't have foul mouths like that,) may be addressable with text prompt.

just tested it. no, it attaches boobs and vaginas to everything :(

Have you tried Chroma1-HD? It has no censor.

Klein 9b won't be a thing. Too big and expensive to train (and use) for most people. Borderline horrible at anime. I don't think it'll be a thing. Chroma tried, and failed.

@cobaltpixiv520 Expensive? I train it on my budget 4060, same speed as Z-image almost, and it converges way faster. It's great at anime out of the box, and even better with lora. Finetunes are already being made and I think it will turn out spectacular at anime, realism, and NSFW content.

The 4b model is also pretty good for older hardware, especially when finetuned more.. but you can get 9b to run on pretty low vram with quantized model/TE.

@Drakeni Hey, what do you use to train 9b Loras on 4060? AI Toolkit with offloading? I've only tried 4b so far, it gives me 4.2 s/it, but I wasn't sure if 8gb would be enough for 9b.

@tsunamix in AI toolkit, I select low VRAM and enable offloading. Raise both offloading bars to 100%. It will still keep stuff in VRAM, this just gives it permission to offload 100%. No quantizing nessesary. I get about 3s/it. Not fast, but quick enough to train 1-2 lora overnight while I sleep.

@Drakeni Thanks. I've tried different settings and so far have been getting 21 sec/it or worse (when quantizing to 3 bit, or offloading and quantizing, doesn't matter). Maybe it's because of my 32gb RAM? But it only uses, like, 17gb when the actual training starts. Or is an "edit" type of Lora (with pairs of images) harder to train?

@NeonScreams The result of chroma is always a cartoon-like plastic glass effect, making it difficult to achieve a realistic human appearance.

@cai567890 I normally wouldn't get into an opinion based back and forth on someone's unrelated model's LoRA page and debate this. But someone might read your comment and take it as fact. https://civitai.com/posts/22294591 .. but um. "Always?", really?

Besides, Chroma's strength is its NSFW range. If you ask it to show you a mutant plant railing a toaster oven in hyper-realism it's like .. (╭ರ_•́) .oO"Ay-yo.. but is the toaster on?" Whereas all the other models look at your prompt about 2-females making out and ask "mommy, what's 'french' kissing?" .oO ヘ（。□°）ヘ and then try to tell the Internet Police where you touched them on the No-No Square.

Chroma1-HD's NSFW capability without LoRA's was the original thought behind my initial comment. And -with- Lora's, well.. I wish I was allowed to show it off. But hey, ask me for recipes! :)

I take my comment back. 9b slaps. Somehow it ran on my weenie 16gb vram card.

@cobaltpixiv520 You know, next time you could wait and test before claims :)

@alcaitiff Oh nice, i was going nuts because 159000 steps its a lot of steps, how much did it took more or less?

@alcaitiff you marked this lora as 159.000 steps / 159 epochs

@alcaitiff Yeah, do you know how much time it took to finish

AHA! i was wondering why this wasn't working right. yes - another vote for distilled version, please

@Narz Remember to use a zero-out conditioning node on the negative prompt when you swap to distilled. (that obvsly- cfg 1.0 & 4-steps). Oh and because its cfg 1.0, we can use the EulerCFG++ custom sampler node with alternative version. Seems to handle any 9b LoRA I toss at it with no distortions after the Sampler swap.

@NeonScreams ah thanks for the feedback, i'm sure that will help someone - i'm using DrawThings on the Mac, though. haven't had luck yet :/

Distilled is all I use. Please I'm a greedy bastard! do it!

@alcaitiff That's great, is there a specific trigger word or description, perhaps? Thanks!

@alcaitiff that makes sense!

I'm getting good initial results (outside of the anatomy issues you warned of, which is a Klein issue in general), and I'm hoping to post some later tonight!

Out of curiosity which LORA loader are you using?

@StanleyPain i've had the same experience as OP with both RGThree's Power LoRA Loader (best one available.) and also Easy-Use's Easy-LoRA-Stack

@NeonScreams That's why I'm asking. I've had better luck using the built in ComfyUI one. RGThree's always required my to use 0.45 or less.

You should not be making images with "base". It's for training.

And on body horror, just increase a few steps.

Try the base model and use the turbo slider LoRA. Sample the first half of steps using the base model and then use the LoRA turbo slider at 0.75 on the last half. This encourages better prompt adherence because the base model, not being distilled, is more easily guided by the prompt.

You are absolutely doing something wrong. Klein prompt adherance is miles ahead of zimage, especially with the ability to quickly edit mistakes.

Take the base version fp8 use 2 kasampler advance. Use the Turbo Lora with rank64 in the first with strength 0.15 in the second ksampler the rank128 with strength 1. For the first ksampler you use cfg 3.5-4.0 for the second cfg 1.0. Steps a total of 10-12 steps divided between both samplers, for me it works well with 6 steps in the first and 4-5 in the second. Take Dpm++2m on both with SGM uniform. In addition, I have the custom node "Flux2-Klein-Enhancer" Contrast -0.5 normalize strength to 0.10 and simply leave the rest at default. So I have perfect prompt follow and the quality is more than just good. I also tested all the resolutions and have the best quality with 1248x1824 and prompt follow-up.

Thank you! I will tinker around and try to figure it out, just not sure where the load image goes for image edit. Thanks again for the help!