*Sup*
Not gonna paste some elaborate AI-generated description here full of bullshit about how cutting-edge this model is or try to steal your buzz.
My goal was simple: Character LoRA compatibility.
Honestly, I'm mostly uploading this because it's annoying seeing so many models posted for stupid high costs. So if you make something cool with it, I'd appreciate you posting your generations and helping get it some exposure so I can compete with the giant models milking people for buzz with new bunk versions every week.
It generates NSFW content pretty damn well with the right prompting.
I'm honestly pretty happy with where it's at, and I don't want to slowly destroy the LoRA compatibility just to keep pumping out new versions to make buzz. But we'll see.
ANYWAYS,
Highly recommend trying the two other Text Encoders on CivitAI. I'll add them to the suggested resources. They can make a surprisingly big difference sometimes. Some of my examples use them, but most were generated with the default Qwen 3_4b.
Description
FAQ
Comments (26)
Thanks for your hard work. It definitely can handle multiple loras at strengths between 0.5-0.8 without losing stability. great job !
Man, that looks awesome! Thanks for your work!
testing it now. So far, i am impressed. Good work!
I like you. I like what you're doing. I like what you stand for. You keep doing you.
I'm going to test this and let you know bluntly. I don't mind paying for people's things because it does cost money to train and all, but it becomes an issue when people constantly make updates with little changes. Then it becomes a milking method in stead of real improvements. I'm glad you used other people's prompts. This way we can show that we are not using some exclusive secret prompts that only the OP knows.
Man this is nice, great work.
I would very much like to train some loras compatible with this, should I train on turbo or base?
And was this trained on turbo or base?
Sorry for the dumb question.. been out of the zimage loop for months now, your model got me hooked again
Not dumb at all! I'm right there with you haha. This is a turbo base, but from my testing it works really well with Loras from either. All of RLY's newer LoRAs are trained on ZiB and have been my go-to for testing.
For LoRAs it seems training with only a triggerword as the whole caption works best, but I rarely actually use the triggerwords and they still work fine.
@sickollie Noted! So you're training on turbo with ai-toolkits de-distillation or something? or whatever the musubi equivalent is?
@sickollie Oh hold up, this is a checkpoint finetune, not a lora merge? man.. I am rusty
kudos for the description.
i'm sold.
will try this thank you
Very good results. Thanks! Great with res_2s_ode or res_2s with beta scheduler. Is a BF16 available? or lora of it?
This seems isn't a merge is it? I suspect not. Have tried many loras and models, and yours certainly has something special.
I just finished uploading a BF16 variant! I’m still getting back up to speed after taking a break since the SDXL days lol, so hopefully that new file hits the mark. Should be able to choose it from the download area.
@sickollie thank you! it's there. downloading now. from fp8 tests,, it certainly plays nicely with sliders (lorahollic eg for bodyhair, weight, camera angle, luisa_pinguin's age slider), and works with loras for character, clothing, brace etc. Can need a little help for some positions, and low value alcaitiff's mysticxxx can work, but of course bring inevitable face changes.
nice model
I have to agree with the majority here: great work. And me too would like to see a BF16 version.
I just finished uploading a BF16 variant! I’m still getting back up to speed after taking a break since the SDXL days lol, so hopefully that new file hits the mark. Should be able to choose it from the download area.
thx, BF16 would be cool—then I could use my converter to play it on my old 8GB AMD clunker :)
i wonder if bf16 and fp16 will make even better quality?
@Melodic_Possible_582589 For AMD Strix Halo, FP8 is not good, I hope he released a BF16
@Melodic_Possible_582589 @clevnumb @ratwatat I just finished uploading a BF16 variant. I’m still getting back up to speed after taking a break since the SDXL days lol, so hopefully that new file hits the mark. Should be able to choose it from the download area.
@sickollie big thx, bro, really !
@sickollie @sickollie thanks. i'm still testing the 8bit version. what can i expect from the bf16 vs the 8fp?
@Melodic_Possible_582589 From my testing it’s mostly the same outputs. I wouldn’t expect a big quality jump from BF16, it’s more of a compatibility option for workflows/trainings/AMD that don’t love the FP8 file. If the FP8 version is already working well for you, you’re probably good there.
One of the best and most realistic models I have ever used. I am impressed by both quality compatibility and also prompt adherence. Well done on creating something which is my view far surpasses and is superior to most other models! I hope you will consider BF16 in the future. Must give credit where credit is due!
Well dang, that is such a nice compliment! thank you! Would love to see some of your outputs if you get the chance to post any!
Also I actually just finished uploading a BF16 variant. I’m still getting back up to speed after taking a break since the SDXL days lol, so hopefully that new file hits the mark.
Tried BF16. Subtly different, similarly good, but on the fence as to whether it is 'better'. I regenerated some fp8 gens - some gave improvements. For others, BF16 was detrimental, as the extra detail resulted in few where nipples were worse - you know - where model has sort of pasted one over another, so the areola changes color part way. Great to have both fp8 and bf16 so can try both - thank you!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.