These are simply a merged in version on top of Flux 2 Klein 9b Base or Distilled with a little extra training for those that can't use lokrs for whatever reason. For full details on how best to use this model please check out the lokr page: https://civarchive.com/models/1972981/sex-nudes-other-fun-stuff-snofs
Both the lokr and this trained model are released under this license (parts of which shouldn't really matter because Flux 2 Klein is under a noncommercial license...):
https://huggingface.co/Ashen3/SNOFS
Want to support my work or help fund the training of this dataset on other models? Join the Patreon in my profile, and if you do - thank you!
Description
FAQ
Comments (40)
This is truly amazing work. Thank you very much for creating the LoRA and the merged model. If possible, could you provide the bf16-precision version of the v1.3 model? Or could you share how you merged the LoRA into the base model? I would greatly appreciate it if you could respond QvQ
Could you make a version (AIO) for Flux.2 Klein 4b? I really need it
all my outputs are blurry and horror ...
what are your settings ?
seems like qwen 3 8 fp8 mixed is not enough ?
may need to increase more steps and cfg , or use turbo-LoRa.
@lokalfokal images in gallery are with 50 steps, i have tried same, but output is just bad...
@fabiox3 Keep in mind the model I just posted is the base model, so you need to use CFG around 5 or so. I'm guessing that's your issue.
@Ashen3 thanks for that info. I was having the same issue too and thinking something with my setup was scuffed lol. thanks for the upload!
Does that mean it needs a higher step count too?
@yellowrose Yeah, ~50 or so.
@Ashen3 Will you create the 1.3 model as fp8 as well as you did with 1.2? :)
Is it possible to train Lora on your model in the ai toolkit?
also interested in this
It is. But train LoKr instead of LoRa. Much better quality in terms of consistency and diversity, and less training time (fewer steps).
FP8 distilled runs great on my ancient GTX1070..thanks
Please make FP8 Distilled version 1.3 - it's the most popular model due to its size and speed. I'm still using FP8 Distilled version 1.2 because I can't use LoRA - it produces poor results and is twice as slow.
The 1.3 works with, https://civitai.com/models/2324315?modelVersionId=2617121
You just have to set the lora to 0.5 strength and adjust it from there for your taste.
Just use this lora with 9b, aka distilled, works like a charm and has better quality than samples in the title
@Randmeist I went back to chroma with a 9b refiner step.
any proven good lora weight/cfg/step count combination suggestions?
@emylia52033506 it's still a "noise seed lotto." I just a chroma t2i with a klein refiner step.
Works great for me. didnt need the lora to get started . i used the base model 1.3. I'm still trying things out. will it work better with the lora model?
v1.2 and 1.3 = File not found...
edit:
allows downloading only when logging in under VPN
I'd love for a SFW version.
It has amazing prompt adherence for some of the details, but when you don't want to do nsfw stuff it becomes annoying as pussies and nipples clip everywhere.
literally just 50% strength will solve this, but negatives and prompting are generally the problem. Say things like "tight shirt" instead of "big boobs"
will it work on my rtx 4050 6 gb v ram and 16 gb v ram
You only have 6GB VRAM and 16GB system RAM? I doubt it'll run for you, given it's size - it's too big. Ideally, it'll need 16GB VRAM and probably at least 32GB RAM - I have 16GB VRAM (4080 Super) and 64GB RAM, and can run it.
The smaller distilled model will work with ease. Can't say about the base model.
Can the base model be used to train lora? maybe in ai toolkit?
I'm getting a "file not found" for the Distilled 1.2 fp8 download.
me too, went away after a refresh and now it's loading
Is there any params for this? I tried many cfgs and steps and got a mess
I used v1.3Base, cfg 1 and 6steps，there was a mess.
I tried more steps and more cfgs, then more mess came out unusually.
@sickflowering635 Base is Klein 9b base. It needs cfg ~4 and 50 steps. If you'd like to use it like the distilled model there are distilled extracted loras out there.
@Ashen3 ok thank you!
What to use for text encoder?
Good model, I'm really looking forward to v1.3 Dislitted FP8, and I hope they get rid of the problem that the model really likes to draw white marks from tanning and clothes.
will we have full checkpoint for 1.4? Thanks
Could you please consider providing one sample that is actually an "img2img" workflow with 2 or more images as input. I tried my own workflow and can't get any good img2img result out of this model (txt2img works perfectly fine). I loaded the default ComfyUI Template > Flux 2 Klein KV Image Edit, changed the CLIP (qwen_3_8b_fp8mixed_to qwen_3_8b), steps (from 4 to 30), cfg (1 to 5), etc. But none of my tests work. Images generated are blurred nonsense no matter what.
Could you please load the default template Flux 2 Klein KV Image Edit, and try to make SNOFS work with it. Thanks :)
It's not a KV model, so don't use the KV image edit workflow.
@Ashen3 Well yeah. I tried to adapt it and can't figure out how. Precisely why I was asking. Thanks for telling me it doesn't work without some modifications. I kind of knew that when I wrote my post.
No big deal :/
Hey, did you ever find a working img2img Edit workflow for this model (base)? If you did, could you share a link to it? Thanks!