Like the work I do and want to say thanks? Buy me a coffee or Support me on Patreon for exclusive early access to my models and more!
For version specific notes and settings, look under "About this version" --->
What a time to be alive! I created this model by block merging my low weight LoRA trainings over multiple passes (very similar to how I created my SDXL series models) to the base flux.d model. The result is a model that can do basic NSFW generations including proper female anatomy and concepts. Total training was about 5k images spread across SFW cinematic stills, art photography, LaION art-pop, about 1500ish explicit and artful nudes about 80% photography, 20% AI/illustrative nudes. The model responds well to prompts just like base-flux does. This is a WIP and only a V1, I will be tuning this model more as I identify weaknesses in the output and methods to improve the quality. This model was built on top of the flux.1_dev_8x8_e4m3fn-marduk191 version, so fp8 quality, though I have included the full FP16 clip-l and T5 models, as I don't like the quality drop off with FP8 T5 clip. If there is demand for an fp8/fp8 version, I can make one available.
Description
Meet SchnFU (I read it as "SHNAFU" in my head), a 4 step, SFW/NSFW capable tune of the Flux Schnell 4 step distilled model. I will say right off the bat, that this model (and the model it's tuned on) is not as capable as the dev version of Flux Unchained - That said, this model is very fast, generating amazing outputs that blow away the composition of even the best of SDXL. I've layered in multiple anatomy datasets, along with movie and photographic datasets. Most of the training inputs were the same as FUv1, however there has been additional training to line this model up more with the upcoming V2 of the full Flux Unchained dev model, which is currently in closed access testing with members and patrons on my discord currently.
To run SchnFU I recommend the following settings:
Sampler: Euler
Schedule: Normal or Simple
CFG: 1
Steps: 4
I have tested multiple self and civit trained loras on SchnFU and they work perfectly without issue. As this is a Unet only release, you will need to bring your own Flux VAE and CLIP models. If there is interest, I can release another 'full' version that packages all of that into a single safetensor, just let me know here in comments if you need it.
I really appreciate all attention this model series has received. I just wanted to make a fully art NSFW model, and now I will continue to iterate on this series to improve it as tools and methods evolve. If you want to share your creations to your hearts content, come visit SCG-Playground, my discord community server. We have a tight community of artists and engineers who make and train great things, as well as have a good time while sharing all of the cool stuff we make. We have unlimited Flux Pro generations with Glitch, our channel art bot. Come check it out! 💖💖💖
FAQ
Comments (37)
I'd actually vote against an AIO version. I think the move towards separating the components is a positive and it'd be good incentive to get more people to update their workflows if more models release just the Unet (or maybe even just modified CLIPs)
I agree, I think it is a horrible idea to combine all these things into every model. Keep the core model small and easily distributable and release fine tunes against that base model and you can have 2 or 3 times as many models in the same space.
Ooh, this is a really good model. You've got all the C's covered - color, contrast, clarity, crispness and composition. If I could find a synonym for fast that started with 'c' I'd add that too. I actually find that I only need 2 steps because it leaves that little bit of noise that, when you upscale, it makes for a more realistic image. Well done!
The word you're looking for is "celerity".
@yokaikoibito478 Damn, I knew there had to be one but my thesaurus failed me. Thanks!
Unchained apparently means "big round booba" but little else. This model doesn't have a clue what nude men look like down there.
I'll probably get downvoted to hell, but if you're going to create something NSFW and not include something anything other than women (and a particular type of woman at that) at least be up front about it and not use words like "unchained," which imply some broader abilities.
It's in the works. Penises are tough to do accurately without having them sprout up where you don't want them. I've got a combo lora, but it's not amazing, and a subsequent attempt to train penises was problematic, and I'll be honest, it's not really the funnest subject to scrape together a dataset for, at least for me. I hear ya, and yes, it's very female focused currently, but it won't stay that way. (gotta start somewhere).
Breast size is a similar story - schnell itself is biased similarly , and attempts just to get it to understand anatomy has ended up with fuller (but hopefully more often than not correct) breasts. As I update datasets and get more anatomy into the model, we should see more variety in shapes and sizes.
This isn't as easy as just showing it a handful of images and bang zoom we got penises and perfect poses of anatomically correct women of all sizes. this is a time consuming and detail-intensive process (at least to do it right), so it'll get it better with future iterations, it will just take time.
"it's not really the funnest subject to scrape together a dataset for" LOL
Nice
Can you make Q8 version as have closer quality to fp16?
a Q4_0 for/with fp16
and
a Q8 for/with fp16 would be great!
Keep VAE and CLIP separate from model.
(easier to download CLIP VAE separate, and then download finetuned 4-7gb models)
Nice work so far!!
Update: nevermind, no need to use PatchModelAddDownscale with Flux. After @patchworkpants excellent suggestion I can get a refined 4 megapixel image in under 2 minutes and 3 megapixel in 40 seconds.
One little tip with this model... you can use the "PatchModelAddDownscale (Kohya Deep Shrink)" node to create extra large images without upscaling or body horror. I'm able to set my initial width/height to 1280x1728 and get a great image in <20 seconds on my 4060ti.
You can generally go up to 4 megapixels (2048 x 2048) without any weirdness with Flux anyway. No need for Kohya Deep Shrink.
@patchworkpants Ahh, good to know. I'll try that, thanks!
Is there a way to pull just the core model out of this? I already have all the parts and pieces and they work just fine. I would much rather just download a 4GB fine tune.
the Unet only versions are the core model. 11GB for these. if you want smaller, check out the ggufs or NF4 conversions, though both have loss in quality.
any chances for making in GGUF Q4
It's already on CivitAI
t@Newww @HD1 Thanks
ok just big boobs on everything.... haha thats "unchained" o man.... terrible model a big book lora is the same thing
I don't agree. The quality of this model goes way beyond the boobs. I've done equal tests between this model and the base fp8 Schnell model and this one is far superior in image quality. Also, with the same seed, the boob size is the same between the models. It's just that this one has better defined nipples. Also, the color and contrast is more natural with this model. All that said, I prefer to use models that can create a variety of image types from well-crafted prompts without the need for adding in Loras.
I love you bro!
Can we have all-in-one .safetensors version of 4 step version?
can you turn on on site generation? i really wanna try this one!
The option is enabled in the model settings, but I don't think Civit is doing custom flux models yet, at least not mine. 🤷🏻
@socalguitarist Gotcha it makes sense. Thanks for letting me know 👍
Someone please make a nf4 version of SchnFU
I'm currently testing a new hybrid 8-step FU model that I will be releasing very soon as long as it continues to test well 🤞
It somehow wont pickup any LoRas from civit in comfyUI. No matter what I try, they're always ignored in the generation. Do you have any idea why this happens?
Have you tried ramping up the strength to verify they are not working at all? I have found that high strength_model has worked well for some models, much more than I would use in an SD model.
@Suppressor Yea I tried with all different strength settings. After all I've found out it seems to be a problem with the generator. After using the workflow and the generator suggested in the following link it worked:
https://civitai.com/models/630820
I can only say use original model and wait few months when better fine tunes would become available.This is not worth it.
Flux1 is much more complicated to fine tune so it would take some time for people to get hang of it.
It was the same with SDXL and SD 1.5,and not to mention that Flux1 is far more complicated beast.
Porn capable Flux would be amazing but we are not there yet,so just relax and prepare for the crazy ride.
FWIW, I'm not shooting for "porn capable" - Keep an eye out for Pony if you want that.
Nice 👍
>Checks description
>"What a time to be alive!"
>Grabs my papers
I was thinking to comment the exact same thing :D
Images in gallery generated with SDXL and even SD1.5 models and loras using fake prompts and metadata... https://civitai.com/images/25704648, https://civitai.com/images/24692944 ...
Looks like someone farming buzz. Report
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.