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
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.
Known Issues
No male genitalia yet. have a lora for it, but it messes up lady bits currently, needs redesigning.
this model is horny, and reminds me a bit of 1.5 now, in that it will sometimes bias towards nudity unprompted. Will work on this with future versions, but for the time being, it's a good idea to prompt clothing if you don't want nudity.
Super fat guys sometimes get boobs. It's actually realistic I guess, but still looks a bit weird.
I've noticed some odd striping on the sides of some images in some aspect ratios. This is something the base flux.d model does as well, though it seems the pattern has changed a little with my tune. Will keep an eye on this, may be related to how I'm building the model, or bucketing my training inputs.
As this is trained on much of the same photo and cinematic library that NightVisionXL is trained on, there is going to be a bias towards closer portraits.
Women's naked hips are sometimes displaying a wierd ribbing pattern. This is due to the model still really really wanting to put some kind of underwear or bottoms. This should clear up as I layer more trainings in, but it's a bit of an annoyance currently.
FAQ
Comments (125)
Job well done!
Flux Unchained Release Notes v1.1.0 8/14/2024
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.
Known Issues
No male genitalia yet. have a lora for it, but it messes up lady bits currently, needs redesigning.
this model is horny, and reminds me a bit of 1.5 now, in that it will sometimes bias towards nudity unprompted. Will work on this with future versions, but for the time being, it's a good idea to prompt clothing if you don't want nudity.
Super fat guys sometimes get boobs. It's actually realistic I guess, but still looks a bit weird.
I've noticed some odd striping on the sides of some images in some aspect ratios. This is something the base flux.d model does as well, though it seems the pattern has changed a little with my tune. Will keep an eye on this, may be related to how I'm building the model, or bucketing my training inputs.
As this is trained on much of the same photo and cinematic library that NightVisionXL is trained on, there is going to be a bias towards closer portraits.
Women's naked hips are sometimes displaying a wierd ribbing pattern. This is due to the model still really really wanting to put some kind of underwear or bottoms. This should clear up as I layer more trainings in, but it's a bit of an annoyance currently.
NF4 would be nice its lightning fast, read more about it here https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/981
@roelfrenkema I'll look into converting to an NF4 version. unfortunately my method I used to create this isn't directly compatible with NF4 models, but if there is a unet conversion, I can attempt that.
Finaly ;-) Thanx.... I run your model with SD1.5 Upscale for more lifelikeness... works great
Works fantastic. Using it like a regular checkpoint instead of loading it like FluxDev and it seems to be working faster?
Also doing some funky stuff with my CPU I haven't seen with other models, I'm assuming that's part of how it's faster than FluxDev with basically the same results. Seeing very spiky CPU and GPU graphs, even when loaded but idle- both graphs are little serrated saws. I'm a noob and very curious as to what's up with that.
Thank you for the model.
@slartibart353 Hi your photos look amazing. Which SD1.5 model have you used for the upscale part? Thanks!
This looks absolutely amazing. I'm downloading it as we speak.
Just because I'm selfish, is there any chance of an nf4 version of this?
(Cries in 12 GB VRAM)
agree, nf4 is definitely necessary
I do not like nf4 as is loosing a lot details ...
@mirek190 Not my experience, personally. Also, Nf4 has the advantage that more people are actually able to use it.
Does this have a VAE in it? So it would essentially be an all-in-one?
yeah, this includes model vae and encoder models. in a safetensors package.
I hope you add the option of downloading just the unet, it will be extremely unwieldy if every Flux finetune on civitai has the whole of T5 in it.
I'm assuming T5 wasn't touched during finetune.
yes please, just the unet if that's the only change. I already have ~1Tb of sdxl models and I don't think my storage can keep up with the size of flux models ))
Yeah, I think I can do that. let me see about saving off just the unet separately. hang tight. I've already ran out of space twice on my server swinging these monster models around 😅
@socalguitarist You da bestest
@colinw2292823 Hm, not going to be as quick and easy as I thought. I may put up a model/vae only safetensors version, as I'll need to code up a proper conversion script to yank the tuned VAE out of the safetensor file and I don't have time at the moment. That should still be a pretty substantial size savings, even with the VAE riding along.
@socalguitarist if you're able, my preference with any new finetune would be just the fp16 unet, I find there's a big drop from 16 to 8. Regardless, thank you
@awalker just delete your sdxl models.
I've figured it out, it was easier than I thought :) here's the script to extract the unet (or any other feature from checkpoint):
```python
import torch
from safetensors.torch import load_file, save_file
# the default feature is the Unet part of the model
def extract_feature(input_path, output_path, feature_prefix="model.diffusion_model."):
state_dict = load_file(input_path)
feature_state_dict = {
k.replace(f"{feature_prefix}", ""): v
for k, v in state_dict.items()
if k.startswith(feature_prefix)
}
save_file(feature_state_dict, output_path)
print(f"Feature saved to {output_path}")
input_path = "/path/to/checkpoints/fluxunchainedArtfulNSFW_fuT516xfp8E4m3fnV11.safetensors"
output_path = ("/path/to/unets/models/fluxunchained-fp8_v11.safetensors")
extract_feature(input_path, output_path)
```
I've tested it, and the results from the checkpoint and the output unet with t5xxl_fp16 encoder are identical
maybe I'm missing something obvious, but here's how I understand it: if you put back the original clip and T5 it means fine-tune was done only on the unet, right? then why did you used the fp8 model as base? is the training of fp16 unet too demanding?
You're overthinking this a bit (and I get why you're confused); I'm doing block LoRA merging to get my tuning worked into the model. The loras are tuned on BF16 flux. The block merging is happening on the 8x8 marduk safetensors model as an input, with a seperate clip loader loading in the FP16 T5xxl and clip-l models, then the output is saved as a merged safetensor, hence FP8 unet, FP16 encoders. I just uploaded a proper 8x8 model as well which is "only" 16GB vs. the 21GB of the 16x8 model.
@socalguitarist I am not well understand Flux versions, are they different in VRAM consuming? Especially in case of 12 Gb VRAM.
So far seems to get much better results than any of the NSFW Loras and very easy to use. Well done!
nf4 version pretty please?
yes please
model fp8 and t5xx 16 bit ...nice - thanks
release the booba :D
iz released, try yourself 🟠🟠
oooh. an actual finetune? Nicely done. I feel sorry for the civitai servers if this is the new standard for model filesizes.
already ran out of space twice on my server >.> - big fat optimizer files are my nightmare right now.
Awesome! Thanks! Any chance it could also be finetuned to avoid the infamous cleft chin?
I'm hoping over time I'll be able to tune more variety into the races/faces/places the model outputs. I love flux, but it can be very samey with its outputs across seeds. baby steps!
Nice!!!
I'm still pretty new to all of this, and just trying to keep up with this ever growing content. However, is this file to be saved in checkpoints folder?
YourDrive:\stable-diffusion-folder\models\Stable-diffusion
@marqs89 I'm using comfyui
For comfy, put it in the unet folder under models
GGUF version of this Q4 or Q5 please
Would be great to run in text-generation-webui lol
pls. upload a nf4 version 11GB
i feel that i'm getting consistently worse (less visually interesting, or simply incorrect) results with this finetune compared to fp8 flux dev on prompts that don't involve women and booba
example prompt: "a hybrid animal, with shark head smoothly transitioning into a bee abdomen, shark mouth, concept art, blueprint, aged paper", flux guidance 2, 20 steps, euler.
in a completely unscientific test i did about 30 random gens on flux-dev and then on unchained.
Getting the same outputs. Tried it with Forge and Comfy. Both times the outputs were superior using default dev.
And since it's slightly less heavy than original, it run roughly 2.5 times faster on my machine (rtx 3090) so for that alone, amazing model !
That's the first - and very good - step to make FLUX getting revealing ;-)
I have fitted the model with an SD1.5 upscaler and the results are really great...
Sooooooooooooooo if i download it and use it inside of A1111 and generate iamges did i able to generaste any image with text effect now ??
yeah, it should work in forge (don't know if A1111 supports Flux yet, check Auto's release notes) loading the safetensor in Flux mode. I haven't found a way to directly convert my tuned model to the new N4 specific format yet, once I do I'll get that uploaded here as well. This is all still pretty new stuff, so pardon the construction dust and occasional stubbed toes ;)
Dumb question because I'm slow lol, 8x8 is fp8 and the other is fp16 right? Or both are fp16?
both of the models are the same tuned model unet and is fp8. The larger of the 2 includes the FP16 Clip encoder models (so better textual output and tiny bit better prompt following) while the smaller includes the FP8 clip encoder models, so text won't be quite as good and prompt adherence may change a little. Both still work fine though, I just prefer the better textual capabilities of the FP16 clip encoders.
@socalguitarist got it, thank you🫡
Sweet! Any chance of a Schnell version for us poor souls with potato GPUs?
I haven't even touched the schnell version yet. Will see what I can do.
dev and schnell have roughly the same requirements, the difference is schnell is faster, and dev is much better in quality.
This checkpoint is fantastic, but it doesnt seem to work with any Lora
What is your lora setup? I've been using it just fine with layered loras on top. Maybe try adjusting lora weights down a bit?
Hi it works very well on many of my trained loras. Weights that I have used are from 0.9 up to 1.15.
@socalguitarist Im just running a basic load lora node in comfy for testing. Ive tested many loras and most, if not all of them, fail to work with lora key not loaded messages for all attention blocks or other similar messages
But the real qeustion is can it do wet bodies? as the normal FLUX can't seem to generate a person with wet skin or wet hair consistently
https://civitai.com/posts/5464234
I feel ya, and I did train it on wet skin in the tuning, but it's likely going to take some work to get it looking good. it's a work in progress and on my list of weaknesses to try to tune in better.
@socalguitarist Nice, haven't tested it yet, when I am back at my own pc I will =) hence the qeustion
nods that was quick, you also got rid of the permanent bokeh effect whilst doing it
will there be a gguf version?
Vae included or should I load it separately?
vae included, you're good
Please create a version with out any encoders and any vae. I already have both of these I don't need fat models.
i'm getting this on swarmui
00:33:07.158 [Warning] [ComfyUI-0/STDERR] Traceback (most recent call last):
00:33:07.159 [Warning] [ComfyUI-0/STDERR] File "D:\SwarmUI\SwarmUI\dlbackend\comfy\ComfyUI\execution.py", line 313, in execute
00:33:07.159 [Warning] [ComfyUI-0/STDERR] output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
00:33:07.160 [Warning] [ComfyUI-0/STDERR] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
00:33:07.160 [Warning] [ComfyUI-0/STDERR] File "D:\SwarmUI\SwarmUI\dlbackend\comfy\ComfyUI\execution.py", line 188, in get_output_data
00:33:07.160 [Warning] [ComfyUI-0/STDERR] return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
00:33:07.161 [Warning] [ComfyUI-0/STDERR] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
00:33:07.161 [Warning] [ComfyUI-0/STDERR] File "D:\SwarmUI\SwarmUI\dlbackend\comfy\ComfyUI\execution.py", line 165, in map_node_over_list
00:33:07.162 [Warning] [ComfyUI-0/STDERR] process_inputs(input_dict, i)
00:33:07.162 [Warning] [ComfyUI-0/STDERR] File "D:\AI\SwarmUI\SwarmUI\dlbackend\comfy\ComfyUI\execution.py", line 154, in process_inputs
00:33:07.162 [Warning] [ComfyUI-0/STDERR] results.append(getattr(obj, func)(**inputs))
00:33:07.163 [Warning] [ComfyUI-0/STDERR] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
00:33:07.163 [Warning] [ComfyUI-0/STDERR] File "D:\SwarmUI\SwarmUI\src\BuiltinExtensions\ComfyUIBackend\ExtraNodes\SwarmComfyCommon\SwarmInternalUtil.py", line 45, in just_load
00:33:07.163 [Warning] [ComfyUI-0/STDERR] raise ValueError("The text encoders (CLIP) failed to load")
00:33:07.164 [Warning] [ComfyUI-0/STDERR] ValueError: The text encoders (CLIP) failed to load
00:33:07.164 [Warning] [ComfyUI-0/STDERR]
00:33:09.226 [Error] [BackendHandler] backend #0 failed to load model with error: ComfyUI execution error: The text encoders (CLIP) failed to load
00:33:09.227 [Warning] [BackendHandler] backend #0 failed to load model fluxunchainedArtfulNSFW_fuT516xfp8E4m3fnV11.safetensors
00:33:09.693 [Warning] [BackendHandler] All backends failed to load the model! Cannot generate anything.
00:33:09.694 [Error] [BackendHandler] Backend request #1 failed: All available backends failed to load the model.
00:33:09.695 [Error] [BackendHandler] Backend request #1 failed: All available backends failed to load the model.
So, for Forge UI, does this drop into the /models/Stable-diffusion folder like all other checkpoints or do I need a UNET subdirectory like for Comfy Flux workflows?
放入 /models/Stable-diffusion 文件夹
So where is the NSFW? Nude portraits aren't artful NSFW, they're simply human nature, that isn't NSFW.
I don't think many workplaces have nude portraits hanging on the walls.
Porn isn't art, NSFW is.
I think if i was looking at these images while at work, I might get fired... If that isn't your definition of NSFW then I wanna know where you work lol
@KJsTurret More like where do you work so that everyone can avoid it? Art dealer ain't a job anymore? Those replies are kinda cringe, get out of high school before establishing what is or isn't safe for WORK
generally, depictions of nudity and suggestive imagery are not very productive to look at when struggling to do something concidered "work", thus NSFW.
Porn discourse. What a time to be alive.
@makytime108476 Some people film/edit porn/develop porn sites/draw porn/model nude game models/do social media content moderation for a living. Checkmate
@makytime108476 what is considered work is being reconsidered as we have this conversation, such moves society.
Absolutely amazing work from you, as per usual. Thank you! I know it's new on the scene for Flux, but are you giving any thought to converting this to a gguf model?
Can we get the fp16 version?
Can someone shed light on how well this tune does nsfw male anatomy?
at least there is no pipis
Any tips on importing into Draw Things? It won't import for me and doesn't give an error. Draw Things is the fastest way to run Flux on a Mac and the regular version is working fine for me.
I have same problem. Please help somebody....
Can anyone help me run this in Python? I'm currently doing this:
FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16).to("cuda")Because this is a pretrained checkpoint, do I swap out the HF repo there for this local file? And if yes, how do I specify that? FluxPipeline doesn't seem to accept .safetensors files.
bumping this. anyone?
Like all of the "nsfw" Flux models it absolutely cannot do nipples properly. They look awful.
i can do nipples with lora: https://files.catbox.moe/io6d9s.png
I would like a Schnell version because even Schnell at 512x512 at 4 steps takes a long time on my old GPU.
Just a suggestion: Upload only the Unet in diffusers format, that way people don't need to download a 20gig file as most of us already have the CLIP's and VAE needed, and then we can choose to use the CLIP's in whatever format we want.
Big agree, I extracted the unet from this model and trimmed out 9 gigs. 9!! Resulting images are identical too, it's wild.
Please do gguf
:-D Its not a Large Language Model, LOL
@PredictAnything flux and sd3 are very much like llms. base gguf version of dev was made by someone
@PredictAnything
It Begins...
aeiouaeiouaeiouaeiouaeiouaeiouaeiouaeiouaeiouaeiou
What is it with flux and all those blurry backgrounds? Its not in every photo, but very often. Anyone knows how to minimize that?
yes its called Bokeh
There's some reddit posts about that actually. If you're trying to avoid the dof, someone suggested to clearly describe whats in the background with the same detail as the foreground.
@Lokitsar This doesn't work most of the time, Flux has a tendency to focus on the subject whenever it's mentioned that there's a subject
it's 'artistic'....
Hey.. so this is really cool but a couple of things.
There is a dev/schnell merged which gives you 4 step+ outputs and guidance. Consider that. Check it out: https://civitai.com/models/628669?modelVersionId=702853
It's good. You can always add more steps but not have less converge.
Caveat 2.. you gave us merged FP16 T5 and FP8 unet.. why? Text models suffer very little form being quantized, but the unet suffers that much more. We may also want to convert to NF4 or GGUF and starting from a quantized model is not a good place to be.
Plus we all have T5 and Clips from the previous flux models.
Hah, I worked with what I had to start with. tools are improving now. All of these concerns I can address now that I have proper conversion tools.
Thank you so much for this checkpoint. It's really the best Flux checkpoint available at this time. Have you given any consideration to converting this to NF4 format? I love the output from this model, but it's just so slow on my potato, er, PC.
NF4 V2 is definitely one of the fastest!
Yeah as much as I like Q4 gguf, it's about 50% slower. I hope we also get NF4 BNB V2 Dev fine tunes 🙏
I can't geenerate a woman with small boobs.
edit
I can't get this to do any skinny women either. Ballerinas are real people too!
It's a mess right now when it's comes to NSFW
working on it. Got the model format stuff all figured out now, so should be able to pour time into developing it out better. new versions coming!
Good, I don't like skinny women with small boobs anyway xD
Nipples are still not good or realistic. This is what happens when the model went so far to censor basic anatomy.
Lora merging/tuning barely start, so it's still soon to say if the heavy boobies censorship will prevent anything accurate, this one overall is okay, a bit hit and miss true, I've seen better anatomy in decent sd 1.5 models, but it's just a start
It is not censored. That's literally how all nekid gals look here in the japz. no jk
Bro, Flux hasn't been out for even 20days yet. Full fine tuning tools aren't even really out . You can wait for boobs. In 10yrs you'll be wishing you could still find "real" nipples on the net.
Bruh, just through nipple fix at 0.25 and you are golden. Can't believe you idiots come on here to complain. SMH.
If you upscale with a good SD15 model then it will fix a lot of those censoring issues and you'll wind up with a much higher quality image. Or you could use a Lora on its own or in combination with upscaling.
The problem with flux is that the nipples are 100% Japanese female nipples (look like a rounded XL cherry being popped from the areola) and the other female private part looks like those very unfortunate Asian "males" that some evil people call clitorimen. Until that obvious problem is resolved, FLUX will continue to lack in the ecchi department.
Flux Topless lora fixes this perfectly. I use this with combination of my own character loras.
@zunaosp3943601 will give it a try. thanks.
@ishadowxx
Are you a racist?
@cafelait What's your problem? Go fix your head.
@cafelait it depends. If the people who removed the inclination to Asian faces that SD 1.5 models had are also racis' then call me in. I am just doing an observation, buddy. Chill.