This is my experimental test checkpoint built using the custom base flux.1dev-abliterated. Unlike my UltraReal Fine-Tune - which has undergone extensive training - this version has been lightly trained as a test, yet still manages to deliver refined NSFW outputs. It focuses on enhancing simpler features (such as breasts) to create a more balanced and realistic interpretation.
Key Features & Notes:
• Lightly trained experimental checkpoint, ideal for exploring subtle nuances and alternative fine-tuning approaches.
• Enhanced rendering for NSFW outputs with a focus on simple yet challenging features, helping achieve natural-looking details.
• Serves as an alternative to the UltraReal Fine-Tune, offering a slightly different style that might suit diverse creative needs.
• Compatible with additional LoRAs and realism amplifiers, allowing you to further tweak and refine the final output according to your workflow.
• This test version paves the way for future updates - if the initial results continue to show promise after further testing, expect newer versions with more in-depth training.
A big thank you to aoxo for creating the original custom base flux.1dev-abliterated. For more details and insights into the base model, please visit:
• https://huggingface.co/aoxo/flux.1dev-abliterated
• https://medium.com/@aloshdenny/uncensoring-flux-1-dev-abliteration-bdeb41c68dff
P.S: if compare with default flux.dev i noticed: better anatomy, better text, in some scenes better hands (but in some even worse). Honestly i decided just to test what base checkpoint could be better then flux.dev
Description
FAQ
Comments (49)
nice if this goes well this could be the turning point in flux
I get this error when trying to use the pruned 11gb version
"UNETLoader
Error while deserializing header: HeaderTooLarge File path: C:\models\unet\UFT.sft The safetensors file is corrupt or invalid. Make sure this is actually a safetensors file and not a ckpt or pt or other filetype."
Hi. For .gguf variants you need to use custom node GGUF loader
@Danrisi Oh I didn't realize the first 2 were gguf, they show up as safetensors, so I thought they're regular diffusion models.
you have to rename the model file, change it from .safetensor to .gguf
@AIsterExploder yeah I realized that a little too late and ended up downloading the fp8 Unet file instead. I somehow never like gguf, so I deleted that and downloaded the unet instead.
I'm glad the abliterated models arrived on CivitAI already :^)
Great work , it's on par with the other checkpoint in terms of quality , with the aidmafluxpro detailer I get great results, but there is a same face issue with it , this lora helps with the faces : https://civitai.com/models/766608?modelVersionId=857446
I tried this model (not your version)
aoxo/flux.1dev-abliteratedv2
and wasn't too impressed. Reminded me of perturbing a good model and getting inferior results. Seems fine-tuning on of these models is an oxymoron, but that's me. Hopefully your model is better. These kinds of models aren't my thang.
Yeah, I checked both versions, v1 and v2, and it seems like something is wrong with v2 (it keeps generating naked women instead of following the prompt). So I decided to stick with v1. Not sure if I'll continue training further since this model doesn’t seem to train well - even with very precise training, limbs end up broken. Honestly, for me, the default flux.dev just needs more training steps, but at least the limbs aren't broken
@Danrisi I appreciate your honest reply. FWIW, I strung (3) realistic style loras together with the base Flux 1D model and am getting very good results. I set the strength to 0.8 for each and may have to dial it back a notch. You can see them over on https://civitai.com/models/832683?modelVersionId=1244911 whenever they get around to posting them. They seem to have a little bit of your vintage style look to them.
V2 is definitely bad, but V1 simply needs some fixing up. It isn't a normal fine-tune: he removed Flux's refusal mechanism via a process coined as "Abliteration". Abliteration causes slight damage to the base model, so what really needs to be done, is some fine-tuning on top of the Abliterated model to correct the errors. This might be possible by using add/subtract nodes in ComfyUI and adding some of the regular Flux-Dev model back into the Abliterated model.
@condzero1950 Can you share the names of these LoRAs, if you don't mind? Is "FLUX [pro] 1.1" one of them?
@mmdd2543 Sure. I've used them with the base Flux 1.D and they seemed to work very well. I also used them with Danrisi UltraReal Fine-Tune with good results. One other thing you might be interested in is the use of the FLAN T5 vs the base Flux model T5. With the FLAN I can get sometimes (I think) better results...But...sometime prompt adherence may suffer. I'm still not sure if the mixture of Loras with the models could also account for some of the prompt adherence issues. I'm using CFG 3.5 for most of the images. Haven't tried bumping this up higher yet.
<lora:aidmaFLUXPro1.1-FLUX-v0.3:0.7>aidmafluxpro1.1
<lora:aidmaHyperrealism-FLUX-v0.3:0.7>aidmaHyperrealism
<lora:FLUX-dev-lora-add_details:0.7>skakker labs
It's a masterpiece. In a sense, there are still some hand errors and some minor issues, but I'm looking forward to it
This model seems great. Looking forward for the updates. Thanks for that!
Someone stole your model in TA again lol
I get this error in forge,
Traceback (most recent call last): File "H:\stable-diffusion-webui-forge\modules\sd_models.py", line 89, in init self.metadata = cache.cached_data_for_file('safetensors-metadata', "checkpoint/" + name, filename, read_metadata) File "H:\stable-diffusion-webui-forge\modules\cache.py", line 114, in cached_data_for_file value = func() File "H:\stable-diffusion-webui-forge\modules\sd_models.py", line 81, in read_metadata metadata = read_metadata_from_safetensors(filename) File "H:\stable-diffusion-webui-forge\modules\sd_models.py", line 268, in read_metadata_from_safetensors assert metadata_len > 2 and json_start in (b'{"', b"{'"), f"{filename} is not a safetensors file"AssertionError: H:\stable-diffusion-webui-forge\models\Stable-diffusion\ultrarealFineTune_v1.safetensors is not a safetensors fileTry putting the models in the UNET folder instead. If it doesn't exist, create it :)
@Lathos as I remember there are no unet folder for forget. All models are used from Stable-diffusion folder. Also I still don't know what's wrong with forge's pipeline that there are so many problems
@Danrisi Ah right, my bad... Wait, I've been doing this thing wrong the whole time?!
I will open an issue in forge and see if someone over there can shed any light on it
If you are using the "Pruned Model fp8 (6.46 GB)" version, change the filetype from "safetensors" to "gguf" and then it should work. This is what OP said to someone else with a similar problem, and when I tried it the problem was solved for me as well. Hope this helps you.
@nunyabizness1Thank you.
This is my favourite creator for Flux. But I hope you can improve the details and accuracy of this model, such as the size and shape of the breasts, hands or fingers, hips, waist, feet, anus, and female genitals so that it resembles or even surpasses Stable Diffusion. In addition, what also needs to be improved is related to the actions taken by the character, the poses taken by the character, non-monotonous faces or ethnicities, and NSFW things.
If it can be like that, then people will flock to use your model to create AI influencers or the like.
Hi, thanks a lot =) BTW, I barely trained any NSFW content with this model. That’s actually a feature of the base model - the author just made Flux 'unforget' how anatomy looks. As for customization, I’m not sure it’s possible without training T5, which isn’t feasible at the moment
Didn't work for me in forge.
The file became corrupted after trying to load it.
I'll keep in tune for any update or workaround for this issue. I'd really like to try the model
If you downloaded version 6.46g, change the file extension to .gguf
@yfjc It absolutely worked, thank you!
I have mixed feelings on this model. When I don't explicitly ask for a nude person it seems to do OK. But if you specifically ask it to give you a nude person you get anatomy issues (like in one of my photos below). I also had a hard time producing good text. Now I did use (3) realistic Loras in combo with this model hoping to alleviate (you know) too much "Obliteration" which may account for some of my difficulties. Again, these type models don't generally appeal to me, but in the interests of distilling Flux 1.D I was willing to give it a try. I'm not sure how well these Abliterated models can be fine-tunes. I'll leave that to the folks doing this sort of thing.
Can you provide gguf format
Might get interesting with this..
https://old.reddit.com/r/unstable_diffusion/comments/1jdhcr8/introducing_t5xxlunchained_a_patched_and_extended/
Eventually.
Hi, yeah i just saw this post and already put in bookmarks =)
but thanx for remind ❤️
Will it work for me?
RTX 4070 12
AMD Ryzen 7 7700X
32.0 GB of RAM
Hi. yeah, 12gb vram should be enough to launch
But i just noticed yesterday that i have broken naming of files. i have all 4 files in safetensor, but 2 of them should be gguf (just need to change extension to gguf in pruned fp16 model and in pruned fp8 model (depends what quant you want to use))
AssertionError: You do not have CLIP state dict!
RuntimeError: Error(s) in loading state_dict for IntegratedAutoencoderKL: size mismatch for encoder.conv_out.weight: copying a param with shape torch.Size([8, 512, 3, 3]) from checkpoint, the shape in current model is torch.Size([32, 512, 3, 3]). size mismatch for encoder.conv_out.bias: copying a param with shape torch.Size([8]) from checkpoint, the shape in current model is torch.Size([32]). size mismatch for decoder.conv_in.weight: copying a param with shape torch.Size([512, 4, 3, 3]) from checkpoint, the shape in current model is torch.Size([512, 16, 3, 3
t5xxl_fp8_e4m3fn.safetensors
vae-ft-mse-840000-ema-pruned.safetensors
clip_l.safetensors
VAE / Text Encoder I used these.
I'm already scared.... I want to run this on POCO F7 via LocalDream or Termux
Hi, There's an updated model from aoxo
https://huggingface.co/aoxo/flux.1dev-abliteratedv2
Hi! Yeah, I know about v2. I ran into the same issue others mentioned - every image, no matter the prompt, ends up with a naked woman. Hopefully v2.1 fixes that. If I can afford it, I’d like to train ultrareal version of it
this being lean to breasts means it will not understand the prompts like the original ultrareal v4 , then v4 is better ? or not ?
Truly perfect, getting much better results for general use, NSFW, and LoRA stacking than the base v1 abliterated model, or non-abliterated based checkpoints
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.