MCNL is meant to be a convinient way to uncensor Qwen with many concepts included. Please keep in mind that loras that are trained on one concept only will generally have better quality than a multi purpose one like this one.
--UPDATE--
V1.0 is out, it adds a lot of new concepts and improves overall quality while being 4 times smaller in size!
Trigger words: nsfw, cum_on_face, blowjob, cowgirlout, creamp1e, penis, l1ck, missionary, nipples, reversecowgirlpov, vagina
Some context for the trigger words, "cowgirlout" is for the cowgirl position viewed from an outside perspective, "l1ck" is for a woman licking a penis, "reversecowgirlpov" is for the reverse cowgirl position from the man's pov. Everything else is self explanatory.
Enjoy gooners
Description
FAQ
Comments (31)
Can you tell us a bit about how you trained it? Everyone is trying to figure out the right parameters, etc...
Muon optimizer with "muon_lr: 0.0001" and "adam_lr: 0.00008" (since Muon only takes care of the hidden weights, while adam takes cares of everything else). Lora rank 128, although this high rank might be the cause of the checkerboard pattern that can be seen every couple of generations with this lora.
Basically I used Muon for it's ability to train multiple concepts at a time effectively, my advice would be to continue using AdamW like with any other model, and do the same settings as Flux when you train your loras. Only thing that you need to change from normal Flux settings is to lower the lora rank, since Qwen is a bigger model, and you simply don't need as big of a rank.
jorkingtoncityshallwe thank you! so you used a custom script? I don't know where I would be able to select two different optimizers at the same time...
rowmoz497 yes I modified ai-toolkit. As I said I recommend you to just use normal AdamW
jorkingtoncityshallwe Could you please share your implementation of Muon, especially if it's already on git? I went to look through the recent fork updates, but didn't find any attempts to implement Muon. But if your implementation is not on git or public, I would still appreciate any tips you could share. Namely: the minimal mods for it not to error on a non-distributed GPU run. Nothing ui-related necessary for me, at least for now. I just want to try to run it via ai-toolkit at all, for a certain project/idea. (Though FYI I do have a fork of ai-toolkits at https://github.com/AlekseyCalvin/ai-toolkit-soonr, with a few mods/extensions on there. But if I added your Muon implementation there, I promise to credit you).
Now, I know that, in theory, adding an optimizer to ai-toolkit could be relatively simple. But Muon, despite all the fanfare around it, has yet to be merged into either Torch or Transformers (though implementations for both already exist as WIP)...
So, I'm genuinely curious: Did you base your ai-toolkit implementation on Keller Jordan's original Muon git? Or some other variant/integration? To that note: there are already some intriguing (and further optimized) optimizers extrapolated from Muon out there (not in relation to ai-toolkit, but in general) – such as the accelerated Flash Muon (https://github.com/nil0x9/flash-muon) and Muon FSDP (https://github.com/samsja/muon_fsdp_2) for distributed set-ups... Anyhow, hope to hear back from you!
A_C_T_soonr hey my implementation is not on github, I've only implemented it locally and it will remain this way because I don't want my github account associated with this civit account LOL, also implementing Muon is not that hard, I'm sure Ostris would do a better job than me if asked nicely :).
As for your single GPU question, there you go, a super simple example of what you can do: https://pastebin.com/LVkNGbTP
And yes my implementation is based on Keller Jordan's original Muon implementation on github.
Edit:
Fixed the issue, it's working as intended.
Struggling to get any decent results on this, even when copying your exact settings. It does seem to introduce an easier time to get some previously censored stuff to generate, but I was unable to generate a vagina or penis, at all.
This has been tested only on comfyui with the fp8 model, so if you're using another tool or other quant, that might be the issue.
works great with Q8 and the latest comfy nightly update.
This is with the fp8 in comfyUI. Strange.
I have this working on Q5/Q4 variants. had to increase the lora weight to 1.5 and cfg to 3.5-4, then it works. also it needs 40 steps or it's likely to be glitchy. I have shift at 1.5, but it's dependent on how you want it to look. it works best with euler_ancestral or uni_pc/uni_pc_bh2 (normal scheduler, though beta might work) for me.
Ye I got it working. I had mixed something up in my workflow! I will edit the comment.
It seems really big for a Lora. Have you tried one at rank 32? Might be more than enough and it might run faster too.
What kind of dataset did you use for it? How did you handle the heads - in order not to influence the other loras (unless it does?)
I am already training a lora at rank 32 actually :D. As for the dataset, it's just a bunch of different datasets for each concept, I just made sure that it is diverse. So as long as we don't have the same person in the dataset many times, it shouldn't influence the faces too much (especially since the captions describe everything in detail, including facial features, which should in theory result to a more flexible lora).
With that said, it still has an influence on other faces most likely, it just shouldn't be that big, and if there are other loras for specific faces, I don't think that it would be a huge issue (especially with the new lora that I am training).
jorkingtoncityshallwe looking forward to your rank 32 then. So far your actuel LoRA doesn't work for me, it "burns" all my generations with artifacts and grainy / blocky pixels, even when reducing the LoRA strength.
sparkleboy599 It was the same for me and I was tearing my hair out what was the problem. Tried the Q4 model, the fp8 model - same problem grainy or blurry images. Then I saw another user that said switching to the nightly version of comfyui fixed the problem and after switching and updating to nightly version through Comfy Manager it fixed those problems - no more images with artifacts, banding or grainy. Try it if you haven't already, but word of advice - don't update the python dependencies from the portable update folder. It bricked my comfy installation and had to revert to earlier backup venv.
spellweaverbg good to know! So is it safe to switch to the nightly version if we don't try the update dependencies etc. ?
sparkleboy599 Yes, and you could always switch back using the Comfyui Manager.
Anyone try is this Lora works with new 4steps Lightx2v for Qwen? https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Lightning-4steps-V1.0.safetensors
I'm curious as well. that Qwen Lightning lora is amazing. Ten times faster than base Qwen model with no quality loss!
Works fine with the 4 steps lora
Hey. Would you mind uploading this to tensor? If you don’t feel like it, I can do it myself. You licence allows that, if I read it right. Just wanted to ask first.
I will upload the new version which will be done today to tensor as well. Just to be clear, the license requires you to credit me IF you do end up re uploading it anywhere.
nsfw content is literally prohibited on tensor.art (https://tensor.art/event/NSFW&CelebrityAdjustments). Uploading it there would be against TOS, not to mention the fact that it doesn't let me upload any of the preview images that I have for the model
Due to pressure from Mastercard etc they very reluctantly changed rules and enforcement. Preview pics must be sfw, and the nsfw channel (113) is hidden. Search is also restricted. Otherwise, nsfw stuff is unaffected, as far as I can tell.
To be clear, their official rule is still "nothing nsfw", but they really don’t care about that. Same as huggingface. Tensor publicly wrote that they want to fully enable nsfw again as soon as they can.
If you do not want to risk your account, I can just upload it and credit you. If you ever want to upload it yourself, you can just write me and I’ll remove it.
aiaiaiai829569 you can go ahead and upload it as a reprint with a link to this civitai page as the original.
Currently uploading it. Is the model size full or pruned? Is it fp16/fp32/fp8 or something else? Sorry if I missed where you wrote that.
aiaiaiai829569 the lora its self if BF16 and I suppose it would be classified as "full" since it was trained on all layers.
Done: https://tensor.art/models/896444506990184056. I hope that including the trigger words is not over the line for public-facing nsfw stuff, but we will see.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.