The dataset for all concepts combined consisted of around 600 images

doggystyle was the one concept I forgot to test and you're kind of right. I just tested it, and it is very hard to get a doggytyle image (the few times that it generated it successfully it didn't look very good either) and it seems to be getting confused with the reversecowgirlpov concept (since it keeps making the man lay down instead of stand up or stay on his knees). The issue seems to be that the 2 concepts are kind of similar visually, and also the fact that I don't have many images for the doggystyle concept.

I will just remove it from the list of concepts/triggers, in the future I might make a v2 with more balanced data that would hopefully fix this issue.

"still needs time to cook more" the lora has 10 concepts... if I make a lora for only one concept (e.g. vagina), it will render them well almost always. The purpose of this lora wasn't maximum quality, the focus was quantity, it was a fun experiment for me, and a fairly convinient way to do a lot of different nsfw stuff decently. Nipples looks fine because they're fairly simple, while vaginas have lots of detail, folds etc, which makes them hard to incorporate in a lora with so many concepts.

The truth is though, qwen requires pretty heavy hardware to train, and each lora is much more expensive to make than other models I've trained. Which is why I decided to do one lora with many concepts, since making many loras for each one would've cost me a lot more.

Mind you I do this completely for free, I don't ask for donations, I don't have a patreon or any donation page setup. If you want super good vagina rendering, or super good cum rendering or whatever else this is simply not the lora for you, since the focus is on generalization, not on one concept expertize.

brnlittokhoes3110 Ostris just released a new update that enables you to train on a 24GB card, but the quality is quite a bit worse because his method leverages 3bit training with a specialized lora to "restore" the quality. The lora that restores the quality isn't magic, and ofc it isn't perfect. So loras made with this method end up having more artifacts overall and it probably won't be able to learn small details. Here's his video on it https://www.youtube.com/watch?v=MUint0drzPk

fox23vang226 yes if you're training only one concept it can work, did you even read my previous reply? My lora contains 10 concepts and it has been trained for almost 31k steps, training it for longer wouldn't help much, in my case it's not the amount of steps that is the issue

fox23vang226 v0.1 had a higher network dim/alpha (which is why it was also significantly bigger in size), and it didn't perform much better, if anything v1.0 which is about 4 times smaller seems to be better.

Higher learning rates would just start cooking either all concepts or specific concepts only (depending on how high/low it was set), the learning rate that v1.0 was trained on was close to the max possible without cooking any of the concepts, but perhaps I could've done a bit higher to save on training time.

Honestly I think if I make a v2.0 I would try to train it without quant (kinda pricy because I would probably need 80gb vram, perhaps more, not sure), I would probably try a dim/alpha somewhere between v1.0 and v0.1, and I would try to improve the dataset to be a bit more balanced.

Anyway good luck with your lora, I hope it turns out well

this lora has not been trained on any dildo related content.

Currently qwen image edit is not supported in any training tool to my knowledge. Also qwen in general costs too much to train properly, I could consider it once support for the model is available, but I don't promise anything.

@Skuuurt I know, it wasn't supported when I first wrote my comment :D

the replies in this comment section might help since they were experiencing the same issue and managed to solve it https://civitai.com/models/1851673/mcnl-multi-concept-nsfw-lora-qwen-image?modelVersionId=2105899&dialog=commentThread&commentId=898889

I have been able to use this lora along side with my style loras without an issue. It all depends on the loras you're using, just play around with the weights until you get a good result. Also if you're using a heavily quantized version of the model and the 4-8 step loras you're more likely to have issues. FP8 with 20ish steps will be consistently better.

First of all you can train the text encoder on ai-toolkit, you just need to enable the option. Also yes this lora has trained both unet and text encoder.

As for this question "This means, youre not touching one time the clip.
Means -> You cant train new trigger words.". This is completely false, even if you train the unet only, your trigger word will still work, training the text encoder as well ofc makes the model understand what it's doing better. (this applies for one concept training only btw, for mutliple concepts I believe training the text encoder as well is pretty crucial, but I could be wrong). I won't go into detail why trigger words work without training the text encoder because this reply will become too long, but literally just train a lora yourself on a single concept with a trigger word and then use the lora with and without the trigger word, there will be a clear difference (as long as you didn't overtrain it).

@LDWorksDV You know what you might be right, I will do some testing, if ai-toolkit's text encoder training doesn't work then a custom script would probably be best, and shouldn't be that hard to do (although having a pretty GUI is always nice haha). As for my sdxl based stuff, I used onetrainer, which I personally like quite a bit and seems to at the very least have working text encoder training :D.

Honestly though I don't think that I am gonna bother with qwen again cuz that thing takes insane hardware to train.

@LDWorksDV yeah I just tested it with the same prompt/seed and different clip strenght, the resulting image was exactly the same. So you're right ai-toolkit's text encoder training doesn't work properly. Honestly kind of impressive that this lora kind of works considering that the text encoder is essentailly clueless about all the nsfw shit lol.

I have not used "Differential Output Preservation" btw, but in theory that should be just a way to preserve existing concepts when introducing new ones (so basically prevents overfitting and "forgetting" of concepts).

@LDWorksDV so just a little update if you're curious, I managed to train the text encoder on the same dataset as the lora, the results were the following:
- in some cases completely mangled/bad looking results were completely fixed with the nsfw text encoder (fairly rare)
- in some other cases it would make the output a bit worse (fairly rare as well)
- the overall nsfw "understanding" and prompt adhesion seemed to be slightly improved

I do want to note that I tested the text encoder on it's own for captioning my dataset and it was fairly accurate so the training its self was successful. BUT I do think that all the effort was not worth it, since the improvements are overall very minor, only in some very rare cases you get a big improvement. I might try training the text encoder on a bigger dataset at some point in the future (since this lora's dataset is not big enough for fully finetuning a text encoder I think :D), but for now I am not gonna bother with qwen and will continue working on my YARI model.

@LDWorksDV you can just use LLaMA-Factory to train the text encoder, just make sure to select the correct model, and make a dataset that's compatible with the required format (everything is explained in the readme of the github repo)

sure, train your own deepthroat lora lmao

@clevnumb A diffusion model can only generate what it has already been trained on (and also kind of mix and match the concepts it has learned). If you want it to reliably generate some specific concept like a deepthroat then you have to train it to do that.

there is no need to send me the crash report as there is nothing for me to do, it's up to the developers of draw things to support qwen loras, in comfyui it works just fine.

Oh cry me a river dude, make a better one yourself or use something better then. I simply don't have good enough hardware to keep working on Qwen and smartasses like you make it much more unlikely that I will ever touch this model again

@deditz111802 first time I see anyone censor "wtf" to "wt*" lmao, look dude if you wanted to actually give constructive criticism you wouldn't have worded it in such a way, "most of my outputs come out pixelated" is constructive criticism, "Output quality is ....." just comes off as passive agressive bs

@amazingbeauty hey man at least I contribute to the community, this is def not my best work and I admit it.

Perhaps you can start talking shit once you make something yourself, cuz currently you have made 0 models and you've made 1 post in which you sound like a moron, while begging people to make a specific lora for you lmao.

@deditz111802 "a normal feedback comment" you delusional passive agressive rat, "Output quality is ....." is as passive agressive as it gets, you legit have no clue what constructive criticism is.
Also idk why you're so bothered by how I answered to him, it's not like I was particularly nice to him either, I literally told him that he's a moron.
Anyway feedback noted I should start being more mean to people so I can make you happy

The metadata and the workflow is attached to all the preview photos so you can see what I have used, generally 2.5 CFG, 20+ steps if you have patience, Euler/Simple work fine and 3.1 shift.

Honestly I haven't bothered with qwen and this lora in general for a long time, there are some pictures generated by other people in the gallery with penises so try to copy their prompts I guess. Or try combining this with a penis specific lora.

this was trained on the first version of qwen, lightning loras/versions can have a bad impact also if you use some super low quant that could have an impact too. All gens that you see in the preview images are with fp8 qwen (the first version released).
Even when it works though it's not that great as you can see from the preview images, if all you care about is nudity I'd recommend SDXL based models.