V5 Update

Changes

Hi guys. This version is an improvement in three ways. One is that I slightly improved the dataset. The next two have to do with training. This time, I upped the LoRa size to rank 32. This might make the LoRa slightly less flexible, but I think in terms of quality it was worth it. Next, I learned how to simulate larger batch sizes by using gradient accumulation steps. This allowed the training to better approximate my entire dataset instead of getting fixated on the details of one or two images.

Instructions

All images were made with the sgm uniform scheduler and 3.5 guidance. Civitai rounds 3.5 to 4 so if you see that, it's wrong. As with previous versions, this LoRa modifies the clip so makes sure you're incorporating that into your worklfow.

Please let me know what you think and if you have any suggestions. Buzz me if you want to support my work. Lastly, enjoy the LoRa! I believe this is my finest work yet haha

V4.5 Update

Changes

Hi guys. I was trying to get my lora to work with Adel_Ai's Anime Artwork and was having trouble. That's when I realized the clip strength for my lora was too high and causing issues. After reducing the clip strength by half, I was getting some good results.

In the showcase there are some side-by-side comparisons of images generated with and without his lora. If you use his lora with mine, my recommendation is to use the settings in the image below, which is how I generated the showcase images.

Overall, by reducing the clip strength, this should be a slight improvement on V4 with or without the Anime Artwork lora.

My goal is to reach a level of quality so that people won't ask, "who generated this?" or even "who drew this?" but rather "what anime is this from?"

V4 Marauder

Changes

I've made improvements to the training method as well as the dataset. This version is slightly more realistic as opposed to stylized. So stick with the previous version if you prefer that.

I've found the sampler dpm++ 2s a to be incredible when paired with the sgm uniform scheduler. Also increasing guidance helps a lot. Most of the showcase images were generated with 4.0 guidance. Guidance improves quality significantly, but sometimes it causes Flux's censorship to come through. You will usually see this as white nipples. Dropping down to 3.5 guidance usually solves it. A few of the showcase images were generated with 3.5 guidance, but civitai rounds 3.5 up so they all show as 4.

V3 Town Car

Changes

This is an incremental update. It is an improvement on my previous training workflow. The lora should now work at 1.0 strength for both clip and model, though you may have to play around with the strengths for certain prompts/lora combinations.

For this version, I have found DDIM to be the best sampler, although occasionally Euler A outperforms it for certain prompts/lora combos/resolutions. I still haven't determined when to use which sampler, but DDIM works a majority of the time. In general, for schedulers I would use DDIM uniform with DDIM and Beta with Euler A.

As before, use 3.0 guidance since that's what this lora was trained on.

If other settings work better for you, please let me know. I'm still experimenting with it.

I am fairly happy with how this turned out so I'll probably be taking a break to focus on character loras for a while. If you have any questions or want to carry on my work towards cracking Flux nudity, feel free to DM me.

V2 Grand Marquis

Instructions

This lora works best on the full flux dev model since it was trained in bf16. I'm not sure what will happen if you try to use it with the fp8 version of flux, but you are welcome to try.

Some tips I have for running flux dev locally in comfy are to enable "--lowvram" and to use "--reserve-vram 3.0" The number 3.0 is how much vram to reserve. You can replace it with whatever number will keep you from running out of vram. With these settings, I was able to generate each of the above images locally in about 1-2min. Less than 2min for a high quality image is what I consider acceptable.

In training the lora, I also trained the clip. So please apply this lora to the clip as well. I posted an image showing what I mean (the Drake meme).

For some reason, (I'm not sure why) the lora works best with the sampler Euler A. Euler A, 55 steps, and the beta scheduler consistently yielded good results. Also use Guidance 3.0 since that's what the lora was trained on.

For model strength, I had good luck between 1.2 and 1.4. Clip strength was not as clear but anywhere from 1.0 to 1.4 seemed to work pretty well. Please experiment and let me know what works for you. Remember that this is an art, not a science.

Intent

Since Flux's release, my goal has been to create a nudity lora that can be used in conjunction with character or style loras with a minimal amount of conflict/disruption. I believe this lora represents the pinnacle of that effort.

Method

I used SLDR's photorealistic nudes lora (Version 2) as a starting point. Please support his work if you haven't already. Then I trained it another 24k steps on 80 high quality images depicting anime nudity. The training was done in full bf16 by using block swapping which is what distinguishes it from V1. After that, I shrunk the lora down to a Rank 8 lora for flexibility.

Conclusion

At times, uncensoring Flux feels like a test of wills between us and Blackforest Labs. The censorship works in ways I still don't fully understand. However, Flux may be the last decent image model we get that will be runnable on consumer hardware. And in terms of potential, I believe we've barely even scratched the surface.

To me, this lora represents the farthest anyone's gone towards uncensoring Flux while incurring the least amount of disruption to the model. But I challenge you to go further. Use my lora as a stepping stone if you need to (the same way I used SLDR's). Reach out if you have any questions. We can do this. We are so close.

V1 Crown Victoria

Instructions

Because of the way this lora was trained, it works best with the fp8 version of Flux and weight d type e4m3fn.

All images were generated at either resolution 1200x1600 or 1536x1536. I also use separate prompts for T5x and Clip L.

The parameters were as follows.

Max Shift - 1.15

Base Shift - 0.5

Sampler - deis

Scheduler - Beta

Guidance - 3.5

Lora Strength - 0.6 to 0.8

Intent

My goal since making the Photorealistic NSFW Merge (PNM) has been to replicate its flexibility and versatility, but with anime. This lora can either be combined with style loras or character loras to act as a sort of nudifier.

The goal of this lora was NOT to create high quality "anime themed" artwork with gradient shading and dynamic lighting. The goal was to create reasonably believable nude anime screenshots.

Method

Instead of starting from scratch, I used PNM as a starting point. I did this by merging it into the Flux Dev model, essentially creating a finetune that already understands nudity. Now I needed to teach it anime nudity. I trained this finetune using 80 high quality images depicting anime nudity and hand written prompts. I trained for about 18k steps.

After training, I had a lora designed for that finetune of Flux, but not Flux itself. To fix this, I baked my lora into the finetune so that it now had both loras in it. Then I used Kohya's Extract Lora function to compare the finetune to base Flux and extract a lora that functionally contained two loras. The end result worked with base Flux. Because this is a lora trained on top of another lora, the strength needs to be reduced. I have found somewhere in the 0.6-0.8 range to be best.

Conclusion

This lora has accomplished its stated goals. From making PNM, I learned that the key to a flexible lora is reducing the lora rank to be as low as possible without significant quality loss. I have found that concept loras, such as this one, can be reduced to rank 16 before quality loss becomes an issue. Character loras can generally go down to rank 8.

In some of the example images, I used my own character lora (Anko Mitarashi) as proof of this concept. That lora is rank 8, and there seems to be no conflict between it and the nudity lora - validating my conclusion.

I believe that future NSFW loras should use this method of training on top of other loras. Training the censorship out of Flux is difficult and time consuming. We should not be reinventing the wheel each and every time.