Thank you for your feedback and thoughts! I always welcome positive and constructive comments like yours.

That said, I’m sorry but I don’t have plans to create models with combinations other than the “ALL” version.
The reason is that having too many combinations would increase the release workload and make future updates and management more difficult.

I’ll continue working on improving detection accuracy, so feel free to let me know if you have any other questions.

@Wizzi Thank you for your reply.
The next update is planned to improve the detection rates for cross-section and x-ray, but I will also try to increase dataset for the compositions you suggested.

@Wizzi Glad to hear it's working better for you!

> Only thing I want to ask is if it is intentional for nipple detection to detect see-through or not.

Yes, the detection is intentional. The low detection rate is due to the limited number of see-through samples in the dataset.

Also, improving the detection rate for see-through nipples was planned for a future update.
If you don’t mind, could you share why you prefer not to detect them?

@Wizzi I see, thank you for your reply.
It's interesting to hear that you're using a dedicated LoRA for specific body parts—that's very helpful to know.

Currently, my approach is focused on improving the overall detection rate, including see-through cases. So I hope you can understand that some extra adjustments might be necessary for your specific use case.

From a technical standpoint, it's always possible to exclude see-through detection, so depending on feedback from other users, I may reconsider the direction in the future. If that happens, I’ll be sure to let you know.

@Wizzi Hi! Thanks for the quick review—your feedback is always helpful.

> but in poses woman is kneeling when looked at from above or when she is for instance on all fours with breasts hanging, recognition is pretty inconsistent.

I didn’t include many images with that kind of composition in the dataset.
To confirm, are you referring to a composition like the one you posted on Civitai about 10 days ago?
I downloaded the image and checked it, and you're right—the model didn’t detect it properly.

By the way, I also tested with an image using the following prompt, and it was able to detect nipples as long as they weren't heavily obscured by arms or other elements.
The model I used was WAI-NSFW-Illu v14:
1girl, from above, all fours, nude, large breasts, sagging breasts, nipples

@Wizzi I see, thank you for the detailed information.

Adding around 10 to 20 images with similar compositions and actions to the dataset should improve detection rates.
Since the images I generate myself tend to have some bias, your feedback on detection performance is extremely helpful.

I'm planning to increase the variation in the dataset for the next release, so if you notice anything else, feel free to leave a comment.

@Wizzi Thanks a lot for the enthusiastic feedback!

Curved penises, partially covered areolas, and bouncing breasts in motion are either not included in the dataset or only appear in very small numbers 😓

I'll try generating some sample images myself when I have time, so if you're willing, it would be really helpful if you could share the specific prompts and (publicly available) models you used.

> Posting it as a picture, will delete once you see to not clog up example pics.
I checked both the model gallery and your profile, but I couldn’t find any images that match the description. Is it unpublished or maybe still in draft?
If possible, please send me a direct link.

Feel free to leave comments anytime if you notice anything else. I keep notes of all feedback and make sure to test them properly when I have the time.

@Wizzi Hmm, it looks like I still can't see the image.
I double-checked my visibility filters, but everything is set to show all content.

In any case, thank you for the tag information. Most of the tags you mentioned weren't really included in the dataset.
Up until now, the datasets I’ve added were mostly focused on foreplay and intercourse, so I think I’ll need to add more compositions that focus specifically on breasts or genitals going forward.

If you’re open to it, I’d love to show you some sample images when I start preparing the next version—just to check if they resemble the ones that weren’t detected properly.
That said, I haven’t decided exactly when I’ll start the next round of work.

Lastly, thank you again for taking the time to provide such detailed feedback!

@MeloChan Glad to hear you're enjoying the models!

> these models (V2) working great but missing some close or small area
Could you let me know which specific areas you're referring to? If you don’t mind sharing an example image or the model and prompt you used, that might help identify the cause.

> Also i was wondering how these models can be trained? Any tutorial ref link?
I've been asked about that by a couple of people before, so I’ll drop the links to those comment threads here:
- English: https://civitai.com/models/1313556?dialog=commentThread&commentId=737373&highlight=737762
- Japanese: https://civitai.com/models/1313556?dialog=commentThread&commentId=768265&highlight=798196

For both CVAT and Ultralytics, I referred to the official manuals and Github issues.
Feel free to ask if you have questions—I can try to answer based on what I know.

@MeloChan Thanks for the detailed info. I’ll try to include the poses you mentioned when preparing future datasets.

Just to double-check, I’ve written out the poses and actions as Danbooru-style tags. Let me know if any of them seem off:
hug, upright_straddle, girl_on_top, kiss, sex, vaginal, group_sex, piledriver_(sex), upside-down

To put it simply, it's definitely possible, but it would take quite a bit of effort.
Specifically, it would require annotating the breast areas (using polygons) in all 900 images, and the same would need to be done for any future additions to the dataset.

That said, I'm currently prioritizing improving the detection accuracy of this model, so I can't say when or if I'll get around to adding that.

If detection is low, try lowering the threshold as much as possible.
While lowering it too much can increase false positives, in my local tests it seemed to work fine even at 0.01.

If you're still not getting detections at 0.01, it might be due to an style or pose that wasn’t included in the training dataset.
If possible, feel free to share the image, or the model and prompt you used—it might help identify the cause more accurately.