Female underwear that resembles male boxer briefs. I'm not sure this trained properly: Epoch 6 was totally missing pictures, and the style seems way too into real/3D. No idea what happened.
Trigger: G_B_B_P, [color] panties,
Trained with: crotch seam, [color] waistband, print waistband, clothes writing, crotch focus, ass focus, panty focus, short shorts
Not trained with but helps: boxer briefs
Negative: bulge (may be needed if you used boxer briefs to enhance shape), lowleg panties, bow panties (generally needed when doing panties in Illustrious or Pony)
Anima version has tag clean up and trigger change.
Description
Comments (12)
About the style, it seems to overtrain quickly with anima (especially with 3d styles and no te training). Did you use a lot of photorealistic images in the data? This lora was trained with photorealistic images/renders only and has kinda the same issue (even if it's not that bad I think).
Yes, but the others in the series also did and they don't appear to have this issue. In-fact, this is the heaviest on illustrations in the series.
@NanashiAnon That is weird. Unfortunately I do not have experience with training anima LoRAs with realistic images. Only the above mentioned and the Blues Brothers outfit LoRA had a few images in the dataset, but most were anime/illustration. What was the ratio of illustrations and realistic images for this LoRA? Like 1:1? I guess you tagged the realistic images with "realistic" and/or "photorealistic"?
I tried the briefs anima lora and that one does not seems to have the same issue with style. Though there seems to be another issue with both of these LoRAs, especially the briefs lora, that is if I do not prompt a known character it will give head out of frame shots, even when using "black eyes, looking at viewer, cowboy shot" and at lower weight (down to 0.6). Even when I prompt a known character I need to lower the weight to 0.7 and specify hair color and eye color to get their face and not a panty focus image.
My conclusion: Given you tagged everything correctly, I think the LoRA is storing to much unwanted weights. The network dim is probably too high, so that everything recurring gets absorbed. I'd recommend to lower the network dim, 8 or even 4 may be sufficient (I used 8 for the clothing LoRAs I did with anima).
@RisingV Interesting. Did you try negative "lower body, head out of frame" for the brief one? All picture but one was zoomed in for that one with one of those tags, but Illustrious almost never ran into that issue and rarely needed them added.
Dim is something I'd have to check for Anima, but I set the final one in the main series to train last night (should be up in an hour) and I don't really want to be blindly experimenting with such a high cost training. Might retrain Mission Vao with a couple extra images this weekened just to test, but I don't think I'll be dumping into retrains if that doesn't make better results.
@NanashiAnon Yes just tried: https://civitai.red/posts/29823291
It helps, but even then I need to lower the weight to 0.8.
Anima is just learning much quicker than illustrious at the same learning rate/settings (at least in most cases). I tried a bit around for a character lora using 3d images with different settings and the character and style train in very fast even at network dim 8 (like 20-30 epochs (400-600 steps) with 20 images dataset). This is regardless of the settings I use, but especially when not training the text encoder.
@NanashiAnon by the way, the lora has a weird name, when I add it to Forge Neo prompt "training_1913851-20260712035038856".
@RisingV Do other trained on site Anima models do that? They don't have a real default name and never did (I reported it, but nothing was ever done) so I had to rename it after training.
@NanashiAnon I am not sure. I downloaded only a few other anima loras and most were probably not made onsite. I'd have to find a creator training onsite. Let me check..
@NanashiAnon So this lora is also giving me a weird prompt name "training_4326959-20260610221610686". This other lora is probably also trained onsite, but is two months old and does not give me a weird name.
@RisingV If the dim should be 8 or 4, do you have any recommendations for the alpha in this case?
@NanashiAnon I use dim=alpha all the time. Never used anything other with illustrious or anima. Alpha=0.5*dim might work too, but I never tried. Those are usually the recommended values. Dim=alpha turns off the effect of the alpha value. I recently wrote an article about anima training settings, trying to explain all the settings I am using and I think the best explanation I found for the alpha value is this one: https://github.com/bmaltais/kohya_ss/wiki/LoRA-training-parameters#network-alpha
@RisingV Thanks.