@EKLL Thank you for testing it! To clarify what I'm experiencing — the issue isn't that sidelock tags have no effect, but rather that the length modifier seems to apply to the overall hair instead of the sidelocks specifically. For example, going from "long sidelocks" to "very long sidelocks" changes the overall length of the hair (i.e., for a girl charactger, the hair on their back; from back-length+ to waist-length+ respectively, which is what "long hair" and "very long hair" does), while the length of the sidelocks themselves remains unchanged almost 100% of the time.

In other words, it behaves as if the tag is being read as a general hair length tag rather than targeting the sidelocks. The sidelock length itself doesn't seem to be a variable the model responds to.

I tested further and can be more specific now. Starting from a character with no sidelocks or very short ones:

• "long sidelocks" — does add some sidelock length, but only to just under shoulder-length. Per the official Danbooru definition, "long sidelocks" should be shoulder-length or longer, so it's falling short of the tag's intended range.

• "very long sidelocks" — no noticeable change from the above. Should produce waist-length or longer sidelocks per the tag definition.

• "absurdly long sidelocks" — also no effect. Should theoretically produce beyond-body-length sidelocks.

(for reference, this is what I am referring to to describe what the tags should be doing: https://danbooru.donmai.us/wiki_pages/very_long_sidelocks; https://danbooru.donmai.us/wiki_pages/very_long_hair)

So it seems like the model responds very partially to "long sidelocks" but doesn't scale further with the stronger tags, and the result is still below what the tag is officially defined as. The upper end of the scale appears to be essentially unresponsive.

For context, I'm using JPN 6.9 as my checkpoint, so it's possible the behaviour differs by base model — your results on your own checkpoint may well be accurate for that model.

@EKLL Thank you for sharing the XY charts — they're actually really helpful for visualising the issue.

In the first chart, "absurdly long sidelocks" is producing roughly waist-length sidelocks, but per the Danbooru definition, that's what "very long sidelocks" should be doing; "absurdly long" is defined as beyond body-length. Meanwhile, "very long" itself only produced marginally longer sidelocks and falls well short of waist-length here. And in the second chart, neither "very long" nor "absurdly long" produces any noticeable difference from "long sidelocks" at all regarding the length of the sidelocks.

Across both charts, the general hair length appears to scale with the tags rather than the sidelocks being targeted in isolation. It's possible that because the characters demonstrated have their hair tied back, the effect only becomes visible at the "absurdly long" tier — but from my testing on JPN 6.9, it does consistently affect general hair length as well, as I mentioned previously.

So even in the best case (first chart), the tags are one tier behind where they should be. In the other case, the upper two tiers have no effect whatsoever, which is what I am experiencing in my testing. The behaviour is also clearly inconsistent across and even within base models.

This is what got me thinking about a LoRA that specifically targets sidelock length — whether through trigger words for each tier or a slider — without bleeding into overall hair length or affecting any base models' art style. I think there's a real gap here worth filling for the community. But of course, no pressure at all. Just wanted to share the observations!

@EKLL 我也发现有这个问题，前几天还正常的，现在文件下一半之后就会非常慢最后下载失败，换账号换网络换设备都会这样，我甚至是在墙外不需要翻墙直接能够访问c站的。怀疑是c站的问题😫😫