Feature: 列出所有可用的Textual Inversion, Hypernetwork, Lora模型

[yi03]

Scope

Stable Diffusion WebUI (AUTOMATIC1111)

Describe the problem related to the feature request

Textual Inversion, Hypernetwork, Lora这三种小模型在最新版的sd-webui里都可以在prompt里来选择模型，比如在prompt里输入<lora:amberGenshinImpact_flexibleV1:1>就表示选择了一个名叫amberGenshinImpact_flexibleV1的Lora模型，权重是1。所以需要一个列出当前所有可用的Textual Inversion, Hypernetwork, Lora模型的功能。
这三种模型里我个人的话是Lora模型用的最多，有时会用Textual Inversion模型，几乎不用Hypernetwork模型。
其中选择Textual Inversion模型在prompt里不需要加<和>，选择Hypernetwork和Lora模型需要在prompt里加<和>，而目前的bot会将<和>转成&lt和&gt，所以在选择Hypernetwork和Lora模型之前需要先解决这个bug。相关讨论：#161

Describe the solution you'd like

向bot发送指令，bot列出所有可用的Textual Inversion, Hypernetwork, Lora模型。

Describe alternatives you've considered

No response

Additional context

No response

[MaikoTan]

Also note that the <lora:xxx> or whatever part SHOULD NOT be fed to the translator. So there're two ways to achieve this:

1. maintain the a1111 way, we extract all <lora:xxx> out before the prompt being translating, and then append these extra net at the prompt end.

concern: is the order of extra nets affect the result? And how do we deal with embeddings which has not a special syntax?

2. add new option to the command, like --extra lora:xxx,hypernet:yyy, and then we novelai-bot would just translate the prompt as always before adding extra nets for the user.

3. Hybrid??

---

Feel free to comment out any thought you have! 🎉

[yi03]

Also note that the <lora:xxx> or whatever part SHOULD NOT be fed to the translator. So there're two ways to achieve this:

1. maintain the a1111 way, we extract all <lora:xxx> out before the prompt being translating, and then append these extra net at the prompt end.

concern: is the order of extra nets affect the result? And how do we deal with embeddings which has not a special syntax?

2. add new option to the command, like --extra lora:xxx,hypernet:yyy, and then we novelai-bot would just translate the prompt as always before adding extra nets for the user.
3. Hybrid??

Feel free to comment out any thought you have! 🎉

刚试了试，lora的顺序对结果有一点影响，不过大多时候影响不大。

考虑翻译的话，这两种我觉得都可以，看你们怎么方便吧。
不过如果方便的话，即使不对<lora:xxx>格式的翻译做特殊处理，也希望能不转义<和>，因为这样的话，我这种平时不开翻译功能的人就可以直接复制sd-webui格式的prompt，不用再改写成--extra lora:xxx,hypernet:yyy的格式了。

[yi03]

Textual Inversion的话，可能也可以设置成形如--extra ti:zzz的格式来避免翻译吧。

[MaikoTan]

刚试了试，lora的顺序对结果有一点影响，不过大多时候影响不大。

Oh, I mean when you have a very long prompt, and there are extra networks everywhere, compares to a long prompt, but extra networks are all at the end. Would they behave different?

Par exemple : (let's say genshin-impact-amber here is an embedding)

A girl sitting under the tree, <lora:aaa>, on the grass, <lora:bbb>, <hypernet:ccc>, she has brown hair and closing eyes, smiling, genshin-impact-amber

compare to

A girl sitting under the tree, on the grass, she has brown hair and closing eyes, smiling, genshin-impact-amber, <lora:aaa>, <lora:bbb>, <hypernet:ccc>

Note: we would still maintain the relative position between extra networks, but put them at the end.

考虑翻译的话，这两种我觉得都可以，看你们怎么方便吧。 不过如果方便的话，即使不对<lora:xxx>格式的翻译做特殊处理，也希望能不转义<和>，因为这样的话，我这种平时不开翻译功能的人就可以直接复制sd-webui格式的prompt，不用再改写成--extra lora:xxx,hypernet:yyy的格式了。

Yep, so personally I would like the hybrid mode, which would support these two ways both.

[MaikoTan]

Also, webui call Textual Inversion as Embedding in some place, are they referring the same thing? Also Hypernetworks vs Hypernet.
Because I want to define a common / general but also short names for them.

[yi03]

Textual Inversion, Aesthetic Gradients, Dream Artist 都是embedding（虽然后两个用的人比较少），用法也一样，所以确实--extra embedding:zzz这样可能更好，嫌太长的话可以缩写成--extra emb:zzz之类的。
Hypernetworks 和 Hypernet 是同一个东西。

[MaikoTan]

So this feature COULD be implementable with these APIs, but a weird thing is that the /sdapi/v1/embeddings API seems to return Textual Inversion only?

Also there is an API that returns all thumbnails for these extra networks.

[yi03]

So this feature COULD be implementable with these APIs, but a weird thing is that the /sdapi/v1/embeddings API seems to return Textual Inversion only?

Also there is an API that returns all thumbnails for these extra networks.

这三种embedding的用法是一模一样的，只是训练的时候不一样。这三种embedding训练出来的pt文件也都放在同一个目录下，所以返回了Textual Inversion肯定也会返回另外两种，可能是文档没更新之类的吧。
缩略图我猜是这个东西

[MaikoTan]

Okay, it looks like I put them different places so I have only Textual Inversion returned. Thanks for pointing out!

And I would not implement the thumbnails feature at least for now, maybe in the future, I don't know.

The last thing is, would you mind testing whether the order differences affects the result much? I mentioned it at #165 (comment)

[yi03]

Okay, it looks like I put them different places so I have only Textual Inversion returned. Thanks for pointing out!

And I would not implement the thumbnails feature at least for now, maybe in the future, I don't know.

The last thing is, would you mind testing whether the order differences affects the result much? I mentioned it at #165 (comment)

我现在不太方便测试，大概今晚七点我会把测试结果发在这里。

[yi03]

Okay, it looks like I put them different places so I have only Textual Inversion returned. Thanks for pointing out!

And I would not implement the thumbnails feature at least for now, maybe in the future, I don't know.

The last thing is, would you mind testing whether the order differences affects the result much? I mentioned it at #165 (comment)

试了试，似乎有一点区别，不过区别不大。

[MaikoTan]

Okay, thanks for the testing.

I'll go with this way, and implement it as soon as possible.

[KotoriKoi]

关于这个问题我想我可以给出一些建议，由于我不熟悉node.js和koishi，我在之前改写了另一个基于python和nonebot2的novelai-bot项目实现了这个issues的大部分功能要求。

首先关于ti(Textual Inversion)，它分为正面ti 和负面ti，我使用了--pt和--ft，并且提供了简写-p和-f来调用它，之后只需要将它们分别添加到正面tag和负面tag即可，像这样：

由于ti的位置会影响图片的效果，而一般我用ti来定义某个角色，所以我把它放到了用户输入的tag前面，另外请注意ti不需要经过翻译。
如何列出ti？我试过了sdapi中的GET方法，但是它有弊端，当添加新的ti后需要刷新webui才能获取到新添加的ti（也可能是models，我不知道ti是不是也这样，另外目前sdapi中也没有lora的GET方法）。所以我传入了一个本地路径来获取它们，像这样：

这是功能展示：

然后是lora，目前来说我还没有看到负面lora（但实际上它可以放到反向提示词，并且你正反都放入同一个lora权重都是1它们会抵消），另外它的位置不会影响效果，效果取决于权重，但我喜欢把它放到最前面。我仍然使用传入路径的方法获取它们，然后定义一些命令来调用它，像这样：

其中列出lora的方法和ti的方法一样。
最后是Hypernetwork，我不太使用它，但它和以上用的的方法差不多。
然后是关于该issues没有提到的一个功能，像这样：

实现它只需要在post中加入styles即可，具体可以参考sdapi

希望这可以给你帮助。

[MaikoTan]

Thanks for the sharing and explanation!
I'll take a look and try to implement this as soon as possible.

[yi03]

Okay, thanks for the testing.

I'll go with this way, and implement it as soon as possible.

今天又想到了个feature，就是不知道放不方便加个备注功能。因为很多lora的名字都很长，lora模型的名字也经常是某个角色的英文名，看不懂是啥意思，就想加个备注。
比如说列的时候备注名和原名都列出来，调用的时候可以直接加--extra lora:备注名来调用。

[yi03]

Okay, thanks for the testing.
I'll go with this way, and implement it as soon as possible.

今天又想到了个feature，就是不知道放不方便加个备注功能。因为很多lora的名字都很长，lora模型的名字也经常是某个角色的英文名，看不懂是啥意思，就想加个备注。 比如说列的时候备注名和原名都列出来，调用的时候可以直接加--extra lora:备注名来调用。

我啥比了当我没说，我忘记可以本地改lora文件名了。