Type | |
Stats | 4,311 |
Reviews | (1,006) |
Published | Jul 15, 2023 |
Base Model | |
Training | Steps: 18,000 Epochs: 30 |
Usage Tips | Clip Skip: 1 |
Trigger Words | wireframe |
Hash | AutoV2 F77F288839 |
If you find our articles informative, please follow me to receive updates. It would be even better if you could also follow our ko-fi, where there are many more articles and tutorials that I believe would be very beneficial for you!
如果你觉得我们的文章有料,请关注我获得更新通知,
如果能同时关注我们的 ko-fi 就更好了,
那里有多得多的文章和教程! 相信能使您获益良多.
For collaboration and article reprint inquiries, please send an email to [email protected]
合作和文章转载 请发送邮件至 [email protected]
By: Ash0080
English
About V4
Actually, when I was training this model, I named it V3Fix because its training dataset is exactly the same as V3, but I fixed some errors. However, it was trained using a completely different strategy, and now it is a style-agnostic LoRa model with very minimal impact on the style of the base model.
Therefore, it can now be applied to almost all types of base models, including 2D, 2.5D, 3D, and Photo. Moreover, you no longer need two models, whether it is for txt2Img or img2img. One model is enough. because its usage has undergone such a significant change. To avoid confusion, I renamed it V4.
For V4, setting the weight to 1 is sufficient. Reducing this weight can to some extent weaken the strength of the lines (but may also cause loss of lines), and I recommend between 1 and 0.7.
For other features, such as HiRes, face correction, and using tips, please refer to the V3 instructions for now. By the way, the Hires Denoising Strength is no longer limited to 0.33, so you can use larger values more normally.
Img2Img has been greatly improved because it is much more tolerant of image style than V3. Now, it is much better at applying topology to images.
What?
This is an experimental model that serves as a tool for adding topology lines to models. Its ultimate goal is to bring everything into a 3D world. ( although that's still quite far away )
Perhaps in the future, it could be combined with NERF to create something amazing, but I don't know for sure!
Why?
Having worked in the gaming industry for many years, although I am no longer a practitioner in the art-related fields, I still believe that AI has the potential to offer much more in game development. Of course, this does not mean that AI can replace excellent artists, but I believe it can help them reduce some of the repetitive work in their daily tasks. This series of GAME DEV TOOLS is based on this idea.
Topology is a crucial skill that 3D modelers need to master, and professional practitioners even use the quality of topology to judge whether a modeler is a novice or an expert. This is not a secret in the industry. However, even for very experienced modelers, they will often keep some topology diagrams as a reference, which can indeed improve efficiency and reduce errors. Therefore, I attempted to create this tool, hoping that AI can learn topology.
Fortunately, I succeeded (to some extent).
TXT2IMG & IMG2IMG
This tool is divided into two models, one for text-to-image and the other for image-to-image.
How to use it
TXT2IMG is relatively straightforward. The canvas size should be around 1024 pixels on one side, such as 1024x1024, 768x1024, 1024x768, etc.
If it is a close-up, it can be reduced to around 768 pixels. Because wireframe needs enough pixels to be drawn clearly,
add wireframe, <lora:topo_v3_t2i:0.6> to the end of your prompt.
If you try other models, the lora weight will suggest between 0.6-0.9, but it is best not to exceed 0.9.
IMG2IMG is a bit more complex. I am currently only sharing some images, which were all searched from Google. It needs to be used with ControlNet, and I will write a separate tutorial and come back to update the link later.
(updated tutorials)
A more detailed tutorial for txt2img
https://civitai.com/articles/1065/can-lora-learn-topology
Use Guide for img2img
https://civitai.com/articles/1110/user-guide-for-topology-img2img
Capabilities
You can use high-res! 0.33~0.44
You can use ADetailer to retouch the face! (If you are drawing a full-body picture, you may need to use it.)
It can draw a gray model by adding "monochrome" to positive prompts.
It can draw a color model by adding "monochrome" to negative prompts.
It can change the wireframe color to some extent by using "green_skin", "red_skin" and similar prompts (not perfect).
Model Selection
As it is a style model, all models are suitable, even 2D models. However, the 2.5D model without noise offset works best (I must say that I have some objections to the current proliferation of noise offset, and I hope everyone can reduce its use, especially when training Lora, because I found that it will overlap with the model and make the picture dirty. It can be said that there is no reason to use it in 90% of LoRa training! This was originally a very good invention, but it has been abused by many people who do not understand it, which is very regrettable.)
I can't draw what I want?
Because it is LoRa, it has very significant limitations. Currently, TXT2IMG can draw things including but not limited to girl, boy, dog, pig, frog, rabbit, Mandalorian. It is almost impossible to draw any mecha content (it will become a completely different style).
However, IMG2IMG can have a broader range of applications, but its drawing quality is relatively lower (because it is IMG2IMG).
You can help me make it more useful
I am looking forward to more professional collaboration. Although the IMG2IMG model can meet some work requirements and make it more like a tool, even if I try to minimize the influence of the style, I cannot completely eliminate it. This is a limitation of LoRa itself. Therefore, I am looking forward to collaborating with professional game development teams so that I can create models that are more in line with actual needs and further improve their quality and functionality (such as... many ideas).
中文
关于V4
这个其实在我训练的时候我给它起名叫V3Fix, 因为它的训练集数据和V3完全相同,我只是修复了一点错误。不过,它是用完全不同的方法训练,它现在是一个去风格化的LoRa模型,对底模的style影响微乎其微。
因此它现在可以被套用在几乎所有类型的底模上,包括2D, 2.5D, 3D, Photo, 并且,你不再需要两个模型,无论是txt2Img还是img2img, 一个模型就可以了, 因为它的使用方式发生了如此大的改变,我还是将它重命名为V4比较好吧,避免造成大家的混淆。
V4的权重给1就好了, 减小该权重可以一定程度上减弱线条的强度(但也可能会造成线条丢失),推荐1~0.7之间,其它包括 HiRes,修脸,使用tip,暂时请参照V3说明,我之后有空会再补充个教程。
Img2Img得到了非常大的改善,因为它对图片风格的宽容度比V3强太多了, 现在它变得更擅长给图片套上topology
是什么?
这是一个实验模型, 用来给模型添加拓扑线的工具。将一切都拉入3D世界(虽然这个目标还很遥远)
也许未来可以与nerf结合做点什么? 我不知道
为什么?
我曾经在游戏行业工作了许多年,虽然我现在早已不再是一个美术相关的从业者,但我仍然觉得AI在游戏开发中,应该存在着更多的可能性,当然这并不意味着AI可以取代那些优秀的艺术家,但我认为AI可以帮助他们减少一些日常工作中的重复性劳动,这个GAME DEV TOOLS系列都是基于这一想法而产生的一系列工具。
TOPOLOGY是3D模型师需要掌握的一个非常重要的技能,专业从业者甚至通过拓扑的好坏来判断一个模型师是新手还是老手,这并不是什么行业秘密。不过即使对于非常成熟的模型师来说,往往也会将一些拓扑图放在手边作为参考,这确实能提升工作效率和减少出错。因此,我尝试做了这个工具,希望让AI能学会拓扑。
TXT2IMG & IMG2IMG
这个工具分成两个模型, 分别用于文生图和图生图,
使用
TXT2IMG比较简单,画幅单边在1024左右,比如1024x1024, 768x1024, 1024x768等等,
如果是close_up可以降低到768左右。因为wirefame需要足够的像素才能画得清楚
在你的prompts末尾添加: wireframe, <lora:topo_v3_t2i:0.6>
如果你尝试其它模型,那么lora权重会建议在0.6~0.9, 但是最好不要超过0.9
IMG2IMG稍微复杂一些,我暂时只放一些图,这些原图都是从google中搜索的。
需要搭配ControlNet, 我会单独写一篇教程,晚点回来更新链接
能力
可以Hires! 0.33~0.44
可以ADetailer修脸!(如果你画的是full_body, 你很可能会需要用到它)
可以画灰色模型,在positive中加入 monochrome
可以画彩色模型,在negative中加入monochrome
可以一定程度改变wireframe颜色, 使用green_skin, red_skin之类的(不是很完美)
模型选择
因为是style模型,所有模型都适用,甚至2D模型也可以,
不过2.5D模型, 不带noise offset的模型效果最佳(插一句,我对目前泛滥的Noise Offset颇有意见,我希望大家能减少它的使用,尤其是训练lora不要使用,因为我发现它会与模型叠加,让画面变脏,可以说90%的LoRa训练都没理由用到它! 这本来是一个非常好的发明,但被很多不明所以的人滥用了,非常可惜。)
我想画的东西画不出?
因为它是LoRa,它有非常大的局限性,
TXT2IMG目前可以画的东西包括但不限于,
girl, boy, dog, pig, frog, rabbit, mandalorian
目前几乎不能画任何mecha内容(它会变成别的完全不同的风格)
但IMG2IMG可以有更宽泛的应用,但它的绘制质量要差一些(因为是IMG2IMG)
你可以帮助我将它做得更好
我期待更专业的合作,尽管IMG2IMG模型能够满足一些工作需求,让它更像个工具。但即使我尽可能的减弱风格的影响,但我没法完全消除它,这是LoRa本身局限。所以我期待着与专业游戏开发团队的合作,这样我可以做出更符合实际需要的风格的模型,也可以更进一步提升它的品质和功能(比如... 更多)