XSMerge-RealisticVisionV3-ForArchitectural(建筑向RV3)
注意:本模型并非原创模型,只在其基础上对目前的行业使用习惯进行了简单的微调,原模型作者链接:Realistic Vision V3.0 - v3.0 (VAE) | Stable Diffusion Checkpoint | Civitai
使用建议:
自RV3.0作者发布模型以来,很多建筑、室内向的AI研究者反应模型在这些方向的表现有些单一,以至于部分prompt如同陷入了过拟合等类似的情况(列如出图时玻璃的框架几乎都是黑色金属,需要特别标注prompt才能改善等等),并不能很好的进行多方位的图面表现。我将RV3.0模型从底层参数上进行了微调,让其模型更偏向于建筑等我所在研究的行业中,以达到让更多使用者有更好使用体验的目的。我在微调后进行了merge,是想要达到很好的泛化性,但似乎丧失了一些真实感。在出图时看似图面的质量会略微低于RV3.0一些,但是可以通过我所制作的轻量模型(Lora等)进行改善,甚至使用后出图会高于原本只用RV3.0模型的出图。
Since the release of the model by the RV3.0 authors, many architectural and interior AI researchers have responded that the performance of the model in these directions is somewhat simple, so that some prompt seems to fall into similar situations such as overfitting (for example, the glass frame is almost all black metal when drawing, requiring special marking prompt to improve, etc.). It is not very good to carry out multi-directional graphic performance. I fine-tuned the RV3.0 model from the underlying parameters to make the model more oriented to the industry I studied, such as architecture, so as to provide more users with better experience. I did merge after fine-tuning to achieve good generalization, but it seems to have lost some of its realism. In the drawing, it seems that the quality of the drawing surface will be slightly lower than RV3.0, but it can be improved by the lightweight model (Lora, etc.) I made, and even the drawing will be higher than the original drawing using only RV3.0 model.
RV3.0の作者がモデルを発表して以来、大勢の建築・室内向けのAI研究者は、モデルの表現がやや画一的なため、一部のpromptが過適合に陥っていると反応している(図の場合、ガラスのフレームはほとんどが黒金属で、promptを明記する必要があるなど)。多面的なグラフィック表現がうまくできていない。RV3.0モデルを基礎となるパラメーターから微調整し、建築など私の研究する業界に近づけることで、より多くのユーザーに体験を楽しんでもらうことを目指した。一般化を意識して微調整してmergeをしたのが、実感がなくなってしまったようなってみたい。見た目の品質はRV3.0よりも若干落ちてけど、私が作った軽量モデル(Loraなど)で改善できので、この問題を解決した。使用後はRV3.0のみのものよりも高くなることもある。
配合一些TI, VAE等使用文件
如果您有任何问题,请通过[email protected]与我联系
Suggestions for use form XS:
TI, VAE and other documents for use
If you have any questions, please contact me at [email protected]
XSからの使用アドバイス:
TI、VAEなどのファイルを一緒で使ってください
何かご質問がございましたら、[email protected]にメールしてください。
Description
safetensors
FAQ
Comments (15)
Why upload .ckpk and not .safetensor?
bc the default for merging in auto is stupidly still set to .ckpk, from what i read on reddit they're planning to change the default since the problem became more obvious now that fraudulent "leakers" offered SDXL and 0.9 checkpoints spiked with malware. Stability ai themselfes posted not to open any pickled files and that all future checkpoints from them going onwards will be released in safetensor format. 95% of model uploads here did get that note like 7 months ago but it obviously didnt reached everyone yet especially on the other side of the world where people would have to use a translator to read those posts. i guess XSarchi was too busy doing architecture work or training checkpoints to follow the subreddit regularly. i hope you read this and adapt your uploads in the future, its just one additional click in training but will grant you much more downloads since no one trusts pickled checkpoints anymore.
@JarJar737 I'm gonna take a chance on it, I like living on the edge!
@JarJar737 Thank you for your reply. I read English. I just don't care about so many details, I will upload the files you need. But I don't think even safetensor is very good at keeping the files safe.
@XSarchitectural why'd u think that? i mean, nothing is 100% safe, but its the best we got atm and way superior to checkpoints which can execute arbitraty python code. and AFAIK there are no safetensor exploits out yet, so the first one to find it would be pretty stupid to use his shot hacking some weebs on civitai instead of selling it
@JarJar737 I don't think there is much difference in security level between the two files. Because people who want to do bad things also improve and learn, much like when I go and buy an iPhone, which is better, version 14 or version 12? 14, of course, but 12 can continue to be used, and the virus can be upgraded to version 14 only once in a while. So I think it is very necessary to check the security of the files ourselves, and the website should also be responsible to help users check the files uploaded by each author.
@JarJar737 I think this is also a very interesting question, and I would also like to ask you what your views are on this matter.
translated using DeepL:
嘿,Xarchi,请考虑以.safetensor格式上传你的合并,人们通常对下载pickle tensors犹豫不决,因为它们是一个很大的安全问题,稳定性AI本身最近警告说不要打开任何pickled checkpoints。该格式在半年前已被safetensor取代,但automatic1111的合并标签仍然使用ckpt作为默认。
也非常感谢你在civitai上为建筑界所做的伟大工作,你是ai建筑界的山羊<3
__________________________________________________
hey Xarchi, please consider uploading your merges in the .safetensor format, people are generally hesitant to download pickle tensors, because they are a big security concern and stability ai itself warned against opening any pickled checkpoints recently. the format has been replaced by safetensor over half a year ago but the merger tab in automatic1111 still uses ckpt. as default. take the time to check the safetensor box instead and you will instantly get more downloads.
also much love to you for the great work you are doing for the architecture community on civitai, you're the goat of ai architecture<3
thank you for reminding me. sure, I will upload the files you need.btw, i read english lol
I am Posting a new comment section in the comments section where we can have a chat about this issue and others are welcome to join in and give their own opinions.
All large models have been updated in safetensor format, users who need to download safetensor format.
Thank you! I mistakenly included this with some models I was using to generate people, and the results were only slightly horrific! I had to laugh when I realized what model did that, I can't wait to use it for its intended purpose!
"Talk about File Security"
We welcome all users to participate in our discussion.
Have your say in this comments section.
well i cant say much more than what i read in subreddits:
AFAIK .pt .bin and .ckpt can all include arbitrary python code, so without any security check at all, theres the potential thread of basically any virus/trojan/malware attack imaginable.
The .safetensor format tried to mitigate this risk by being basically just plain data. ive read someone probably much more qualified then myself saying that its theoretically possible to create a safetensor in such a way that it would be able to surpass this barrier by breaking out of memory and then being able to execute code, but ive yet to see anything on the web related to executing such an attack.
for stuff scanned by civitai its like with a normal anti-virus, the pickle scanner will look for code inside the data that could execute malicious commands, but it can only recognize what it knows. so there is always the possibility that someone could craft a file in such a way that the scanner wouldnt recognize their new yet to be catalogued malicious code.
for safetensors the same risk exists, but its definitely more complex to write something that breaks out of the plain data restriction and then execute said malicious code, it takes more layers and a much more sophisticated attack for this filetype, and those who would first find a way (if there arent already thousands of those on the dark web that i dont know about) should definitely sell the code for millions or hack some pedo billionaires instead of some poor as weebs on the internet.
Thus we can conclude that no filetype is safe. nothing is safe with computers, EVER.
its all about your personal risk evaluation and the risks you take.
downloading safetensors from reputable sources with many followers and downloads who even give out personal information and stand behind their models with their own name are no guarantee you wont get hacked, but they surely are the most risk free option atm.
on the other hand "leaked" sdxl checkpoints in .ckpt format, of which staff of stability ai themselves warns against on reddit are a sure fire way to get hacked.
same goes for not yet verified .pt embeddings from people with just one upload. they certainly dont have to be nefarious, but i'd recommend letting someone else try it out first.
This is where we come fake likes and stuff. one could easily make a few accounts to download and heart a malicious model that you just crafted and uploaded to make it seem more legit, look at their followers and the followers uploaded models and generated images if your suspicion is raised.
if you work with stable diffusion and your monthly income depends on it, i'd be extemely careful to only download safetensors from reputable sources.
if you got an off side backup and are a normal everyday consumer, just dont download files in not safetensor formats from people you dont trust.
no matter what, if you join the realm of downloading cutting edge technological files from the interwebs that in any shape or form break anyones copyright, you should always protect yourself as if you're downloading random files from strangers on piratebay. Use a good (better multiple) anti-virus software, regularely update windows/linux, especially security updates. have a backup of your files, and another backup off-site, meaning air-gapped, not connected to the internet or pc, update it monthly or something, not too often. modern malware can be as intelligent to wait on your pc for a certain amount of time in the hopes of being backed up on all your storages so you cant roll back when they encrypt all your files and ransom you to pay them in Monero or some other dogshit memecoin. If you get caught by this scheme, dont pay the ransom, its almost certainly not resulting in them decrypting your files. if they want bitcoin, call the cops, the blockchain records every transaction ever, if you're lucky and they aren't mixing their coins randomly with other criminals, they can certainly be traced back and stopped from scamming more people.
And lastly, take everything you just read with a grain of salt. you never know what exploits may already be available to paying customers on the dark web or governments.
one should think you're safe if you arent downloading shit like cp or bestiality, but its not that easy.
I can imagine downloading russian/ukrainian models midst an ongoing war being much more likely to contain trojans than for example these architectural models.
on the other hand, just bc it seems professional doesnt neccassitate professionalism, theres a lot of money in architecture, if you can hack everyone using your models to then let a scanner automatically rate the best generations and use these to sell as your own architectural ideas to customers could be a great dirty business scheme. Im of the strong belief at this point that XSarchitectural is a fair player and i love his models, just picked this example so you realize the most important thought process before downloading anything from the web: "what could the hacker get out of it?"
if you load illegal stuff you can get blackmailed, if you load legal stuff that has a potential monetary benefit, people will try to get rich off of your work/compute resources.
you will like with everything you do in life do a risk benefit analysis and evaluate weither its worth it to you personally.
also dont trust other sites without a scanner at all and don't support sites that dont credit the creators. It may seem safer to download from a site you pay to use, but most of these sites just scraped civitai and are selling free models without giving credit and sometimes even injecting malware additionally to scamming you.
Stay safe out there, much love y'all
谢谢
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.







