@Zovya I don't think I've seen an embedding that ISN'T .pt. While I refuse to use them and zip files for other things, I'll reluctantly use them for embeddings and just make sure civit and numerous other of my security programs make sure they're clean...

picklescan is your friend.

pip install picklescan

But everything Zovya says is also true. You have nothing to worry about downloading a .pt from Zovya, who is a long-time and well-respected contributor to CivitAI. Been here as long as I have—over a year. I've downloaded many of Zovya's uploads during that time.

@dita Pickle is VERY insecure format (can result in launching ANY code - you can see warning about it even in official Python documentation: pickle), and should be banned on model sharing sites. When you distribute any model, even the smallest one, it should be always be safetensors. Models are not executables, so you shouldn't distribute them in a form that allows arbitrary code execution - period. Antivirus and picklescan helps, but none of them gives you 100% guarantee to detect and stop every kind of malicious code.

@Zovya You shouldn't ask people for that kind of trust. Distributing pickles is like "allow me to run any code on your computer". What kind of code you want to run on users computers? If none (as it should be), then just use safetensors.

@Zovya You're welcome: Examining Malicious Hugging Face ML Models with Silent Backdoor.

@Zovya It is a big issue and people should be aware of the security flaws (enough already lose their information and privacy to big tech as is and many people become victims of identity theft and the life which has lifelong impact) but to my knowledge is it even possible for an embedding to be in safetensors format? Maybe you aren't embedding pickles but what's to say 5% of the uploaders here are?

@Kaladae Embeddings work just fine in .safetensors - and conversion is trivial, too. You can check it out on unaestheticXL, for example.

The only reason a LOT of embedding creators have OPTED NOT to make safetensors is the following:
1. PT can be a risk YES - it's a pickle tensor file.
2. HOWEVER: Embeds are NOT LIKELY, very HIGHLY unlikely especially with the origin being of civitai TO BE PICKLED in such a way that serves bad code. (Hint: UNLIKELY but not impossible)
3. THe major reason: 90% of people i've run into ask "IS THIS A LORA" - So i serve PT files instead, there ARE safe conversions on the web or via colab
4. Trust me when i say this: Zovya and other creators that have been here since nearly the start WOULD NOT and will not add pickling VIRUS code in to execute on your machine...

A PT file willnot serve BAD code for execution unless it is (and yes i realzie i'm not exactyl right nor do i have the info hadny) -- somehow Injected.

Aka: I would have to know what i'm doing for it to do that, and if i upload one of Zovya's as a backup right now to a huggingface repository it WILL do a pickle sc an and tel lme what callbacks it does or doesn't have.

I'd be more worried about an EXE file than a PT file. (At least from civitai)

The thing that keeps coming to mind about this discussion is the fast food meme with a *.pt on a plate and the civitai girlfriend saying "He said no pickles!"

In his situation, virus threat isn't a real threat because/if his real identity is linked to Patreon and Ko-Fi.

Greater risk is that new creators think going against the consensus is cool.

Going against the "consensus" (aka "standard") also means that it's much more difficult to use in all the software that will not read .pt files since they are, by definition, unsafe.

can someone help me convert this safetensors? my pipeline cant load these by default

Can you just rename them to a .safetensor file?

Ive noticed this as well. Tested it on Pony Diffusion XL V6, AutismMix and AWPmix2. You can get better results with score_x_up

If I were to guess, I'd expect that this embedding is the score sequence, bundled into an embedding. If it's less effective, it may not contain the full sequence (i.e., missing score_4_up) or (much less likely) it might contain other terms like "masterpiece" that only serve to muddy or confuse the prompt. If that's correct, the NSFW embedding would be a similar thing for terms like "rating_explicit".

People tend to assume embeddings have to be trained like LoRAs, and many are, but they can also just be a simple collection of other tokens, bundled together with an easy-to-use nickname. The Embedding Inspector extension for A1111 can show you the exact contents of SD1.x and SD2.x embeddings, although it's no longer maintained and doesn't work for SDXL (which is the only reason the above is a guess and not a statement).

Simple embeddings like that are genuinely convenient, as a six or seven character word is a hell of a lot easier to use than "score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up", but this model page seems to be selling it as a kind of easynegative for PonyXL, and I'm 99% sure that's not what it is.

I think they mean Pony Diffusion XL.

@mog I do not understand. (The message this was in response to was removed.)

https://civitai.com/models/257749/pony-diffusion-v6-xl

A real human being.

there are 2 trigger words zPDXLxxx for the possitive and zPDXLxxx-neg for the negative prompt

@derpmagician excellent, I'll use them both moving forward.

they go in the /stable-diffusion-webui/embeddings/ folder. once you load up automatic1111 they will be in the "Textual Inversion" tab. You can add them to your prompt by clicking them there, or simply typing the embedding name into your prompt.

@sixpt55 it doesn't help at all. Nothing close to original pics with this files

@Bugagish Are you using the Pony Diffusion model? These embeddings were created for use with that model and probably won't produce expected results with other models.

@sixpt55 Yes, Pony Diffusion model

@Bugagish Hmm. Sorry, it's hard to ID the problem without seeing. When in doubt, you could try copying all the image generation info from one of the sample images (seed, pos/neg prompts, sampler, dimensions, etc) and try reproducing the image with and then without the embedding to see the difference.

@sixpt55 I placed the files in that folder but they are not showing in the textual inversion , I have many embedding files and non of them are showing there

Because it absolutely doesn't replace them. The score_9, etc is still necessary.

my experience so far is that you still need them or you'd mostly have blurry/bad hands, which this embedding doesn't seem to directly address

I Forget you dont' need the scores adn i still use my score embedding ... >_> whoops.

@smitetrash675 true, it looks much better with scores

the examples dont have numbers because the author uses the embedings instead, every platform has their own way to use embedings

@derpmagician i just need an example how to put it in prompt properly, that's all. Like small guide for a LoRA's models

@Bugagish i dont know how to use it on A1111 but  on comfyui you put it in the description of the image like this:

(embedding:file_name:1) or

(embedding:file_name:1.4) if you want it to have more emphasis

@derpmagician so, it may be useless in a case of A1111 use? It's only for ComfyUI?

@Bugagish there should be tutorials on how to use A1111 with embedings on youtube

@Bugagish It works the same way on A1111 (Tag:1) = normal, < 1 weaker strengh, > 1 greater strength.

I'm not a comfy user, but i've used the ZPDXL ones in general in Auto, it MIGHT be an application of your node wires?

sounds like the embedding does not support the resolution you are using (its above standard and probably not to be recommended unless its for upscaling). But all that does is disable the embedding so you should still be good.

@JohnnyB1 Don't normally work at that res, was just trying to replicate someone else's params. Strange that it's only zPDXLxxx-neg and not the others (which work fine at that res). Still its usable at 1024, which is fine.