@Nerfgun3 does this work at all with realistic models? or is deisgned for anime?

works well with colorful

@Nerfgun3 can u add some Majicmix Realistic examples please?

@aikanojo Here are some examples using majicmixRealistic_betterV2V25
https://civitai.com/posts/595309

Use it with after detailer (adetailer) or uddetailer extensions to target only the hands, or use it with mnanual inpainting targeting only the hands.

@Nerfgun3 I appreciate your good answer on this, I never seem to find a good solution. Your explanation to think of them as style embeddings puts a great perspective on how they operate that makes a lot of sense. It also explains why I would frequently end up with clothing the wrong color, etc.

@Nerfgun3 thanks this is really awesome actually, the fact is I noticed as well that many negative embeddings would lead to unpredictable results and sometimes that is great and actually make the image more interesting but if you want to achieve just the negative commands and no art style change, that's splendid. Thanks!!

If you've already used it to know that it's good, then it doesn't really matter. Either it's safe, or you're already screwed.

@sergiohzph601 this is so true , i dont know how to make good prompts

@S3r6i0 Where would you suggest I start learning how to make good prompts? Thank you

可以尝试搜索以 "Ai 绘图 Pt 文件如何使用".

@Nerfgun3 Hi, ive search enitre comment section but didnt find your post. Could you point to it please ?

upd: BadDream+verybadimagenegative_v1 pair is actually good when you want stronger effect then negative_hand but not so it will alter your image, I dont know why but they balance each other and if you use one consider do some renders in pair. I keep them both in styles and sometime slaping (either pair or negative hand) to promt I want to test. Works great! Yet I feel negative_hand will stay my "by_Default" TI and its still to find usecases for slightly stronger effect of pair but its a great alternative

"Changing the image" is not a bad thing or "destructive". You're saying a lot of things are bad when all they're doing is making a different image. I've tested adding a couple embeddings only after 20% like you suggested and all it did was exactly that, produce a different image. Sometimes it fixed issues, sometimes it added issues. From my own testing, this embedding tends to add shoes when socks would be present with another negative hands embedding. Not something that I'm after, but if you don't mind or that's desired, then go for it. This also depends on the model you're using, it might not do that with different models. You're also not supposed to pick an image, lock the seed and add embeddings. You're supposed to use the embeddings from the start and find your seed after. Otherwise of course you won't like the embeddings. You're starting from something you like and changing it!

interesting experiment! Can you try the same with my bad_pictures TI? I noticed a lot of people just throw a bunch of negative TIs together and call it a day. Please let me know how mine interacts with the other TIs...

@Vivipeg First of all I joined txt2img community just 3 weeks ago, thats why Im questioning every legacy norm and do my tests, yet I am in no way a respectfull source of truth.

I have generate more then 50k images in this 3 weeks, and 90% of them were in photo portrait style, so while my eye already get used with some photoreal models, I have nothing for anime/cartoon models that as I understand your TI is focusing on.

Infact having the same faces over and over again was the reason why I start testing different TI, I was using a pair of 2 extremly strong TIs that basicly overwrites anything that any model and prompt was generating. Its anecdotal but when I scrolling through image reel of this website sometimes I see face and know exactly what TI were used, and this is on top of limited capability to generate new faces of models by them self. Now with this understanding and without token noise my generated portraits are MUCH diversive and sensetive to prompt.

So this is all I got in terms of experience and expertise, this is why while I will leave my opinion below, its would be better if you will see everything with your own eyes.

I deployed small example here https://freezerain.github.io/stable_diffusion_bad_pictures_ti_test/#auto-loc,true,true,false,negative_prompt,cfg_scale,none,none,1,1,1,1,1,1

If you never used infinity table before - small explonation:

On the X axis horizontaly I prefer "negative prompt" variable that in this case are different negative TIs.

On the Y axis vertically I prefer CFG value to increase TIs power.

Use arrow keys left-right to switch between TIs and up-down to visually see effect.

What to look in my opinion:

-Usually I use empty negative prompt to validate that there are no change in image between 4 and 9 CFG, however I have to inject "!" symbol for prompt replace capability, so there are small visual change even without any TI.

-Move through different negative prompts to see how image change relatively to original "empty prompt" (the one with "!" symbol in this case).

-Move through different cfg values to examine changes closely.

-In upper part of the table(do not confuse with table axis buttons) switch between empty prompt and the one that original image were generated (I left only quality modifiers). So you can see how much TI affects the image with already positive modifiers like (masterpiece, 4k uhd quality by greg rakatovski and all that garbage). If no effect then TI not needed

-In upper part use "prompt replace" to inject my generic negative prompt in place of "!" symbol, this is keywords that I tested with prompt matrix that doesnt have duplicate effect on the PAINTED(not photos) style images, for example I dont use (bad quality, low quality) together, because from my testing using them In pair have zero effect on quality, only token noise, and between them (low quality) were more powerfull atleast in my test set so I choose it. Same logic as above, if TI not contribution enough to already filled negative prompt then its not needed.

-Use "Prompt replace" to inject bad_pictures in place of "!" symbol and examine in pair with other TI (this is different from what I tested originaly, I used 2*2 prompt matrix and not infinity table, but data is the same).

-Dont look at pixels and small compositional changes, look for quality changes. Its not important if model will have a bit different dress, its inevitable that you will get different noise from different tokens, the question is will this dress be lit better, will it have more detailed texture etc., majority of changes nor good nor bad, but I consider this as bad as any TI eat token budget while contributing noise. I also looking for "cheap HDR" effect and obviosuly makeup faces, lips, eyes that overwrights original.

-I left ng_deepnegative_v1_75t and easynegative TIs as bad examples (originaly their both with 1 minus). I hope it will be easier to see why some TIs got + and other are not, and remember this 2 bad exmaples is not even the worse. In this case easynegative have very powerfull face, and ng_deepnegative_v1_75t have combination of face, HDR and 0 skin texture (lmao)

-I generated 3 seeds, so you can make sure its not just a bad luck seed

bad_pictures performance solo:

-Slightly leans to asian, very noticable on seed 3

=light Hdr effect, not a problem untill high (>10) CFG values, helps on low CFG with contrast

=some TI have effect of subsurface scattering in this test set, you can see it in ears, I dont know if its good or bad but 50% of TI have it

+more background details

-some kinda problem with teeths, cant comment on this one, futher testing needed

+not very much contribution on good prompt except more detailed background, can be good sometimes as face is stable as fuck and not asian at all

++over all every TIs(except deepnegative) didnt messed up with good prompt but bad_pictures were in top 3. Intrestingly that "negative_hand-neg" changed less then bare "!" symbol, lol

--With negative prompt all TI were still evaluating through CFG, but bad_picture completly stayed the same as without negative prompts. I think its somehow suffocated prompt. This means you will have bad time affecting your result with prompts, this is objectevly bad as having same repeatable image is work for character/style lora and not negative TI.

--All TI (except deepnegative ofcourse) were perfect with filled both normal prompt and negative prompt, even bare bone "!" just look how this beautifull flower bloomed with higher CFG. But on the first seed bad_pictures was the worse of all of them (except deepnegative), its the only TI that make here blond, change strangly facial expression, intorduced half baked shoulders and destroyed flower in hair. But it was only 1 of 3 seeds, so maybe bad luck. And in other seeds I saw less changes then "!", same problem as without normal prompt. Not good even if forget about first seed.

Performance in pair:

First of all if you want to see easily how it working in pair go in advanse setting in top of the page and disable "prompt replace" with keywords, so only "!" and "bad_picture" will stay active, then in mini table with axis settings just above main table swtich X to "prompt replace" and X-Super to "Negative prompt". now you will be able to easily see difference between TI and TI + "bad_picture" using left-right arrow keys.

--you can see how powerfull this TI is by looking at "negative_hand-neg + bad_pictures" It completly overwrites gentle flower negative_hand-neg is, dominating token space entirely. This is very egoistic TI and there are no much sense to use it in pair (tho like any pair of TI ever except "BadDream+verybadimagenegative_v1")

--another example is very good CyberRealistic_Negative-neg that when paired with bad_pìcture is losing background details! and reverse to asian obviosuly

=There some good improvements with filled prompt keywords, however its difficult to see if those improvmentes are actualy balanced to create a pair, I actually dont know, potentially maybe

tl;dr Overall bad_pictures score:

Doing asian faces, usable preferable with <10 CFG, super egoistic - will not work well with any other TI, completly dominating negative prompt making your other keywords useless.

On the other hand with good prompt and no negative prompt performed better then majority of TI, and consistently was adding details to front and back.

I think this TI should be used only with no other negative keywords and very strong prompt but be aware of asianating face and additional contrast.

so I give this TI ±, very specific usecase but not for everyday. Which is not what your actually expecting from TI right? You want just slap it and forget right? lazy dog. no this TI is almost weak Lora in my eyes.

My rating of all other TIs in order of table:

1) ± bad_picture - you can find usecases if you really want to

2) ++++ negative_hand-neg - still king

3) + BadDream - very strong and very close to bad_picture, but more consistent and leaving space for other keywords.

4) ++ verybadimagenegative_v1.3 - same as BadDream, but not as strong, more gentle.

5) ++ BadDream / verybadimagenegative_v1.3 - the only working pair of TI, yet difference is very subtle I dont think someone will care enough, try this if you use BadDream or verybadimagenegative_v1.3 solo.

6) +++ CyberRealistic_Negative-neg - actually solid TI, im impressed wow, I fear that the face is too strong and will be in all the photos, need to test it in bulk

7) -easynegative - just mediocre TI that doesnt contribute quality only steal face and token space

8) --ng_deepnegative_v1_75t - this is worse then in previous test (because in previous test were ng_deepnegative_v1_32t that made my eyes bleed). Just look at results.

9) ++and_a_half_+ UnrealisticDream - very good TI, doing babyfaces but its easier to alter then asian faces by the nature of data sets, so no biggie, everything else is solid.

Thats all for now, if you want to test additional combination just pull this github and edit table layout.

@freezerain Thanks for the overview. I tried the prompt with different models and the asianfy effect seems to be model dependent. Can you try your test with various different models including the base SD1.5 and let me know your thoughts? Here's my results so far: https://civitai.com/posts/304224

@freezerain in fact, there is one negative prompt that you should never use, becaude it completly change the style of a model.

It's ; poor lighting.

I did a few test whith negative Embeddings and without..... And i spotted the one that use this negative prompt.

They should remove it, this prompt is pure crap that alter the style and doen't really upgrade the quality (the lighting is more uniform and all over the place).

It's also one of the rreason why you have diffulty to find goor pairing.... If you put two Embeddings with this negative prompt it's just, wow...

And, i disagre about negative-handng..... it affect the style even worse than easynegative and doesn't fix hand on many model.... I had more ssuces without it than with it.

@Lorim RN i have around 15 grids of 10k test images of 21 or so different TIs in different conditions. I expanded my testing to wider scope. There are around 20% more of test images to generate however if I want to share my result the most difficult part will be to write text explaining methodology and speculating about the result.
I noticed some connection between effect strength and TI vector count, however even without it I think neg-hand is the most consistent of all of them and most "useless". Oh and there are some hands actually in test images and yes a LOT of TI totaly destroying them haha so ironic I actually enjoing it.

I wish I knew about poor lighting before, in theory it should be possible to append it to grid.

If I will be able to finish text part I will post grids here on this web site as separate article so if you wish you can subscribe on me to see notification :))))))) Im honestly not baiting :=)))))

Or save your time and ----DONT USE NEGATIVE TIs AT ALL----

@freezerain yeah i don't whant to use Negative Embeddings anymore.

this why, i'm ding something even more tedious than you : i'm testing a lot fo mdels with different perfect prompt withe same seed on my base model.... then i removing those that :

1) have mesed hands.

2) messed colors

3) poor details

4) messed hairstyle

5) bad anatomy and proportion

but since i have a slow connection, i will take me a long time.

But, in the end, i will merge those models and create my own moel who only need (masterpiece, best quality, absurdres) in positive and (worst quality, low quality), lowres in negative ^^

Can you review these embed reviewer god

bad-hands-5, Unspeakable-Horrors-Composition-4v, EasyNegativeV2, bad-artist-anime, bad-image-v2-39000,

@illyaeater Hey thank you I was not aware about new versions. I can give you for now few links to infinity tables, they already contain Unspeakable-Horrors-Composition / 4v/ 64v, EasyNegative, bad-artist. Not ideal but maybe will be usefull to you for now

https://freezerain.github.io/img2img_ti_comparison/azovya3/#auto-loc,true,true,false,styles,cfg_scale,none,none,1,1,1,1,1,1,1

https://freezerain.github.io/img2img_ti_comparison/epic5/#auto-loc,true,true,false,styles,cfg_scale,none,none,1,1,1,1,1,1,1

As someone who relies on TI's, I thank thee for thine divine contribution! although I don't know what's good with a anime model

@Noxy_ Try this negative prompt ^^

(worst quality, low quality), normal quality, lowres, watermark, signature, greyscale, monochrome, jpeg artifacts, blurry, out of frame, aliasing, imperfect, ugly, gross proportions, disfigured, dehydrated, malformed, mutilated, mutation, mutated, misplaced, error, poorly drawn, what, body horror, anatomical nonsense, engineering nonsens, wrong hand, wrong feet, wrong fringers, wrong digits, extra arms, extra legs, extra hands, extra feet, deformed hands, defromed feet, deformed fingers, bad anatomy, bad proportions, bad perspective, bad drawing, bad lighting, bad details, bad shadow, bad outline, bad highlights, bad colors, bad hair, bad face, bad body, bad skin, bad muscles, bad joints, bad background, bad foreground, bad clothes, thick eyebrows, misplaced genitals, bad genitals, wrong hair color, wrong hairstyle, wrong hair length, wrong eyes color, wrong breasts size, wrong clothes, wrong clothes color, bad makeup, bad jewelry, bad accessories, bad ornament, bad object, bad vehicle, bad building, bad vegetation, wrong colors, extra clothes, imperfect hair, imperfect face, imperfect body, imperfect clothes, imperfect hands, imperfect fingers, fused fingers, fused limbs, fused clothes, malformed hands, malformed fingers, entangled fingers, bad reflection, bad refraction

@Lorim You a legend! I'll try em and edit if the results are bad or not

EDIT: After some generation, it turned out pretty.... mediocre but it does look like a anime though, thanks for the negative prompt

@Noxy_ All you really need is (worst quality, low quality:1.4). You can add more like watermark, signature, logo etc. I also add stuff I generally don't want, like moles, earrings, piercings, simple background.

search for "negfeet"

there needs to be a "thumbs-up" emoji that's just the big toe sticking up made especially for people like this

don't get me wrong man, i'm not trying to knock your foot fetish, i can appreciate the need for one like that every time my models forget to add a leg. even the most niche of fetishes being included somewhere in the training data can improve these models overall 😂 just look at 2.1's current lack of support, in spite of its larger image template and better time of handling larger images, if you doubt me. (i normally use 2.1)

side-tangent - imagine how much better sd2.1 would be if people trained it on their fetishes instead of 1.5... a real human artist knows how to draw the musculoskeletal structure! a real human artist chooses not to do nsfw art, not because they can't, but because it's 'beneath them' or something like that.

ai's the same - i use NSFW models because they're unironically better at drawing the human form. they know what it looks like!

They're not, and it's not easy for creators of TI's to convert them XD it's secret ti's magic but also if pepole DO THAT then y'all go "BUTS THATS NOT A TI ITS A LORA"

@duskfallcrew pt is bad. I hope you get pickled for spreading misinformation.

@upscaleanon537 <_< "Smoking is Bad MMKAY" - Look if you're making a PT file on A1111 and your instance is fairly safe and set up the FILE ITSELF IS NOT INHERITLEY BAD - it's when someone maliciously uses that PT file.

Now now, I don't actually generate shiz local so you're asking someone to hack vast for my stupid comment?

IN THEORY A PT file is like a CKPT file: It's not a virus in itself until someone calls on the code -- and some files are worse than others, some files have MORE CODE than they should.

That's like asking if a ZIP file is dangerous or an EXE file -- or a JPG file

@duskfallcrew Of course a pt file in itself isn't bad, but they can contain malicious code, hence why people should use safetensors when spreading files, since you don't know. If you make the pt yourself, of course it's fine, that's not the point I'm making, nor what the OP is either.

Just like EXE's, you shouldn't just run any exe you find on the internet and think it's fine. You have 0 guarantee what code is in it.

There's a reason why almost everyone wants safetensors.

If you want to put your faith into completely random people, go ahead. But don't mock others for wanting a little bit of safety.

@upscaleanon537  The thing is CKPT is yes dangerous, it can get pickled - and YES a PT file could for Textual inversions - and i've myself convereted mine SOMETIMES to safetensors, but largely people don't understand that they're not LORAs.

I apologize for the humorous mocking tone, i was sort of not joking AT YOU but in general.
Let's be clear here, you're not stupid or ignorant, nor was your request -
It's just that honestly, the fearmongering tone got me a bit and you're misunderstandintg that I HAVE ACTUALLY CONVERTED TI's to Safetensors and gotten people going "Omfg it's a lora not a TI" - the average user doesn't understand it.

You by far are clearly smart enough to ask for a TI as a safetensor, sure - and that's valid that you want that - i'm sure even I COULD HELP on that (even though this isn't my TI) - and i acutaly know of a colab that can convert these.

And i don't find random EXE's i'm on a mac.
I'm 40 years old, I don't click fart button ads and get viruses, that i left ot my 60 some year old father i don't even talk to anymore lol.

Look i apologize for my snarky tone, but you have to understand the AVERAGE USER DOESNT GET how a TI can be a safetensors.

@duskfallcrew 

we have a tagging system, where it highlights "embedding", "lora", "base model" and so on,
we have the capability, of explicitly selecting, what type of that we want to see and what not,

on top of that, the "average user" should be capable of reading,
so, if you write at the very top of your description, what it is - and what it isn't

and the user still actively chooses, to ignore all of that,
than, it's on them, not on you ... it's, quite literally, their own problem!

and, while i don't know about any of the technical details involved,
and don't realy understand, what it would actually mean / involve, to properly convert these stuff into safetensors,

i do know, that pretty much any major player out there actively (!) warns against the use of pickles,
from invoke ai, warning on reddit against pickle's
up to stable diffusion, explicitly saying, that pickles are unsafe and can execute code

and here's the point,
yes, the person who provides it may not have any harmfull intentions,

but how would the average user know, what the intentions of the distributer are?
wouldn't it - by principle - always be best, to choose the safest option for everyone?

also, as a sidenote to end this text ...

if everyone continues using pickles for inversions becouse of the argument "some don't understand" ... than those some will never have the possibility, to learn about it ...

Back to the actual question which was never answered, is there any possibility of uploading a safetensors version of this?

All the Checkpoints, Loras, TIs, and VAEs I download are safetensors. It's just as a matter of principle. I have no intention of making exceptions here and there even if I should be able to trust the source.

@Exyloz If they ARENT getting back to you, let me know I"m able to convert these to safetensors, i just don't want to re-upload it to Civit, look i get your point and everyone else's - it's just MANY PEOPLE mistake TI's as safetensors to be Loras, and it's really hard xD -- PT's are bad when they GET pickled but if they're scanned by CivitAI, they should be "IN THEORY" fine, but i get why people don't trust them, so i'm here volunteering to convert a few to safetensors, just NOT to reupload them to Civit without permission

@duskfallcrew Who cares if they mistake them for loras? If they're too stupid to read, it's all on them. It says it's an embedding once in the title, twice in the sidebar, and several times in the description. Making them pickles for a reason like that is like saying, "Let's not put safeties on guns because most ignorant citizens won't know which way to flip the thing".

For now, you can find a safetensors version here: https://huggingface.co/Exyloz/safeConvert

@Exyloz Ok but TI's are less likely to cause an issue, CKPT's are the most common issue - and cool but ya know instead of RE arguing something we'd already covered -- yo ucouldv'e just offered the same thing you just did and literally just didn't re-hash shit XD

@duskfallcrew You might notice, I came in trying to move the conversation back to the original topic. I was completely ignoring that irrelevant conversation you started. You're the one that brought it up again, and even directed it at me. Had you not done that, I would have simply posted the link like you said.

@Exyloz Fair call, look i'm autistic as hell and i don't catch everything.
The topic at hand could've just been "SCREW YOU GUYS HERE'S A LINK" - i was merely nervous about doing it myslef asi t's not my TI, but fair call - and thank you for converting it- I'll be trying to catch up with my 100+ Ti's lmao

@Exyloz Hey thanks for that!
I actually didnt know anything about this whole pickle/safetensors thing and im really new to sd and civitaai.
So thanks for your advice! I will only download safetensors on civit from now on.

@Exyloz Just wanted to say thankyou for taking the time to convert and share those embeddings - feel much better using them in safetensors format! :)

Every embedding I have downloaded has .pt as it's extension

Usually these go into the negative prompt, though your preffered stable diffusion UI may work differently

xl dont support embeddings for now...

SDXL isn't compatible with 1.5 embeddings.

well which of the two? seems xl needs an alternative or a classic no negative prompt

@duskfallcrew true for most embeddings, but not all. For example, the Jessica Biel embeddings all work in SDXL, as does Marilyn Monroe

Clearly "negative hand" should be photorealistic pussy. Anime pussy is positive hand.

There actually is, it is called NegfeetV2

\stable-diffusion-webui\embeddings

@thepoplezz167 Thank you, worked like a charm!

确实是，有一段时间没上c站了，现在上c站发现全是一堆发小黄图的垃圾玩意。

Agreed honestly I dont mind seeing a few weird or naughty images but when Im trying to see if this is useful or not having a shit load of sex images doesnt help me.

Bro I feel ya, this crap is on every single model page (Checkpoints, loras, and even on embedding pages) and makes it hard to find any relevant pictures that showcase the lora (Checkpoints, and embeddings too). People are even getting away with posting pictures that don't even have the lora in the freaking prompt. It's incredibly annoying.

Agree 100%

Nobody is spamming, it is just cross-posting, which means people are posting their pictures in other topics, but they had this model in their prompt when generating, so it ends up in gallery of this embedding too.

drop it in embeddings folder then activate by putting in negative

@raredragon what do I put in the negative prompt?

@Jemo676 open web UI --> click textual inversion --> click on what u want --> move it to negative prompt --> generate

@raredragon sorry you lost me at Web UI, I just use the generator here. Thank you for your help and information, however!

@Nerfgun3 Good news: There's a tool on HuggingFace to convert embeddings from SD 1.5 to SDXL: https://huggingface.co/spaces/FoodDesert/Embedding_Converter

@kirby90815 May look into that. Long as it's easy to use. Not like trying to install Pytorch, and make a1111 actually aknowledge that Pytorch is installed. Bit of a pain, I kinda wish there was a way to make these Loras, vaes, embeddings, etc. work with all 3 mainstream base models.. I keep almost getting completely set to make a scene, but then, nope, that one thing I need to make the scene work is for the other model type.. Also, it seems a lot of the creators improperly label the assets, so I only find out after trying to run image generation.

@kirby90815 that converter is currently not working.

@silvdra00563 Shit.. I've been thinking about that thing recently.. Cuz any embedding I attempt to use, hasn't worked cuz SD1.5.. But I couldn't remember where I saw it at. Well, I just saw now, but apparently it's not working(as you said).. Any idea how long it's been not working for? Or if it will be back up again?

@Lazman I have no idea how long it's been inoperable.

Also no idea on how to message the creator either; which is an inconvenience (for me at least, not sure about anyone else).

@silvdra00563 Actually, I looked at the site again after asking. Under 'container logs' it says "===== Application Startup at 2024-11-03 22:42:35 ====="

So, I suppose it looks like it may have gone bunk on November 3rd of last year.. I wonder if there's an offline version of the tool..

Edit: I logged in and wrote a message to the guy on his other thing, so, hopefully they get back to me with something useful.

@Lazman Sounds good, and thank you for taking the time and intiative.

@silvdra00563 No prob! Looks like it paid off. they got back to me, and A: it's solved up. B: also got the code in-case it aint. https://huggingface.co/FoodDesert/Boring_Embeddings/discussions/6#678ea8cd6c337bdf15e8af28

@Lazman Nice, I'll check it out.

Thank you, sincerely, and best of luck to you and your own endeavors.

i want to know too

No, it only has one name, based on the filename (when you download from this page it's "negative_hand-neg.pt"). So if that's the filename, put "<negative_hand-neg>" in your negative prompt only. This is the case for any negatives you get (a lot have the problem of confusing keywords shown on here)

@vilifai What file name has to do with what you put in a prompt? I thought you use trigger words in prompts. Excuse me being a beginner.

for me using it on SDXL makes images worse

Don't use SD1.5 embeddings for XL, it doesn't work. Even if your program lets you, it's wrong.

You're talking about a Krita plugin? You'll have to read the docs from the plugin devs and/or A1111 (I think the plugin probably just sends requests to A1111 so install it in the usual place for that I guess)

Assuming you have put it in the right place, it may be because it's a pickletensor rather than a safetensor.