Jennifer is a lovely redhead created entirely on AI generated pics
Description
FAQ
Comments (6)
Though I have seen this before, I don't recall the solution to what causes this. I get the following errors when starting A111 and it attempts to load your TIs...
*** Error verifying pickled file from E:\Automatic1111\embeddings\tv_Jennifer_MXAI.pt
*** The file may be malicious, so the program is not going to read it.
*** You can skip this check with --disable-safe-unpickle commandline argument.
***
Traceback (most recent call last):
File "E:\Automatic1111\modules\safe.py", line 137, in load_with_extra
check_pt(filename, extra_handler)
File "E:\Automatic1111\modules\safe.py", line 84, in check_pt
check_zip_filenames(filename, z.namelist())
File "E:\Automatic1111\modules\safe.py", line 76, in check_zip_filenames
raise Exception(f"bad file inside {filename}: {name}")
Exception: bad file inside E:\Automatic1111\embeddings\tv_Jennifer_MXAI.pt: tv_Jennifer_MXAI/byteorder
---
*** Error loading embedding tv_Jennifer_MXAI.pt
Traceback (most recent call last):
File "E:\Automatic1111\modules\textual_inversion\textual_inversion.py", line 203, in load_from_dir
self.load_from_file(fullfn, fn)
File "E:\Automatic1111\modules\textual_inversion\textual_inversion.py", line 184, in load_from_file
embedding = create_embedding_from_data(data, name, filename=filename, filepath=path)
File "E:\Automatic1111\modules\textual_inversion\textual_inversion.py", line 284, in create_embedding_from_data
if 'string_to_param' in data: # textual inversion embeddings
TypeError: argument of type 'NoneType' is not iterable
I remember now. Certain embedding trainers unnecessarily include a byteorder file and a .data folder with a serialization_id file inside. Automatic1111 sees these TIs as corrupt or malicious.
Opening the TIs with WinRAR, if I delete the byteorder file and the .data folder, your TIs load and work perfectly.
@Bit_Shifter This is bloody awesome! Do you have any idea how to stop automatic1111 from adding that to the embeddings? It's what I use to train them, so it's odd that it adds data that prevents the file from being seen by the software that created it in the first place.
@tvange365 Apparently, this is a known issue that has existed for some time. I just found the following on A1111's GitHub, but no solution...
This was a known issue but was fixed as of pull request 5327 in Dec. '22...
https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/3878
But the issues appears to have returned some time between then and May '23...
https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/10179
It's odd that this reoccuring issue has existed for so long and that some have it, and others don't. I used to create embeddings in A1111 before switching to LoRAs around Nov. last year. That's before AND after this bug was first reported. I had no issues back then. Trying to create one now in A1111, I get the error for every test TI I create.
Now that I think about it, I don't remember seeing or encountering this problem before A1111 v1.5. I'm on v1.6 now and have this error with every test TI I create.
I hope we find a solution soon. I've noticed more and more TIs affected by this.
@Bit_Shifter it really is odd.
From the links you share it seems to be a problem with an update to torch, but the problem doesn't exist using forge or fooocus.
I have seen updates solving one issues causing another to resurface, in many other places, so it's not unheard of. I do wonder what others who train embeddings train them on as they don't have this problem.
I really appreciate the time you've spent investigating this issue. It's been a huge help.
And now to figure out how to train a LORA... Too many videos and damn few written manuals. And I hate instructions videos!
@tvange365 I think the problem is specific to A1111's safety checker.
Give OneTrainer a try. It has all the same bells and whistles as Kohya (and more) but it's much easier to set up and use and produces equal or better results. Do a bit of research on the adaptive optimizers such as DAdapt Adam, Adafactor and Prodigy. Just getting started with LoRAs, they will remove a lot of the learning rate guesswork for you. It also trains finetunes and TIs as well for both SDXL and SD 1.5. Masked training is also another cool feature that is missing from Kohya. It also has far fewer bugs than Kohya.
https://github.com/Nerogar/OneTrainer
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.