Since no one else would do it... I guess let me do it.
Target File: Replaces training/sd-scripts/library/anima_utils.py.
Shameless vibecoded fix for this trainer which was deprecated way too soon.
I'm a bit out of the loop as to why the Dev crashed out, but if I had to guess, it's likely because someone likely called it "vibecoded slop" or something, and seemingly made him crashout.
If that is the scenario, then that is insane.
This trainer if it were made with AI or not, it was the fastest and most efficient trainer for a modern model to this day, and it still takes a level of competence to create such a thing that a majority of people simply don't have.
Ultimately, just show respect to people. Especially people who work with technical stuff, as typically technical stuff is where people who have problems elsewhere in life find their solace and peace.
If you still need the trainer itself, thankfully the dev seems to have left this link alone:
https://huggingface.co/datasets/ThetaCursed/Anima-TrainFlow-Portable/tree/main
Appreciate my effort? I will admit, this was an easy fix if you have intermediate to expert knowledge.
But either way, you can buy me a coffee here:
Description
Comments (8)
Thank you, really like this trainer easy to use no hassle I was extremely disappointed it disappeared from github
Fast as fuck too lol
Very easy on the system RAM and VRAM
Should've never been removed...
What is this fixing, exactly?
New Anima architectures dropped recently.
2.9B and I believe there's another one 3.5B
Something about the layers and tensor matrix being different from Anima 1.0
It's a small fix for the error that occurs when attempting to use this trainer to train 2.9B loras
No clue if it works for 3.5B, but supposedly it uses two text encoders.
So probably need a new trainer all around for that one.
Vice versa, if you attempt to simply use 1.0 Anima loras on 2.9B models, the image will fry
any possibility of making another patch for the Anima 3.8B?
Unfortunately, my hardware isn't even good enough to test if it would work or doesn't work
That's why I haven't done it
I have the will to do so, but simply lack the capability
@Intosica that sucks, I got the hardware but none of the skill or any idea how this things work. I have access to gemini pro thanks to a promotion google was running for 9.99 a month. Send me the info I just tested 3.8B and love the results really want some lora's based on it.
@kitekholin My advise is switch to Claude.
I fell for the Gemini trap too at first.
Gemini was great at the beginning of year, but it's recently lost a ton of competence and intelligence.
You need to turn the source files (everything except the venv files) into a zip file, then tell Claude the ordeal along with giving it the zip file.
Claude will likely ask for a debug log of the model's tensor matrix. Make sure to mention that you need to add a optional second CLIP section that's toggle-able for 3.8B usage.
Best way to get that log may be through a ComfyUI node, or if that option is limited, tell Claude to make Anima Trainflow start printing it through a log in the Web-UI terminal.
You can easily just screencap what I said here, and paste it to Gemini to start the game plan up.
Tell Gemini that you want Claude to do it as to spare the load on Gemini itself.
You could use Gemini, but you'll be doing a ton of manual labor trying to help it through, and it will often try to take shortcuts and will often hallucinate to spare the load on it's own server.
Only use Gemini for instructions and/or planning.
I also highly recommend you cancel that Gemini subscription, just from one AI user to another.
Use container tabs and separate accounts until AI is affordable and worth the price.
With Gemini you don't even need container tabs, just use 2 different emails.
You'll see no difference.
My experience:
Made Gege Node suite with it starting January this year.
Was phenomenal.
Overtime, I can confidently say, Gemini has lost competence to do something similar again.
Furthermore, Gemini despite being made by same people who made TurboQuant will just lose context nowadays, and will completely forget chats literally one message above itself.
Reecently, I tried porting "Scalpel" from that node suite to these new Qwen CLIP models.
Gemini now takes a "let me just wing it" approach, as a oppose to "how about we investigate what's going on to see what the problem is?"
Claude is only public LLM that seems to actually attempt to reason still that I can think of, whilst also accepting file uploads.
Ultimately you can't drag and drop the source files and have it analyze them in good faith either with Gemini.
You will receive a pure hallucinatory shortcut.
I personally expected better from Google as well, but for some reason, they're getting demolished in the AI rat race this year.
I genuinely think it's part of the AI bubble deflating and/or shifting.