Datasets:

---

NeuroDonu

/

PortableVersions

like 12

Hi. Well, I can try to do it closer to the weekend, I guess. But I've seen another repository, I think.

Thanks for the kind reply,
I'm not sure if Musubi-Tuner even have a GUI or some sort of User Interface, I was looking for anything user friendly such as FluxGym to allow to train Wan 2.1 LoRA locally but I'm noob into this.

I'm not 100% sure but I think there is also a GUI for Musubi-Tuner, and another one maybe but I remember tried to install it on my RTX 5090 and I had no luck so I gave up...

Anyhow,
If you know ANY other user-friendly way to train locally Wan 2.1 LoRA locally please consider to make a portable version it will be so great to try.

Thanks ahead! 🙏

I took a look at this repository... It's not as scary as it seems to me at first glance lol. i'll probably make the first test version of webui (simple enough) this weekend just to practice on the pictures for now

That will be awesome!
I still have no clue how to actually start training based on some VIDEO files, people say they're using 1-2 seconds of videos to train a LoRA for WAN 2.1 so I'm curious how to START training with it because it looks overwhelming with options and properties.

If by any chance you'll give it a try (even just for START training) please consider to drop some tips & tricks how you did it?
It will be probably much easier to follow than try to understand any repository's confusing advanced explanations.

Also, once it will work (with your upcoming portable version) I'll be happy to try it myself and share if I learn something.

Thanks ahead, please keep up the good work! 🙏

Hm. Recently i train some flux LORAs... Maybe its help to train wan

I've trained my first Flux LoRA with FluxGym it was easy (unless you'll look on the ADVANCED options which is overwhelming)
But I don't think it's simple as WAN... I may be wrong, we gotta give it a chance.
I didn't find a tutorial for LoRA training, people are mentioning properties and stuff like it's obvious... I'm just a noob when it comes to training.

I don't know if training WAN 2.1 LoRA is the same, maybe it is? it's just that the trainer options looks scary...
but it seems like many people already using it so maybe there are some "Defaults" we can use to train some tests first, from what I read people are using 3-4 videos each 1-2 seconds and they get nice LoRAs, but some do much more than that.. more videos or epochs, etc.. I guess we'll have to try it first.

There is another trainer people are using but I heard it's HELL to install on windows, "Diffusion Pipe" maybe this will help:
https://www.stablediffusiontutorials.com/2025/03/wan-lora-train.html

From my understanding Musubi-Tuner supposed to be easier to install, but training... that's a different thing of course.

Please feel free to share your experience, of course I will do the same so we can help each other but I first need a Wan 2.1 trainer that works on my RTX 5090, maybe whatever you'll decide to make a portable version will be great so we can use the same exact TOOL and share our experience and learn how to train Wan 2.1 LoRA baed on tests we'll do 💪

fluxgym is literally for people who want quick and easy. i used kohya-ss.
Right now, by the way, i'm making portable version of lllyasviel framepack with sageattention and rtx 50xx support. and i'll try to make misubi-trainer closer to the weekend

Thanks for the suggestion, currently I'm using this fork and it works with the rtx5000 series

https://github.com/sdbds/musubi-tuner-scripts

Is it complected without UI?
I still couldn't train anything successfully via the portable version because of the errors, but it was like 99% all setup worked.. until training itself of course.

Will you make a portable version for the Musubi-Tuner-Scripts?
Also, will you consider to make Batch files (.bat) as template to Train based on your experience with it, if it's all working with text?
I can't imagine how complicated it is compare to UI

Any tips will be sure helpful, thanks ahead!

Wow, ui for Diffusion Pipe? So interested

Yeah, it sounds like a GLOBAL trainer because it usually updates with the latest Models, maybe with the UI it will be easier to install / use

I'll try to finish webui musubi before the end of May

That will be great, thanks ahead! 👍

Hi again, I'm just sharing my experience so far (nothing much but maybe you can help)

So I keep looking on the internet to find a way to train locally a Wan 2.1 lora as you know and from what I've read and tried (and failed) so far: most trainers for Wan 2.1 LoRA needs the most annoying impossible almost to install: SageAttention / Triton with specific version to fit to the RTX 5090 and that may be reason why many errors appears.

• In example this one is really clean: (native Windows GUI, none-Gradio)
  https://github.com/Kvento/musubi-tuner-wan-gui
  I had to create venv and the other requirements procedure, had some issues but I had issues when I tried to train, related to Sageattention/Trition ...that's where I stopped trying, I had too many hours even trying to install these before, when I see it I run away...

• The portable Gradio one you made looked easy enough to follow,
  I got errors but also it didn't SAVE all settings but most of them so it was a bit weird but still nice UI look.

• Also the GRADIO one is interesting: (I'm not sure if this is the portable one you made? or a NEWER one)
  https://github.com/kohya-ss/musubi-tuner
  seems promising but I didn't understand how to install the NEWEST version
  installing it was hell (I bet if you'll make a portable with built-in environment it will be 1 click and done) but, I also couldn't train on it because I failed installing Sage/Triton many times before and the only EASY way was with the ComfyUI that had a script to install it for me, I don't remember where I grab it from but RTX 5090 users used to download it and ONE CLICK to follow a guide that installs also Sageattention/Triton with ComfyUI (that's the version I'm actually using).

• Diffusion pipe:
  was tempting to try on windows with all the WSL/UBUNTO etc.. and is more complicated and confusing to install and setup compare to anything I've tried so after spend too much time on it I just uninstalled everything related to it because I got stuck before I even could get anything to work.

So far Musubi-Tuner seems to be the best friend for Windows users, but...
I couldn't start TRAINING successfully even a small test with any of the versions of the GUI, and the one with the script you mentioned I didn't understand how to install at all it seems super confusing.

So I wonder if you can look around Musubi-Tuner (any of the versions) and see if you can make a ONE-CLICK from one of them like you did with the portable but I'm not sure what the error was all about maybe it was almost ready to train I can't tell but it RUNS easy thanks to you without extra installations!

I wrote this to share my experience so far in the last few days so you won't spend time on Diffusion Pipe unless for you it will be easy to install sure thing go for it👍
But if you can make ANY of the Musubi-Tuner to successfully allow me to TRAIN it will be awesome!

As always, I'll be happy to try your next portable and give feedabck or share tips & tricks if I'll ever be able to train locally on my RTX 5090 of course.

Oh, man. I honestly don't want to do musubi anymore. I'd rather do diffusion-pipe than him

Oh my! that was hell for me to even trying Diffusion-Pipe on Windows
With musubi I was "almost" able to train something... also many people who train LoRA move to Musubi-Tuner because they like Kohya-SS work and I guess it just works for them.

With Diffusion-Pipe (which sounds AWESOME because you can train almost any MODEL type!) I wasted like 2 days fighting trying to make it work with so many guides and tutorials... and even CoPilot but I kept fail I guess it's not gonna be working on Windows unless wasting days trying, but if you made it work or ANY other way I can train my first Wan 2.1 LoRA on Windows, please do share 🙏