SECourses Musubi Tuner - 1-Click to Install App for LoRA Training and Full Fine Tuning Qwen Image, Qwen Image Edit, Wan 2.1 and Wan 2.2 Models with Musubi Tuner with Ready Presets
Added 2025-09-29 13:46:00 +0000 UTCPatreon exclusive posts index to find our scripts easily, Patreon scripts updates history to see which updates arrived to which scripts and amazing Patreon special generative scripts list that you can use in any of your task.
Join discord to get help, chat, discuss and also tell me your discord username to get your special rank : SECourses Discord
Please also Star, Watch and Fork our Stable Diffusion & Generative AI GitHub repository and join our Reddit subreddit and follow me on LinkedIn (my real profile)
=======
Latest Zip File : SECourses_Musubi_Trainer_v18.4.zip
This is the initial release of the App
App screenshots gallery : https://www.reddit.com/r/SECourses/comments/1n2qo9g/secourses_musubi_tuner_v1_published_a_training/
Stage 1 Research results gallery : https://www.reddit.com/r/SECourses/comments/1n8bcrj/qwen_image_lora_trainings_stage_1_results_and/
Currently 1-click to install on Windows, RunPod and Massed Compute
We also have 1-click necessary model downloader script
The model downloader will download qwen_2.5_vl_7b_bf16.safetensors, qwen_image_bf16.safetensors and qwen_train_vae.safetensors
Please use model downloader to not have any issues because your selected models may be wrong
Moreover, the Musubi Tuner automatically does FP8 and FP8 scaled conversion while loading BF16 model into RAM
I specifically developed this model downaloder with like UGET method so that it is both ultra fast and ultra robust - not using Hugging Face downloader anymore
Moreover the downloader script verifies SHA 256 hash of the models and prevents any possibly corrupted model downloads
This is an interface app that is based on famous Kohya Musubi Tuner
It has all the features of Kohya Musubi Tuner + extra features with huge amount of details and very easy to use
Musubi Tuner Official repo : https://github.com/kohya-ss/musubi-tuner
The installer will install with Torch 2.8 and CUDA 12.9 and pre-compiled xFormers, Triton, Flash Attention and Sage Attention libraries
We literally support all of the GPUs like RTX 3000, 4000, 5000 series or, A40, L40, A100, H100, B200 etc
Flash Attention and Sage Attention may not work on RTX 2000 or 1000 series but I have plan for them hopefully soon with newest libraries compiling myself
Currently it fully supports Qwen Image model LoRA training + Qwen2.5-VL image captioning + Qwen Image Edit Training+ Qwen Image Full Fine Tuning / DreamBooth
The followings are also fully supported and their research in progress
Wan 2.1 LoRA training - example config :
Wan_2.1_text_to_video_test_lora_8500_mb.toml
Wan 2.2 LoRA training - example configs :
Wan_2.2_Text_to_Video_Low_Noise_Test_LoRA_Config_10250_MB
Wan_2.2_Text_to_Video_Both_Low_and_High_Noise_Test_LoRA_Config_12000_MB.toml
Musubi Tuner tab is from original fork - so this is not tested or supported please use Qwen Image LoRA and Image Captioning tabs
Please check all screenshots in details to see what the app really supports
28 October 2025 Update V18.7
Bug fixes and a new amazing feature called as Image Preprocessing and FP8 Model Converter
Scaled FP8 Model Converter is used to batch convert Qwen Image Fine-Tuned / DreamBooth models into scaled FP8 from BF16 weights
Extremely useful if you don't have over 60 GB GPUs and also reduces size to half
Almost same quality with using intelligent FP8 Scaled weight conversation
What this tool does is that, it preprocess your given folder image with Kohya Musubi tuner actual training code
So your pre-processed images is the images that are actually being trained by the trainer
Extremely useful to see your resolution, aspect ratios and what bucketing does
Moreover, currently Kohya doesn't use exif data image orientation so if you might get mis-oriented images surprise, use this tool to see and you can use fixed dataset as well
Detailed debug options added to latent caching section so you can enable debug and 1 by 1 look at the processed images
You can use pre-processed dataset as your training dataset if you wish saves time and resources
21 October 2025 Update V18.0
Qwen Image Edit Plus (2509) Fully supported now
Just load the DreamBooth or LoRA Config and then follow the below steps
Original files are included inside Qwen_Image_Training_Configs folder check each one
Extract latest zip file, overwrite your existing files and run Windows_Install_and_Update.bat to update
Model downloader now supports downloading Qwen Image Edit Plus (2509) model set download as well
Run Windows_Download_Training_Model_Files.bat to download desired model
Qwen Image Fine Tuning configs completed finally
Working on updating inference presets in SwarmUI
The logic of Qwen Image Edit Model training shown below
21 October 2025 Update V17.9
Qwen Image Full Fine Tuning / DreamBooth research finally completed
Hopefully full tutorial coming and more information will be shared soon
We have made configs starting from 5750_MB (6 GB GPUs) to 84000_MB (96 GB GPUs) GPUs
The lower VRAM requiring configs will use more RAM since we do block swapping and also they will be slower
I have prepared 3 different config set
200_epoch folder - best quality - lower learning rate - more epochs (more epochs takes more time linearly)
150_epoch - good quality - a little bit higher learning rate - medium epochs
75_epoch - lower quality - higher learning rate - lower epochs - faster training
All qwen configs moved into Qwen_Image_Training_Configs and seperated as LoRA_Training and Fine_Tuning_Training-DreamBooth sub folders
All configs updated including LoRA and below 20 GB configs modified like this
T5 text encoder caching will be made on CPU with BF16 - still really fast
After tutorial for this hopefully I will research Wan 2.2 training for ultra realistic image gen + video gen
20 October 2025 Update V17.8
Windows_Download_Training_Model_Files.bat updated and now you can download Wan 2.2 Text to Video and Image to Video models
I have converted official FP32 Wan weights into BF16 so use our downloader to download BF16 weights of Wan 2.2
BF16 yields much better quality at training and also slightly better or sometimes significantly better quality at inference
Example Wan 2.2 train configs added
Wan_2.2_Text_to_Video_Low_Noise_Test_LoRA_Config_10250_MB.toml
Uses 10250 MB VRAM
Wan_2.2_Text_to_Video_Both_Low_and_High_Noise_Test_LoRA_Config_12000_MB.toml
Uses 12000 MB VRAM
Example configs are using BF16 weights of Wan 2.2 now
Massed_Compute_Instructions_READ.txt and Runpod_Instructions_Trainer_READ.txt updated to have Wan 2.2 model download options
Wan training research results coming soon hopefully
Qwen Image fine tuning completed i am finalizing to publish hopefully
3 October 2025 Update V17.5
Please extract latest zip file, overwrite older files
Delete SECourses_Musubi_Trainer\musubi-tuner folder and run Windows_Install_and_Update.bat to update
Switch to low memory branch file removed since it is now merged
Bug fix on Linux systems - RunPod & Massed Compute for Wan training
Wan models training directory explanation improved : https://pasteboard.co/y5fI2Sf8Bc0L.png
29 September 2025 Status Update V17
Wan training support added
I think it is supporting all of the Wan trainings that Musubi Tuner supports
So far only Wan 2.1 Text to Video model LoRA training tested and a demo config put as : Wan_2.1_text_to_video_test_lora_8500_mb.toml
Windows_Download_Training_Model_Files.bat updated to download Wan 2.1 Text to Video models
The models will be downloaded into Training_Models_Qwen and Training_Models_Wan according to your selection
Model downloader updated for RunPod and Massed Compute as well
You can read full changelogs on app and below
Extract zip file, overwrite and run Windows_Install_and_Update.bat for update
27 September 2025 Status Update V16
Qwen Image LoRA trainings research and development completed fully
I have done over 50 full trainings to find out best parameters and workflow
You can see training checkpoints and files here : https://huggingface.co/MonsterMMORPG/Qwen_Image_LoRA_Training_Research/tree/main
To access files you can message me and i can let you with a special price but normally this is not needed by you
I have updated all the configs after recent experiments
I feel like 6e-05 is working slightly better than 5e-05 so updated configs to that
The latest experiments put into Final_Qwen_LoRA_Test_Results.zip file
I have trained different learning rates up to 1000 epochs on 28 images dataset and you can see grid results in the attached Final_Qwen_LoRA_Test_Results.zip file
So if you need faster training, you can possibly increase learning rate to like 8e-5 or 9e-5 and do like 100 or less but it is R&D depending on your needs
The app is now fully supporting Qwen Image full Fine Tuning / DreamBooth
I added a demo config file : dreambooth.toml
This demo config toml uses 5 GB VRAM only when you also run Windows_Switch_Low_RAM_Branch_Temporary.bat
There is a pull request of Kohya which reduces used shared VRAM signficiantly but it is not merged yet so we switch to that branch with that bat file
When you run Windows_Download_Training_Model_Files.bat again it will switch back to main branch
I am starting to fully research Qwen Image full Fine Tuning and also adding Wan 2.2 and Wan 2.1 training capability to the app hopefully as soon as possible
Installers updated to Torch 2.8, CUDA 12.9, Flash Attention 2.8.3, Sage Attention 2.2, xFormers 0.0.33
I have pre-compiled libraries for both Windows and Linux and working amazing
Windows Requirements
Python 3.10.11, FFmpeg, CUDA 12.9, cuDNN 9.12 or above, C++ tools, MSVC and Git
Don't worry CUDA 12.9 works with all GPUs
Follow this requirements tutorial video exactly : https://youtu.be/DrhUHnYfwC0
Follow its updated post with links and screenshots exactly : https://www.patreon.com/posts/click-to-open-post-used-in-tutorial-111553210
5 September 2025 Status Update V14
Some bug fixes and Configs_v1 added that you can use
New option added for sample generation Disable Automatic Prompt Enhancement
Thus you can provide fully customized prompt txt
Space character having paths will now work but still dont use space character in any path
.e.g my awesome images - is wrong but my_awesome_images is right
We support as low as 6 GB GPUs
Tier 1 is better than Tier 2 and Tier 2 better than Tier 3 and so on
There shouldn't be any big difference between Tier 1 and Tier 2
Tier 2 and Tier 3 also should be pretty close in terms of quality
To update download latest zip file, extract and overwrite older files and run Windows_Install_and_Update.bat
Some Raw Generation Results of 2_september_best_config.toml
No Uscale, No Face Restore
All 8 Steps with 8-steps LoRA and Preset We have
Trained for 200 Epochs, 28 images = 5600 steps
All are 1328x1328px native res
31 August 2025 Update V11
Now you can immediately and properly stop batch image captioning as well
Some Critical config save load bugs fixed please upgrade
e.g. : FIXED: Critical checkpoint removal bug - Checkpoints were being deleted immediately after saving when save_last_n_epochs=0
Now we give GPU ID = 0 as a default, if you have multiple GPUs like onboard GPU and external GPU, make sure the ID matches to your external GPU ID
It is set in Distributed GPUs and that may look like multiple GPU config but when no multiple GPU enabled, it still uses it
UI significantly improved please check newest screenshots given at the very top
The downloader app improved and now it will show the progress of each model on a single line not in multiple lines as it downloads
Added comprehensive parameter support for Qwen Image training with 100% coverage of Musubi Tuner parameters
NEW: Integrated search bar in Qwen Image Training tab - quickly find any setting without opening all panels
Tab renamed from "Qwen Image LoRA" to "Qwen Image Training" to reflect both LoRA and Fine-tuning support
Implemented Qwen-Image-Edit mode support for control image training (experimental - not fully tested)
Added control image resolution settings for Edit mode (dataset_qwen_image_edit_control_resolution_width/height)
Introduced dataset_qwen_image_edit_no_resize_control option for maintaining original control image sizes - this is for Qwen Image Edit model - not tested yet
Started implementation of Qwen Image Fine-Tuning mode (DreamBooth) - parameter infrastructure in place
Enhanced FP8 quantization descriptions with clearer GPU compatibility information
Improved timestep sampling with better documentation of qwen_shift vs standard shift methods
Added advanced flow matching parameters (logit_mean, logit_std, mode_scale) for fine-tuned control
Implemented complete VAE optimization settings (tiling, chunk_size, spatial_tile_sample_min_size)
Enhanced parameter descriptions throughout the GUI for better user understanding
All critical default values confirmed to match official Qwen Image documentation
Parameter accuracy validated at 100% for all implemented features
Some configs save and load were broken and all fixed hopefully
Please report if you notice error
e.g. one of the fixed one is GPU ID set
Please use latest zip file, overwrite previous files and just run Windows_Install_and_Update.bat
30 August 2025 Update V7
Dataset TOML file generate error fixed
Qwen2.5-VL image captioning turns out working perfect on Windows
It turns out my model file was corrupted even though it was same size
Therefore I have updated the model downloader and now it will check and verify SHA 256 of files therefore it will be 100% accurate
Prompt file selection folder icon issue fixed
Downloader file will use generated venv of installation
Make sure to run it after installation completed
Fixed skip existing captions functionality in Image Captioning with Qwen2.5-VL
Previously skipping was happening after caption generation which was destroying the skip logic
Now properly checks for existing captions before processing, significantly improving efficiency
Added full batch captioning status display in command line with progress tracking and ETA
Enhanced config save/load functionality for better reliability
Improved interface of Image Captioning with Qwen2.5-VL for better user experience
Various error fixes in the Qwen2.5-VL captioning pipeline
Fixed broken config save and load functionality for Optimizer Arguments and Scheduler Arguments
Improved Stop Training button responsiveness - now appears much earlier when Text Encoder caching starts
Enhanced training control for better user experience
A new full dedicated section for Sample generation implemented
It will automatically format your given sample txt file with the settings you set on GUI
So you just type prompts into txt file with new lines e.g.
ohwx man wearing a very nice amazing suit
ohwx man driving a luxury car
Please use latest zip file, overwrite previous files and just run Windows_Install_and_Update.bat
How To Give Accurate Dataset for Dataset Prepare
Example folder : E:\training_imgs_28_1328
So you give above path into UI
Inside this folder example sub folder for dataset images
E:\training_imgs_28_1328\1_ohwx man
Inside E:\training_imgs_28_1328\1_ohwx man
I have my images as below
How To Install and Use:
Windows:
Full tutorial coming soon hopefully
Use same folder logic of Kohya and use Generate Dataset Configuration button it will handle all
e.g. Parent Folder > sub folder like 1_ohwx man
Just use Windows_Install_and_Update.bat for install and update
Massed Compute (Recommend Cloud) :
Please register via this link : https://vm.massedcompute.com/signup?linkId=lp_034338&sourceId=secourses&tenantId=massed-compute
We have a special coupon for all GPUs : SECourses
If you want to learn more about GPUs and prices read this link : https://www.patreon.com/posts/126671823
Select RTX A6000 or Better GPU - like L40S or A6000 ADA or A100 or H100 or now RTX 6000 PRO
Then select our image SECourses from Creator dropdown
Then follow Massed_Compute_Instructions_READ.txt
Same as my any other Massed Compute installer script
Example tutorial for learn how to install and use Massed Compute
(Starts at 12:58) : https://youtu.be/KW-MHmoNcqo?si=G1WbG-Qw4ujWvOtG&t=778
RunPod (Cloud):
Please register via this link : https://get.runpod.io/955rkuppqv4h
Then follow Runpod_Instructions_READ.txt
Same as my any other RunPod installer script
Use the template written in Runpod_Instructions_READ.txt file
Example tutorial for learn how to install and use RunPod
(starts at 22:03) : https://youtu.be/KW-MHmoNcqo?si=QN8X8Sjn13ZYu-EU&t=1323
To See All Screenshots : https://www.reddit.com/r/SECourses/comments/1n4qq8y/massive_updates_and_improvements_to_secourses/
Comments
For the dreambooth training I copied your Tier1_20800_MB config and edited that copy, since that config was for 24 GB GPUs according to the image about it (I have 24 GB GPU and 256 GB RAM). Though outputting a sample at 1328x1328 takes about 5 mins 20 secs, which is quite a bit longer than outputting a qwen lora sample on the same PC which took up to about 2 mins 40 secs at full res. Is there an easy way to increase the speed of output samples for dreambooth training on this pc to more like 2 mins? Should I go with the Tier 1 13500 MB config (editing a copy) or just increase the block swapping? Will increasing block swapping a bit allow for faster output samples while not making training too slow?
cool1
2025-10-29 14:35:45 +0000 UTCyou can do that with extra args --debug_mode image Use `--debug_mode image` to display dataset images and captions in a new window, or `--debug_mode console` to display them in the console (requires `ascii-magic`). but i will add as a checkbox
Furkan Gözükara
2025-10-29 10:27:52 +0000 UTCFor the dreambooth is there an option or could one be added to make it not show every caption from every file in the dataset (maybe when it's caching). Surely showing every caption from the dataset in the cmd window (especially if they're long) will slow down processing of them, so maybe a "show each caption when caching" checkbox could be added please. Maybe the same for images (though there's less text shown there).
cool1
2025-10-29 10:02:35 +0000 UTCtrue this is a known issue. sadly i couldnt fix this problem yet and i havent seen anyone fixed. a class bleeds into others
Furkan Gözükara
2025-10-25 16:48:44 +0000 UTCSubject: Issue with LoRA - Trained Character's Features Bleeding onto Other Characters in the Scene Hello, I have trained a character LoRA using your tool and am facing a specific issue when using it for image generation. The Problem: When I generate an image with my main character, the character itself looks perfect and just as intended. However, the LoRA's influence seems to "bleed" onto other characters present in the same scene. For example, if there are background characters or other people in the image, their faces and sometimes even their body structures start to resemble the character my LoRA was trained on. I have tested this LoRA with the "Qwen Image" generator and have also tested AI-Toolkit of ostrich, but the problem persists everywhere. This leads me to believe that the issue might be with my training dataset or the training settings, rather than the specific image generation tool. I would be grateful if you could provide some guidance on the following questions: 1. **Training Process:** What should I be mindful of during the LoRA training process to prevent this "character bleed"? Are there specific settings, dataset preparations, or captioning techniques I should use to ensure the LoRA only affects the intended character? 2. **Commonality:** Is this a common problem that many people face when training character LoRAs? 3. **Potential Solutions:** Are there any established methods or solutions to fix this issue, either during training or during the image generation (prompting) stage? Any help or advice you can offer would be greatly appreciated. Thank you for your time and assistance.
Yogesh Patel
2025-10-25 16:21:20 +0000 UTCwell i did 10s of training all worked. hard to know without seeing everything. how many images you had? what captioning used?
Furkan Gözükara
2025-10-23 09:44:11 +0000 UTCQwen Image Lora works great!!! But the Full Fine Tuning has no effect. I used the /200_epoch/Tier1_84000_MB.toml . The trained model produces the same images as the original model. What could be wrong? I use ComfyUI...
Klaus
2025-10-23 09:12:18 +0000 UTCyes increase block swap count by 2. and try again until you wont get error. are you using all images 1024x1024?
Furkan Gözükara
2025-10-01 15:07:26 +0000 UTCdo you have recommendation for "CUDA error: out of memory" for 4090 24GB ram and PC with 64GB ram? using Tier1_21700_MB_VRAM_Need_More_RAM.toml
Ec Jep
2025-10-01 14:59:01 +0000 UTCI didn't have chance yet. usually native resolution of yielding model is best. this was case for both SDXL and FLUX.
Furkan Gözükara
2025-10-01 13:56:48 +0000 UTCHello ! thank you for your work ! on Qwen training have you ever tried 1024 instead of 1328 ? and if so, did you noticed any change in final quality ? When you train with 1328 resolution , how many it/s do you have ?
Baekdoosixt
2025-10-01 07:22:19 +0000 UTCi will add it. currently you can use this installation. just activate venv and execute scripts. but it takes time for me to test and prepare configs. please wait me to. we support everything kohya support since we use it directly
Furkan Gözükara
2025-09-17 05:18:19 +0000 UTCHello Furkan, sorry but the headline of this post is misleading. I just installed this, only to discover that WAN 2.2 in not supported by the GUI. Could you at least publish non-gui sample training scripts for the mentioned Wan 2.2 LoRA training for both Images AND video clips as training data? 24GB VRAM and 64 GB Ram should be enough for clips. Info from Kohya: FP8 support and memory reduction by block swap: Inference of a 720x1280x81frames videos with 24GB VRAM, training with 720x1280 images with 24GB VRAM
Jason Dawn
2025-09-16 07:16:32 +0000 UTCi think try to find official comfyui qwen lora workflow. it should work right away since it works in swarmui
Furkan Gözükara
2025-09-15 21:50:56 +0000 UTCIs there a comfyui workflow we can get from anywhere that would allow us to use the Qwen Image model with a qwen image Lora that we trained using Musubi tuner trainer? I tried changing a flux lora workflow to make it work with qwen but it didn't. Maybe because the "type" is set to Flux but in the drop-down list there's no "Qwen Image" type listed. There's the workflow for qwen with lightning lora but changing that to a different lora doesn't work.
cool1
2025-09-15 17:29:13 +0000 UTCi added all the features. it has debug option so you should enable debug and see how it actually process. also i will add every GPU VRAM config hopefully
Furkan Gözükara
2025-09-02 19:00:16 +0000 UTCThanks for doing the testing. For the different configs that use different amounts of VRAM, could you also (if it's possible) put what the training time would likely be on PCs with different GPUs with those configs (that have enough RAM)? Also you say "With 8 steps Fast LoRA results are thankfully much better". It seems to be yes (though I couldn't get the Lora working swarmui) but in the test results the Fast Loras also change Donald so it looks less like him than without the Fast Lora I think. In the Kohya SS training, in the bucketing section there used to be more options, including a max bucket size (which we could set to twice the max res width or height - if your training aspects were going to be up to about 2:1 ratio). In this GUi, it has only 2 settings - "enable bucketing" and "bucket no upscale". How do we know what sizes it's going to bucket up to on this one and is there a reason we don't get control of that? Is the bucket size automatically getting set so the lowest res side of the max res width or height? Or is it based on the size of the source image (which could be higher than training size, and if so couldn't that slow training down?
cool1
2025-09-02 15:50:21 +0000 UTCSame with RTX3090, except it's even slower with that (1 training step about every 19 mins after it gets to step 2, sometimes more (training step 1 is a lot more)). Though step 3 was 15 mins with "use fp8 text encoder". I assume it could be faster if the res was lowered from max 1328x1328 though I don't know how much and it would be under the normal res of the model and lose some of the advantage of it.
cool1
2025-09-01 12:27:18 +0000 UTCDoes anyone know how to fix `ValueError: Reserved escape sequence used`? The app is generating TOML files with illegal escape sequences, I am not sure how to fix this. I am using folders with names like "1_ohwx" Edit - this appears to be related to the python code where it appends the enhanced prompt filename to the directory path, however it doesn't escape the forward slash "/" that gets inserted in between the directory and filename Edit2 - removing the path to "pompt.txt" fixed the issue Edit3 - current config does not play well with 4090. About 1 step every 10 minutes.
Glen Carpenter
2025-08-31 23:59:50 +0000 UTCThanks for this, look forward to your tests!
Caleb
2025-08-29 02:59:25 +0000 UTCnot ready yet. i will do 10s of maybe 100+ trainings to find best. also training is fast like FLUX
Furkan Gözükara
2025-08-28 23:12:18 +0000 UTCThanks. Do you have example images of a trained qwen lora yet that you can compare with flux1 dev or flux1 krea trained lora? So the Musubi tab doesn't work at all yet - or is it just not tested? If that worked or when you create other tabs for it would we be able to train all types (including the qwen lora, dreambooth, wan image and video) on 24 GB GPUs (though I think someone said even short video lora training would take ages)?
cool1
2025-08-28 23:05:08 +0000 UTC
