Furkan Gözükara patreon

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 UTC

Comments

cool1

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?

Furkan Gözükara

you 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

cool1

For 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).

Furkan Gözükara

true this is a known issue. sadly i couldnt fix this problem yet and i havent seen anyone fixed. a class bleeds into others

Yogesh Patel

Subject: 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.

Furkan Gözükara

well i did 10s of training all worked. hard to know without seeing everything. how many images you had? what captioning used?

Klaus

Qwen 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...

Furkan Gözükara

yes increase block swap count by 2. and try again until you wont get error. are you using all images 1024x1024?

Ec Jep

do 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

Furkan Gözükara

I didn't have chance yet. usually native resolution of yielding model is best. this was case for both SDXL and FLUX.

Baekdoosixt

Hello ! 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 ?

Furkan Gözükara

i 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

Jason Dawn

Hello 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

Furkan Gözükara

i think try to find official comfyui qwen lora workflow. it should work right away since it works in swarmui

cool1

Is 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.

Furkan Gözükara

i 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

cool1

Thanks 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

Same 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.

Glen Carpenter

Does 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.

Caleb

Thanks for this, look forward to your tests!

Furkan Gözükara

not ready yet. i will do 10s of maybe 100+ trainings to find best. also training is fast like FLUX

cool1

Thanks. 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)?