Seems like it is using a seperate branch of musubi for wan22 training
Seems like it is using a seperate branch of musubi for wan22 training
Fye musubi will be also my choice
Fi will make my own gradio interface easier one
JAweomse. By the way, pro tip, I got a "flash_attn2" error running it on my RTX 5090 , however when I use the command line like this, loading CL before the loading the venv, it works just fine:
C:
cd "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\Build"
vcvarsall.bat x86_amd64
cl
C:\musubi-tuner
cd venv
cd scripts
activate
cd ..
cd ..
Ahi jonko
JI am sure the batch commands can be done less sloppy but the point is that CL needs to be loaded apparently
Acan you pls share results and your opinion when you finish wan 2.2 training? thanks
SHey does anyone tried a full finetune with Flux Krea ? Any samples to share/insights ? Curious
�What is the best finetune checkpoint for consistent character to avoid plastic flux look? Is krea or wan or qwan better than flux dev? I was optimistic about krea, was it not as good after all?
EAnyone know How to copy and paste in massed compute?
Cuse mouse copy and paste, not CTRL+C etc... sometimes not work at forst times.
CTraining wan2.2 High-Noise in musubi-tuner in progress.
Next Low-Noise and i will see result of my first wan2.2 lora.
Next Low-Noise and i will see result of my first wan2.2 lora.
Ethank you
Jt2v or i2v?
Jt2v works for me
C@Ella t2v. But I used just 30 photos.
- --task: t2v-A14B
- --dataset_config: dataset/dataset.toml
- --mixed_precision: fp16
- --fp8_base
- --xformers
- --optimizer_type: adamw8bit
- --gradient_checkpointing
- --learning_rate: 3e-4
- --max_train_epochs: 100
- --gradient_accumulation_steps: 1
- --seed: 5
- --optimizer_args: weight_decay=0.1
- --network_module: networks.lora_wan
- --network_dim: 16
- --network_alpha: 16
- --lr_scheduler: polynomial
- --lr_scheduler_power: 8
- --lr_scheduler_min_lr_ratio: "5e-5"
- --output_dir: output
- --save_every_n_epochs: 10
- --timestep_sampling: shift
- --discrete_flow_shift: 1.0
For High-Noise Training: - --min_timestep: 875
- --max_timestep: 1000
For Low-Noise Training: - --min_timestep: 0
- --max_timestep: 875
CToday will try in diffusion pipe.
Jprobably the epochs and learning rate can be tweaked for faster results, ill try it out
Jcan you post some results after training?
�I trained lora on a real person and still get plastic skin and plastic face in comfyui....what am I doing wrong here
Fit works with our configs
Fpeople tested but i didnt have chance yet on vacation
SWell deserved ! Enjoy your holidays 
J@Furkan Gözükara SECourses
For Musubi Tuner Wan 2.2 training, you can add the following arguments to the accelerate command, to make it generate samples
For Musubi Tuner Wan 2.2 training, you can add the following arguments to the accelerate command, to make it generate samples
--sample_prompts prompt.txt --sample_every_n_epochs 1 --sample_at_firstIn the promp.txt file you can add the positive prompt, and also some arguments (within that textfile!)
A blonde girl dancing --w 640 --h 480 --f 1 --d 1 --s 20 --n blurry
JAnyone else trainign Wan 2.2. and somehow realizing the plugging the Lora for the low model into the high model and the Lora for the high model into the low model, giving you better results?
JI checked the MD5 hash with the Hugging Face page, it's strange..
JYes, but managed to get my Cartoon style Lora finished today
J(Wan2.2 T2V Lora) - Cartoon Style Create Cartoon Style images and videos. (see examples) Use "r3tr0c0m1c" and "cartoon" as a prompt keyword. Resolu...
J
J
J
Fnice
JThanks!
Wan 2.2 is very cool, my favorite model to play with.
I just wish they would launch a controlnet version for 2.2 already.
Wan 2.2 is very cool, my favorite model to play with.
I just wish they would launch a controlnet version for 2.2 already.
JMy results so far is 3000 steps for 25 images at 2e4 learning rate, on dim 32 alpha 24, trained on 720x720 resolution.
Just a bit less than 2 hours training per model on my RTX 5090 without any offloading.
Lora strength of 1 works for 960x720 and 720x480. Needs a higher strength for higher res or if prompt gets too long.
Just a bit less than 2 hours training per model on my RTX 5090 without any offloading.
Lora strength of 1 works for 960x720 and 720x480. Needs a higher strength for higher res or if prompt gets too long.
Fnice info thanks
NI was interested in creating a Chorma lora using Kohya SD scripts. I'm thinking the issues I'm having are just my ignorance of command line. Does anyone know of a youtube or idiots guid for Kohya SD-SCripts? It's probably too much to ask for Chroma training, but if there was a Flux I think I can learn on that and eventually wing it. Thanks.
Jwhat does your command look like?
Fthis is literally gui for kohya sd scripts
FYouTubeSECourses
If you want to train FLUX with maximum possible quality, this is the tutorial looking for. In this comprehensive tutorial, you will learn how to install Kohya GUI and use it to fully Fine-Tune / DreamBooth FLUX model. After that how to use SwarmUI to compare generated checkpoints / models and find the very best one to generate most amazing image...
Aim getting very bad results
Aon wan 2.2 preset not working
NThank you in advanced. I'm almost ashamed to post it for fear I forgot something simple. I put the read me into chatgpt to try and help me create the command line and the dataset.toml file: accelerate launch --mixed_precision bf16 --num_cpu_threads_per_process 1 flux_train_network.py ^
--pretrained_model_name_or_path "D:/APPS/KohyaSD/TestModels/Chroma.safetensors" ^
--clip_l "D:/APPS/KohyaSD/TestsModels/clip_l.safetensors" ^
--t5xxl "D:/APPS/KohyaSD/TestModels/model-00001-of-00002.safetensors" ^
--ae "D:/APPS/KohyaSD/TestModels/diffusion_pytorch_model.safetensors" ^
--model_type chroma ^
--apply_t5_attn_mask ^
--cache_latents_to_disk ^
--save_model_as safetensors ^
--sdpa ^
--persistent_data_loader_workers ^
--max_data_loader_n_workers 2 ^
--seed 42 ^
--gradient_checkpointing ^
--mixed_precision bf16 ^
--save_precision bf16 ^
--network_module networks.lora_flux ^
--network_dim 4 ^
--network_train_unet_only ^
--optimizer_type adamw8bit ^
--learning_rate 1e-4 ^
--cache_text_encoder_outputs ^
--cache_text_encoder_outputs_to_disk ^
--highvram ^
--max_train_epochs 4 ^
--save_every_n_epochs 1 ^
--dataset_config "D:/APPS/KohyaSD/TestCFG/C4TS_dataset_multi_bs.toml" ^
--output_dir "D:/APPS/KohyaSD/TestOutput" ^
--output_name chroma-lora-name ^
--timestep_sampling shift ^
--discrete_flow_shift 3.1582 ^
--model_prediction_type raw ^
--guidance_scale 1.0
--pretrained_model_name_or_path "D:/APPS/KohyaSD/TestModels/Chroma.safetensors" ^
--clip_l "D:/APPS/KohyaSD/TestsModels/clip_l.safetensors" ^
--t5xxl "D:/APPS/KohyaSD/TestModels/model-00001-of-00002.safetensors" ^
--ae "D:/APPS/KohyaSD/TestModels/diffusion_pytorch_model.safetensors" ^
--model_type chroma ^
--apply_t5_attn_mask ^
--cache_latents_to_disk ^
--save_model_as safetensors ^
--sdpa ^
--persistent_data_loader_workers ^
--max_data_loader_n_workers 2 ^
--seed 42 ^
--gradient_checkpointing ^
--mixed_precision bf16 ^
--save_precision bf16 ^
--network_module networks.lora_flux ^
--network_dim 4 ^
--network_train_unet_only ^
--optimizer_type adamw8bit ^
--learning_rate 1e-4 ^
--cache_text_encoder_outputs ^
--cache_text_encoder_outputs_to_disk ^
--highvram ^
--max_train_epochs 4 ^
--save_every_n_epochs 1 ^
--dataset_config "D:/APPS/KohyaSD/TestCFG/C4TS_dataset_multi_bs.toml" ^
--output_dir "D:/APPS/KohyaSD/TestOutput" ^
--output_name chroma-lora-name ^
--timestep_sampling shift ^
--discrete_flow_shift 3.1582 ^
--model_prediction_type raw ^
--guidance_scale 1.0
JWhats the error you get?
JI think the --t5xxl needs to be a single safetensor file or pth file, and not a partial file (00001 of 00002)
JAlso discrete_flow_shift maybe rounded number?
Jdid you manually edit the datset.toml? how does it look?
Fwhich preset exactly you using?
NYes, I edited the toml file. Also I am working off a 4090. It looks like it attempts to run. I get the error: File "D:\APPS\KohyaSD\sd-scripts-sd3\venv\lib\site-packages\torch\nn\modules\module.py", line 1362, in convert
raise NotImplementedError(
NotImplementedError: Cannot copy out of meta tensor; no data! Please use torch.nn.Module.to_empty() instead of torch.nn.Module.to() when moving module from meta to a different device.
Traceback (most recent call last):
File "D:\PROG\Python310\lib\runpy.py", line 196, in _run_module_as_main
return _run_code(code, main_globals, None, and Toml: [general]
color_aug = false
keep_tokens_separator = "|||"
caption_extension = ".txt"
[[datasets]]
enable_bucket = true
resolution = [1024, 1024]
[[datasets.subsets]]
image_dir = "D:/APPS/KohyaSD/TestIMG/10_C4TS cats"
num_repeats = 1 A suggestion from chatgpt was to move to bf16. (on the two lines in the bat file) I changes that is and looks like it attempts to load the models. But then another set of errors. It's like moving from one set of issues to another and I'm afraid I'm just working in the wrong direction.
raise NotImplementedError(
NotImplementedError: Cannot copy out of meta tensor; no data! Please use torch.nn.Module.to_empty() instead of torch.nn.Module.to() when moving module from meta to a different device.
Traceback (most recent call last):
File "D:\PROG\Python310\lib\runpy.py", line 196, in _run_module_as_main
return _run_code(code, main_globals, None, and Toml: [general]
Common settings for the dataset
flip_aug = truecolor_aug = false
keep_tokens_separator = "|||"
caption_extension = ".txt"
[[datasets]]
This is the dataset definition for training
batch_size = 2enable_bucket = true
resolution = [1024, 1024]
[[datasets.subsets]]
image_dir = "D:/APPS/KohyaSD/TestIMG/10_C4TS cats"
num_repeats = 1 A suggestion from chatgpt was to move to bf16. (on the two lines in the bat file) I changes that is and looks like it attempts to load the models. But then another set of errors. It's like moving from one set of issues to another and I'm afraid I'm just working in the wrong direction.
Fyou are trying to train flux?
NNo, I would ultimately like to train Chroma.