But a checkpoint with krea Lora works
But a checkpoint with krea Lora works
CI'm currently training a test LORA on the wan22 low noise model, specifically a LORA for a person or character. From what I understand, it's best to train it on a low noise model. It's a fast training, only 30 images in diffusion pipe wsl. It should take about 1.5 hours.
Almk how it went
AI tried wan 2.2 t2i and I was very disappointed
Atext was all wrong and textures / quality same as krea
CI will
Ais this musubi
CDiffusion-Pipe - best for me. using many times and created many nice Loras for wan2.1
Fi think kohya lora extraction is not working accurate
Janyone know how to upload finetuned flux dev model ( dreambooth style ) to replicate so i can call it with a api?
F
Fwe had requested base model support i am not sure if he added
Dcan i run multiple instances of onetrainer(gui) or at least cmd? because i don't seem to find any batch processing in the interface
Dlike if i wanted to train the same dataset with the same settings but different base models (for a lora) it would be nice if it would do them after another over night. right now i run it, wait around 1-2hours and then start a new one
Fyou cant queue
Fdo you have multiple gpus?
Dif we count the 4090 in the cupboard, ready to sell, yes. no, just the 5090 in a pc. i was hoping it would be like with gradio stuff
Dso IF i could open it two times, it would do it at the same time and fk things up because of that? or would one of the instances just say nope
Fyou can open 2 instance
Fand set different cuda device as well
Fbut if you have single gpu i dont see benefit
Fif you do that change folders on each one to different
Fso wont conflict
Dah ok. my thought would be since once instance runs like 1-2 hours , that i could let it do 4-5 trainings over night
Dbecause when i sleep i and one is done, i cant start the next one with a different checkpoint
Fi dont know if queue
Fyou can try
Jkoyha or musubi may perhaps let you train from the command line (batch)
J@Dr. Furkan Gözükara Perhaps interesting for you, trying out now for Wan22
https://civitai.com/articles/17740
https://civitai.com/articles/17740
The basis for this workflow is my original WAN2.1 training guide: My WAN2.1 LoRa training workflow TLDR | Civitai In this new article I will explai...
JSeems 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