Animation I made using my Lora + a new workflow I am working on (for extending videos)
Animation I made using my Lora + a new workflow I am working on (for extending videos)
CSooo many combination..lol. I need go for a walk.. clear my brain.
PROMPT:
"close headshot portrait photograph of white ohwx man wearing a super expensive white suit, a light blue shirt and a black tie in a photo studio lighting hdr. The studio backdrop is a rich, deep black that makes the white suit pop dramatically. Professional lighting creates a perfect balance of highlights and shadows, accentuating his facial features. The suit is impeccably tailored, with subtle textures visible in the fabric. His tie is made of the finest silk, with a subtle pattern that catches the light. His hair is perfectly styled, and his expression is one of confident professionalism.<segment:yolo-face_yolov9c.pt-1,0.7,0.5>photograph of ohwx man"
EULER / BETA
30/40 steps
CFG: 1 , 2, 2.5
PROMPT:
"close headshot portrait photograph of white ohwx man wearing a super expensive white suit, a light blue shirt and a black tie in a photo studio lighting hdr. The studio backdrop is a rich, deep black that makes the white suit pop dramatically. Professional lighting creates a perfect balance of highlights and shadows, accentuating his facial features. The suit is impeccably tailored, with subtle textures visible in the fabric. His tie is made of the finest silk, with a subtle pattern that catches the light. His hair is perfectly styled, and his expression is one of confident professionalism.<segment:yolo-face_yolov9c.pt-1,0.7,0.5>photograph of ohwx man"
EULER / BETA
30/40 steps
CFG: 1 , 2, 2.5
CHey everyone,
Kicking off a new training run for my Qwen-Image likeness LoRA today. This time, I'm diving into some of the newer, more advanced parameters available in Musubi-Tuner to see if they can improve the final quality.
Here's what I'm experimenting with:
LoRA+: Using a loraplus_lr_ratio of 4, which is supposed to make the training more efficient by using a higher learning rate on part of the network.
PyTorch Dynamo: Enabled the inductor backend. I'm curious to see if the JIT compiler gives any extra speed boost on top of the performance gains from WSL.
New Timestep Sampler: Switched to the qinglong_qwen sampler. It's a hybrid method that's reportedly better for style and likeness learning.
Post-Hoc EMA Merging: After the training is done, I'll also be using the new script to merge the best checkpoints into one final model instead of just picking a single one. The goal is a more stable and accurate LoRA.
The training is running in WSL2 (Ubuntu) using the Musubi-Tuner scripts on an RTX 4090.
I'm really curious to see what effect these new parameters will have on the final result. I'll let you know how it goes!
Kicking off a new training run for my Qwen-Image likeness LoRA today. This time, I'm diving into some of the newer, more advanced parameters available in Musubi-Tuner to see if they can improve the final quality.
Here's what I'm experimenting with:
LoRA+: Using a loraplus_lr_ratio of 4, which is supposed to make the training more efficient by using a higher learning rate on part of the network.
PyTorch Dynamo: Enabled the inductor backend. I'm curious to see if the JIT compiler gives any extra speed boost on top of the performance gains from WSL.
New Timestep Sampler: Switched to the qinglong_qwen sampler. It's a hybrid method that's reportedly better for style and likeness learning.
Post-Hoc EMA Merging: After the training is done, I'll also be using the new script to merge the best checkpoints into one final model instead of just picking a single one. The goal is a more stable and accurate LoRA.
The training is running in WSL2 (Ubuntu) using the Musubi-Tuner scripts on an RTX 4090.
I'm really curious to see what effect these new parameters will have on the final result. I'll let you know how it goes!
DI got a question..in one trainer.. when setting the captions from one txt file (ohwx man) does this apply when training a acharacter.. same body chape and size..and face too... or is it only for face... my images have tags and it kind of threw me off abit..
Snice! how many seconds do you think are achievable while keeping the context
J5 seconds, since that's the 81 frames generated per clip.
I noticed that when she turned her back, and then turned to the front again (as a test) she looked different. (when reusing the last frame)
I noticed that when she turned her back, and then turned to the front again (as a test) she looked different. (when reusing the last frame)
JHopefully they will make an IP Adapter for Wan2.2
Snice
I’ve heard that some folks managed to get 7 to even 10 seconds in a single generation, but I’m not an expert in WAN so i dunno if maybe this was some diffrent case
I’ve heard that some folks managed to get 7 to even 10 seconds in a single generation, but I’m not an expert in WAN so i dunno if maybe this was some diffrent case
Jyes it can, but it may distort a bit.
any thing above a single clip length will lose context tho
any thing above a single clip length will lose context tho
Junless we get ip adapters or some kind of forging stuff
Acyber
Aso far flux krea or qwen?>
Awhat do you like
CDamn.. cant choose one. I like both. Just finished second training qwen lora with new parameters. I like result. will post soon.
CJust wanted to share an update: I've finished my new Qwen-Image LoRA training, and I'm really happy with the results! It's a huge improvement over my first attempt.
I think the new, advanced features in Musubi-Tuner made a massive difference. For this run, I used:
LoRA+ to make the training more efficient.
PyTorch Dynamo (inductor backend) to compile and optimize the model on the fly in WSL.
The qinglong_qwen sampler for a more intelligent learning process.
After the training was complete, I picked my top 3 favorite checkpoints. Then, instead of just choosing one, I merged them into a single, final model using the lora_post_hoc_ema.py script. This command averages the weights of the best LoRAs to create a more stable and higher-quality final version.
There's still some work to do on finding the best sampler and scheduler, but I'm very pleased with the outcome so far. The training took a bit longer than expected, around 4 hours on my RTX 4090, which was likely due to the initial, slow compilation phase of PyTorch Dynamo that needs time to warm up.
Here are a few example images generated with the final merged LoRA + 8stepLora.
I think the new, advanced features in Musubi-Tuner made a massive difference. For this run, I used:
LoRA+ to make the training more efficient.
PyTorch Dynamo (inductor backend) to compile and optimize the model on the fly in WSL.
The qinglong_qwen sampler for a more intelligent learning process.
After the training was complete, I picked my top 3 favorite checkpoints. Then, instead of just choosing one, I merged them into a single, final model using the lora_post_hoc_ema.py script. This command averages the weights of the best LoRAs to create a more stable and higher-quality final version.
There's still some work to do on finding the best sampler and scheduler, but I'm very pleased with the outcome so far. The training took a bit longer than expected, around 4 hours on my RTX 4090, which was likely due to the initial, slow compilation phase of PyTorch Dynamo that needs time to warm up.
Here are a few example images generated with the final merged LoRA + 8stepLora.
CDataset just 20 images like last time.
Command in Musumi-TUner to start:
python3 src/musubi_tuner/qwen_image_cache_latents.py --dataset_config dataset/dataset_qwen_test.toml --vae models/qwen_image/vae/vae/diffusion_pytorch_model.safetensors
python3 src/musubi_tuner/qwen_image_cache_text_encoder_outputs.py --dataset_config dataset/dataset_qwen_test.toml --text_encoder models/qwen_image/text_encoders/split_files/text_encoders/qwen_2.5_vl_7b.safetensors
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True accelerate launch --num_cpu_threads_per_process 1 src/musubi_tuner/qwen_image_train_network.py \
--dataset_config dataset/dataset_qwen_test.toml \
--dit models/qwen_image/diffusion_models/split_files/diffusion_models/qwen_image_bf16.safetensors \
--text_encoder models/qwen_image/text_encoders/split_files/text_encoders/qwen_2.so_vl_7b.safetensors \
--vae models/qwen_image/vae/vae/diffusion_pytorch_model.safetensors \
--output_dir output_qwen_advanced \
--output_name Qwen-LoRA-Arek-Advanced \
--mixed_precision bf16 \
--sdpa \
--optimizer_type adamw8bit \
--learning_rate 2e-4 \
--lr_scheduler cosine_with_restarts \
--lr_warmup_steps 300 \
--lr_scheduler_num_cycles 3 \
--network_module networks.lora_qwen_image \
--network_dim 32 \
--network_alpha 16 \
--network_args "loraplus_lr_ratio=4" \
--max_train_steps 3000 \
--save_every_n_steps 250 \
--gradient_checkpointing \
--timestep_sampling qinglong_qwen \
--dynamo_backend INDUCTOR \
--fp8_base \
--fp8_scaled \
--blocks_to_swap 20
Command in Musumi-TUner to start:
python3 src/musubi_tuner/qwen_image_cache_latents.py --dataset_config dataset/dataset_qwen_test.toml --vae models/qwen_image/vae/vae/diffusion_pytorch_model.safetensors
python3 src/musubi_tuner/qwen_image_cache_text_encoder_outputs.py --dataset_config dataset/dataset_qwen_test.toml --text_encoder models/qwen_image/text_encoders/split_files/text_encoders/qwen_2.5_vl_7b.safetensors
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True accelerate launch --num_cpu_threads_per_process 1 src/musubi_tuner/qwen_image_train_network.py \
--dataset_config dataset/dataset_qwen_test.toml \
--dit models/qwen_image/diffusion_models/split_files/diffusion_models/qwen_image_bf16.safetensors \
--text_encoder models/qwen_image/text_encoders/split_files/text_encoders/qwen_2.so_vl_7b.safetensors \
--vae models/qwen_image/vae/vae/diffusion_pytorch_model.safetensors \
--output_dir output_qwen_advanced \
--output_name Qwen-LoRA-Arek-Advanced \
--mixed_precision bf16 \
--sdpa \
--optimizer_type adamw8bit \
--learning_rate 2e-4 \
--lr_scheduler cosine_with_restarts \
--lr_warmup_steps 300 \
--lr_scheduler_num_cycles 3 \
--network_module networks.lora_qwen_image \
--network_dim 32 \
--network_alpha 16 \
--network_args "loraplus_lr_ratio=4" \
--max_train_steps 3000 \
--save_every_n_steps 250 \
--gradient_checkpointing \
--timestep_sampling qinglong_qwen \
--dynamo_backend INDUCTOR \
--fp8_base \
--fp8_scaled \
--blocks_to_swap 20
Awell , i still like krea more with the examples you shown yesterday
Fi dont prefer bucketing
Fbut if you have certain aspect ratio and you will use same after training
Fit helps
Fi am coding an app now for qwen. it will also have wan 2.2. it is based on bmaltais implementation
F
Sthank you soooo much
Fcoding the interface to start testing hopefully
Fit is entire image
Fso whatever in your training image entirely
Fi think results can be made better
Frecently i compared optimizers on flux
Fand our workflow was most realistic one
Doo thanks
also
I started cloud training today on OneTrainer, but I’m getting an SSH key authentication error in PowerShell. My public keys on RunPod and my PC match — I’ve verified this multiple times. I even used ChatGPT to help troubleshoot, but the issue persists.
Currently, I have no pod running before training. When I start training on OneTrainer, I get the authentication error in PowerShell, but on RunPod, the pod launches and shows as running and ready with no errors in the logs.
I’ve been advised to confirm that the SSH keys match, and they do. The last time I trained, everything worked fine, but today I’m facing this repeated error. On RunPod’s side, everything seems okay — the issue appears only in OneTrainer.
ive been at this since last night... ive gotten no where
Could you please assist in resolving this?
also
I started cloud training today on OneTrainer, but I’m getting an SSH key authentication error in PowerShell. My public keys on RunPod and my PC match — I’ve verified this multiple times. I even used ChatGPT to help troubleshoot, but the issue persists.
Currently, I have no pod running before training. When I start training on OneTrainer, I get the authentication error in PowerShell, but on RunPod, the pod launches and shows as running and ready with no errors in the logs.
I’ve been advised to confirm that the SSH keys match, and they do. The last time I trained, everything worked fine, but today I’m facing this repeated error. On RunPod’s side, everything seems okay — the issue appears only in OneTrainer.
ive been at this since last night... ive gotten no where
Could you please assist in resolving this?
Nwan 2.2 lora training?
Nwill it include 14b parameter model @Furkan Gözükara SECourses
T@Furkan Gözükara SECourses are flux trainign in main branch of kohya_ss now?
Ayou mean qwen?
TMy scripts broke because sd3-flux.1 branch is removed now 
Twe used to use it
Fi recommend massed compute
Fit is installed there
Fjust run update and use
Fyes
Fyes for a long time
Fqwen image is in musubi tuner
Fi am making gui based on that
F
Fadding more info the gui settings right now
TI'ts been a while... do we need to manually reinstall torch and xformers still or is it fixed too in main?
CCant wait to see your reasult and advice.. Want to to my FINAL qwen and flux krea LORA or full model
CDr.Furkan, if you making GUI please add all new function like :
--timestep_sampling option with qwen_shift
--qinglong_qwen
---dynamo_backend INDUCTOR
etc..
I personally liked to use command but will be great and easier when we have GUI
--timestep_sampling option with qwen_shift
--qinglong_qwen
---dynamo_backend INDUCTOR
etc..
I personally liked to use command but will be great and easier when we have GUI
CWIll be great also see later optimal config files for example 16gb vram, 24gb vram etc...
CSO you thing my example pic are bad for some reason ?? Can you share you honest opinion ?
A@Furkan Gözükara SECourses when will you drop gui?