Easy to use Lora Trainer(:3000) with ComfyUI(:8000) for testing and VsCode(:8888).
https://runpod.io/console/deploy?template=9j2tjwyxys&ref=0eayrc3z
Has diffusion-pipe pre-installed on the container
File Management UI needs some more work to streamline the process but can be run by going to
1. cd /workspace/file-manager
2. source /workspace/bcomfy/bin/activate
3. npm install
4. node app.js
visit the https://xxxxxxxxxxx-3000.proxy.runpod.net/Current functionality
Current functionality of file-manager
Select a name from the Wan models on Hungging Face paste it to the file manager and it will download to the file (Needs better loading UI)
Uploads traning data = image and text files
Settings can modify the tomal files. Ensure whatever model you add is also added to the ckpt_path = 'Wan2.1-14b'
Manually run traning but hope to create a full ui that can do everything
Use this method for now:
Highly recommed to follow this Youtube tutorial on how to setup traning data. This template is fully setup with everything you need, so just skip to the traning part.
馃搧 File Manager - Quick Start Guide
Download Wan Models
Select a model name from Hugging Face (e.g.,
Wan-AI/Wan2.1-T2V-14B)Paste it into the file manager download field
Click download (Note: UI loading indicators will be improved in future updates)
Upload Training Data
Use the file manager to upload your training images and text files
Files will be automatically placed in the correct input directories
Configure Settings
Use the Settings tab to modify the TOML configuration files
Ensure your model path is correctly set in the configuration
Example: Verify
ckpt_path = '/workspace/diffusion-pipe/models/wan2.1-14b'matches your model
Start Training Manually
Currently, training must be initiated via command line
Navigate to the correct directory and run the training command:
cd /workspace/diffusion-pipe NCCL_P2P_DISABLE="1" NCCL_IB_DISABLE="1" deepspeed --num_gpus=1 train.py --deepspeed --config examples/wan_video.toml
Future updates will include a complete UI for all training functions.
Description
Creating a WebUI to handle Testing Data, Download Models and run traning (coming soon)
FAQ
Comments (8)
Could you release any of the Loras you have made with it?
working on it, will add a full write up tutorial
Hi,
I used your video to get started with WAN鈥攖hanks for that!
I wanted to ask about your experience with character LoRAs and WAN. Somehow, I feel like it just doesn鈥檛 quite get it right. I鈥檓 using a 1.2 strength on my character LoRAs to achieve at least decent results, but even then, it鈥檚 not perfect. I noticed a similar effect in your sample as well鈥攇ood, but not quite there yet.
I鈥檝e been following the WAN LoRA list, but I haven鈥檛 seen any character LoRAs appearing. Do you think WAN is just not as effective for character generation?
Would love to hear your thoughts!
Thanks!
Yeah ive had the same issue. My intial Lora came out well and got good likenees but was unable to relicate the same results. It might be due to having good resolution traning data as well toml settings.
I havent dont much traning and still leaning I'm going to test out having all the images in the same ratio (1024x1024 or 512x512) and playing around with the settings. When I find something good i'll make that default on the template but for now i'm still testing.
Is it possible to train with 16GB VRAM?
The models are huge. I was unable to even with the 1.3b model with this setup
this works for I2V (Image to Video) ? Great work nonetheless! thanks
Incredibly valuable resource especially with Diffusion Pipe included.
]it's actually mind blowing how good this is, can be adapted for training all models on Diffusion Pipe, Flux training etc
Probably the most valuable thing I have ever found for training. Everything I ever needed to organise shit in one package!
How have I only just found this??
Seriously great work, thank you so much!