and by the way Nano Banana Pro is impressive at generating synthetic data... I us gemini 3 Pro to generate the prompt for nano banana pro to generate different angles & expressions, all you need is 3 or 4 high quality reference shots...

replied to obsxrver's post about 2 months ago

thanks again @obsxrver your template is working amazingly well, I already have generated 3 T2V loras, I have a question : as my current objective is simply to face swap and I have already my highnoise lora performing well at generating hair color, hair cut and face shape, do you confirm I only need to generate low noise models with better training dataset to improve the face swap fidelity ?

replied to obsxrver's post about 2 months ago

finally I tried on Vast.ai & ~50 min on a H100 to run the high or low with 16 pics, very easy to use, thank you ! I just failed to setup the cloud upload at the end (I ticked the "Upload LoRAs to cloud storage after training" but nothing happen) , I had to scp to retrieve the files

replied to obsxrver's post about 2 months ago

Thanks @obsxrver for this template, I'm fortunate to have a 5090, would it be easy to create a local run version ? I'm currently looking at the code & scripts, I'm not sure how much this is dependent to vast.ai
Thank you

reacted to obsxrver's post with ❤️ about 2 months ago

Post

2993

(https://github.com/obsxrver/wan22-lora-training)
If you’ve been wanting to train your own Wan 2.2 Video LoRAs but are intimidated by the hardware requirements, parameter tweaking insanity, or the installation nightmare—I built a solution that handles it all for you.

This is currently the easiest, fastest, and cheapest way to get a high-quality training run done.

Why this method?

* Zero Setup: No installing Python, CUDA, or hunting for dependencies. You launch a pre-built [Vast.AI](http://Vast.AI) template, and it's ready in minutes.
* Full WebUI: Drag-and-drop your videos/images, edit captions, and click "Start." No terminal commands required.
* Extremely Cheap: You can rent a dual RTX 5090 node, train a full LoRA in 2-3 hours, and auto-shutdown. Total cost is usually $3 or less.
* Auto-Save: It automatically uploads your finished LoRA to your Cloud Storage (Google Drive/S3/Dropbox) and kills the instance so you don't pay for a second longer than necessary.

How it works:

1. Click the Vast.AI template link (in the repo).
2. Open the WebUI in your browser.
3. Upload your dataset and press Train.
4. Come back in an hour to find your LoRA in your Google Drive.

It supports both Text-to-Video and Image-to-Video, and optimizes for dual-GPU setups (training High/Low noise simultaneously) to cut training time in half.
Repo + Template Link:
https://github.com/obsxrver/wan22-lora-training
Let me know
if you have questions