Trained on Wan 2.1 T2V-14B model locally with 20 publicly accessible images using ai-toolkit trainer.
Use with LoRA strength between 0.7-1. Trigger word is “woman”.
Example prompt (ComfyUI): A woman, wearing a shirt with a white blazer, with diamond earrings, looks directly at the viewer. Her hair swaying gently, her subtle smile widening slightly as she blinks, her eyes sparkling with a slight movement.
Want a custom/private LoRA? Good news—commissions are open! Request yours here: https://ko-fi.com/de_zoomer/commissions
Background
I've been deeply exploring how to create LoRAs with 100% accuracy to the original character. My focus is on quality of both likeness and generations.
I could use most of the knowledge from Hunyuan LoRA creation process, and I've consistently stayed up to date with the latest releases, exchanging knowledge in their communities.
My expertise is mainly with characters, so I’m not as familiar with LoRAs for style or anime, although the process might not differ too much.
If you want your own custom LoRa, feel free to message me! Commissions are open—check out my Ko-fi link above.
Enjoy using my LoRAs and have fun!
Description
FAQ
Comments (5)
Can you share your training config ?
Is this a I2V or T2V model?
That's really impressive for only 20 images.
Any settings or configs would be great.
Nice job on this lora.
Looks like a couple of your generations are a bit over-cooked. I've noticed that lowering cfg by 0.5-1.0 can improve this.
Amazing work on your Ana de Armas LoRA for Wan2.1 T2V 14B! I saw you achieved great results training on only ~20 images, with the final LoRA being around 300MB.
I'd love to learn from your experience as I'm training a similar LoRA on my RTX 4090 (24GB) using Diffusion-Pipe.
Could you possibly share some specifics of your training config? I'm trying to figure out the optimal settings, especially:
What resolution did you train at?
What LoRA rank did you use (to get the ~300MB size)?
What optimizer (e.g., AdamW8bitKahan?), lr, and LR scheduler settings worked well?
Roughly how many epochs did you train for?
What did you set for gradient_accumulation_steps and num_repeats in the dataset config?
Did you need to use blocks_to_swap, and if so, how many? Was activation_checkpointing enabled?
For captions, did you use detailed descriptions or just a trigger word?
Understanding these key parameters would be a massive help. Thanks for considering!
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.