text2image lora to mimic Marat Safin style photography.
QWEN:
use dpmpp_2m + beta sampler, 25 steps
WAN:
Made for Wan2.1-T2V-14B. But I recommend to use with Wan2.2 with a single LOW noise sampler. Don't use 2 samplers (high+low), it will result in wrong composition.
Trained on 55 horizontal images.
Recommended settings for WAN t2i: LOW noise checkpoint, 1280x720 (horizontal orientaion is preferable), 6 steps, lightx2v lora 0.50, fusionx lora 0.50, Sampler: res_2s, scheduler: bong_tangent
lora strength: 1.00
trigger word: Marat Safin style
Example prompt: Marat Safin style. Woman is sitting on the floor by a bathub at bathroom in a modern apartment in a crop top and denim shorts. At first she is looking at viewer then looking down. upper body shot, side view.
wan workflow json is attached as training data.
Description
for qwen-image and qwen-image-edit
FAQ
Comments (1)
This is amazing. When you make lora like this, do you remove characters from the dataset? I'm asking because it's so consistent with the character lora I created







