After a gap of two years, I trained using 20,000 images, and the SD lightning model.
From this I can provide this nice LoRA up weights. Feel free to use.
This repo is for testing.
I need your responses and opinons.
RealVisXL V4.0 V4.0 Lightning (BakedVAE) is used
Thus, Hires fix is a must! 2.0 scale, 0.6 denoise.
I may change the base model, current one have difficulty in generate a full body picture. (I suspect they only train the face...)
v2.1 using the larger base version, pls notice.
Plan:
Testing Optimizing Parameters (Currently train unet only)
Change base model
How to remove water mark at preprocess? (May need help)
DPO + LoRA (Large amount of code, it takes time)
GPU resources is still needed.
Description
https://civitai.com/models/139562/realvisxl-v50?modelVersionId=789646
Based on RealVisXL V5.0 Fp16 version
Training Datasets
32126 images (using yolo + inpaint remove all watermarkers)
A watermark safe dataset.
4 buckets
1280 * 1024
896 * 1280
1024 * 1280
1280 * 896
TODO:
I think DPO is a must, i may do it later. Code is not a problem.
wd-tagger is good, I am think of retag with MLLM.