
this model is trained on 512 * 512 resolution.
I used dataset created by https://huggingface.co/datasets/p1atdev/pvc, nsfw images are removed using another model. Only trained on gsc images. Multi object images are removed in dataset before training (I use blip2 to remove this multi object images, but there might be some left).
Updates in progress:
train this model on a better resolution. I tried to generate images in 1024 * 1024, but got very INTERESTING results, the pattern is similar to what is suggested in SDXL's model card page on hugging-face.
I was using only 2k images from the dataset, while the dataset has 14k images. I'm working on updating my scripts to support training from all images.
example prompt:
"full view of sksgsc pvc figure,1girl,wide shot,jk,detailed cute face,detailed bright eyes, vivid, scenery, masterpiece,flawless, clean,best quality,sharp edges,cinematic film still,realistic"
Any suggestions welcomed!
描述:
训练词语: sksgsc pvc figure
名称: pytorch_lora_weights.safetensors
大小 (KB): 22842
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success