kpsss34/PLAYJOY_2048

kpsss34/PLAYJOY_2048 is a model modified from SDXL. It can operate at a resolution of W2048 × H1024 using a dual UNet architecture. I originally made it just for fun, but the results turned out to be quite decent. However, it may require a relatively high amount of VRAM.

• My dataset: 2048 × 1024 = 10,000 images
• GPU: H100 80GB
• Training time: 4 hours
• Inference steps: 30
• CFG: 3.0–4.0

Sample images below.

For inference, you also need to load the pipeline and place it in the same directory.

Note: You should separate the pipeline py-file and infer py-file to generate images outside the directory, while all models should be in the same folder/directory.

python infer.py \
--model_dir ./model_local_dir \
--prompt "you prompt ..." \
--steps 30